ORIGINAL RESEARCH published: 13 July 2022 doi: 10.3389/fneur.2022.886477 Understanding and Predicting Cognitive Improvement of Young Adults in Ischemic Stroke Rehabilitation Therapy Helard Becerra Martinez 1*, Katryna Cisek 2 , Alejandro García-Rudolph 3,4,5 , John D. Kelleher 2 and Andrew Hines 1 1 School of Computer Science, University of College Dublin, Dublin, Ireland, 2 Information, Communication and Entertainment Research Institute, Technological University Dublin, Dublin, Ireland, 3 Institut Guttmann Hospital de Neurorehabilitacio, Badalona, Spain, 4 Universitat Autónoma de Barcelona, Cerdanyola del Vallés, Spain, 5 Fundació Institut d’Investigació en Ciéncies de la Salut Germans Trias i Pujol, Badalona, Spain Accurate early predictions of a patient’s likely cognitive improvement as a result of Edited by: Swathi Kiran, a stroke rehabilitation programme can assist clinicians in assembling more effective Boston University, United States therapeutic programs. In addition, sufficient levels of explainability, which can justify Reviewed by: these predictions, are a crucial requirement, as reported by clinicians. This article Ana Lúcia Faria, presents a machine learning (ML) prediction model targeting cognitive improvement after University of Madeira, Portugal Niamh C. Kennedy, therapy for stroke surviving patients. The prediction model relies on electronic health Ulster University, United Kingdom records from 201 ischemic stroke surviving patients containing demographic information, Elisabetta Farina, Fondazione Don Carlo Gnocchi Onlus cognitive assessments at admission from 24 different standardized neuropsychology (IRCCS), Italy tests (e.g., TMT, WAIS-III, Stroop, RAVLT, etc.), and therapy information collected during *Correspondence: rehabilitation (72,002 entries collected between March 2007 and September 2019). Helard Becerra Martinez The study population covered young-adult patients with a mean age of 49.51 years helard.becerra@ucd.ie and only 4.47% above 65 years of age at the stroke event (no age filter applied). Specialty section: Twenty different classification algorithms (from Python’s Scikit-learn library) are trained This article was submitted to and evaluated, varying their hyper-parameters and the number of features received as Neurorehabilitation, a section of the journal input. Best-performing models reported Recall scores around 0.7 and F1 scores of 0.6, Frontiers in Neurology showing the model’s ability to identify patients with poor cognitive improvement. The Received: 28 February 2022 study includes a detailed feature importance report that helps interpret the model’s inner Accepted: 13 June 2022 Published: 13 July 2022 decision workings and exposes the most influential factors in the cognitive improvement Citation: prediction. The study showed that certain therapy variables (e.g., the proportion of Martinez HB, Cisek K, memory and orientation executed tasks) had an important influence on the final prediction García-Rudolph A, Kelleher JD and of the cognitive improvement of patients at individual and population levels. This type of Hines A (2022) Understanding and Predicting Cognitive Improvement of evidence can serve clinicians in adjusting the therapeutic settings (e.g., type and load of Young Adults in Ischemic Stroke therapy activities) and selecting the one that maximizes cognitive improvement. Rehabilitation Therapy. Front. Neurol. 13:886477. Keywords: cognitive improvement, AI explainability, machine learning (ML), ischemic stroke, predictive models, doi: 10.3389/fneur.2022.886477 cognitive therapy, web-based therapy Frontiers in Neurology | www.frontiersin.org 1 July 2022 | Volume 13 | Article 886477 Martinez et al. Understanding Cognitive Improvement in Stroke Rehab 1. INTRODUCTION them rely on demographic and admission assessment variables to make predictions of a binary outcome (e.g., whether or not A stroke event occurs when the blood supply to part of the a patient achieves a specified score for a particular standardized brain is interrupted. There are two types of stroke depending scale), but only a few have explored the inclusion of therapy on the problem in the blood vessel supplying the brain: ischemic variables (22). Having reliable evidence of the impact that a stroke (blockage blood vessel) and hemorrhagic stroke (bleeding therapy configuration (e.g., type and load of therapy activities) blood vessel). Out of these two, ischemic stroke is reported for exerts on the cognitive rehabilitation outcome can help clinicians almost 87% of all stroke episodes (1), with an increased incidence plan more effective rehabilitation programs. Moreover, it’s been among adults aged 55 years and under in both the United observed that despite the number of models developed over the States and Europe (2, 3). Among developed countries, stroke is years, clinical adoption has been cautious because of the limited considered the leading cause of death with a mortality rate of capacity of models to explain their operations or outcomes and nearly 30% (4) and a disability rate of surviving patients close the ethical concerns this introduces regarding patient safety to 40% (5). For stroke survivors, the event can result in motor (25, 26). In this context, developing ML models with the and/or cognitive impairments (e.g., body weakness, disturbances capacity to produce actionable predictions and a sufficient level of cognitive functioning), and their outcome can vary from of explainability is crucial for clinical adoption, as pointed out by permanent disability to complete recovery (6, 7). Stinear et al., in their review study (27). In this context, cognitive rehabilitation aims to maximize This article presents a ML prediction model targeting the patients’ recovery and achieve an optimal level of cognitive cognitive improvement after therapy of surviving patients with functioning to promote their reintegration into normal activities stroke. The model relies on patients’ demographic characteristics, of daily living (8, 9). The clinical pathway of cognitive cognitive assessments at admission, and their corresponding rehabilitation starts often with a clinical interview and several therapy records. These records are integrated and pre-processed assessments to diagnose the patients’ cognitive condition. using standardized techniques commonly applied over data This information is then used to formulate a personalized science projects and following clinicians’ recommendations to rehabilitation plan for the patient. Performance information can avoid biased entries. Different models are trained and evaluated, be recorded during the patient’s therapy so the initial treatment varying the ML algorithm, their hyper-parameters and the plan can be adjusted. Although the rising of computerized number of features received as input. Then, predictions of the cognitive programs has shown good advantages in therapy best-performing model are analyzed at global and individual management and systematization (10, 11), a large part of the levels using explanation methods. Through the development rehabilitation process still relies on the clinician’s experience and and analysis of this model this article seeks 1) to enable better availability. Thus, planning the best rehabilitation scheme for a design of rehabilitation therapy by exploring the impact of large number of patients and closely monitoring their progress therapy configurations on patients’ cognitive improvement; 2) to can turn into a very demanding task for a clinician (12, 13). understand the underlying factors that improve/inhibit patients Due to their capacity to find patterns and relationships from improving on cognitive capacities after therapy using across data, machine learning (ML) models have shown feature importance reports at individual and global levels; and promising results in solving complex problems in different 3) to explore visualization reports that could explain models’ fields, including medical diagnosis and prognosis prediction outcomes and bolster clinicians’ trust. (14). In recent years, researchers have explored their integration The remainder of this document is divided as follows. Section into the stroke rehabilitation process, either as prediction 2 presents an overview of ML tools developed for stroke tools or as part of larger clinical decision support systems rehabilitation outcome prediction. In Section 3, the data and (CDSS) (15–17). These ML-based tools can provide treatment methods used in the study are described. Section 4 presents recommendations and predictions of relevant outcomes such as the data preparation, predictions, and model explanation results. the length of rehabilitation, the patient’s performance throughout Finally, Sections 5 and 6 present a detailed discussion and the therapy, and the patient’s cognitive and physical improvement overall conclusions of the study. after rehabilitation (18). For the specific case of cognitive improvement, having accurate predictions early in rehabilitation can help clinicians assemble a realistic therapy plan, adjust 2. RELATED STUDY it to obtain better results, or anticipate additional care after the therapy. Studies suggest that better rehabilitation outcomes can be Several studies have applied ML models to predict the achieved through 1) early rehabilitation and mobilization, 2) cognitive condition of stroke survivors after therapy. These higher-intensity therapy, and 3) personalized therapy plans (28, models often target scores from standardized scales such as the 29). However, finding the best therapy options for an individual cognitive portion of the Functional Independence Measure (c- patient can be a very challenging task, even for a well-experienced FIM), the Glasgow Outcome Scale (GOS), the National Institutes clinician. Unlike traditional expert systems, which are restricted of Health Stroke Scale (NIHSS), or the Disability Rating Scale to hardcore knowledge and a set of rules, ML-based tools (DRS) (19–22). Other studies have developed models to predict can learn solutions outside the expert knowledge and adapt outcomes for specific cognitive domains such as memory (risk themselves to each patient’s condition providing more accurate of dementia) and language (aphasia recovery) (23, 24). Most of and personalized solutions (30). In this context, ML models are Frontiers in Neurology | www.frontiersin.org 2 July 2022 | Volume 13 | Article 886477 Martinez et al. Understanding Cognitive Improvement in Stroke Rehab an appealing option to assist clinicians during the rehabilitation tasks, predicting optimal use of rehabilitation resources. process. The potential benefits of applying these new tools have The comparison and evaluation procedures involved robust drawn enormous attention to this field in recent years (27, 31). parameter tuning, varied resampling methods of the traditional Most of the literature focuses on precision and personalized variables, with and without rehabilitation configuration variables, medicine in the area of diagnosis rather than the prognosis of as well as model-dependent and model-independent ranking cognitive improvement; however, the following studies represent techniques to assess variables’ importance. This study highlights some recent advancements in research into this area. the importance of utilizing variables describing the cognitive The authors of the study by Sale et al. (19), assessed the rehabilitation configuration. It provides therapists and clinicians predictive value of common inflammatory biomarkers and with actionable predictions facilitating their decision-making other blood biomarkers in predicting cognitive and motor for the rehabilitation intensity and duration and more specific improvement upon completion of rehabilitation treatment. A interventions for patients. two-step procedure was implemented consisting of feature selection among the biochemical and hematological parameters based on mutual information criteria and the application of a 3. MATERIALS AND METHODS Support Vector Machine (SVM) classifier. This study confirmed previous reports of the prognostic value of biomarkers on motor The methodology followed in this study is presented in the function and cognitive performance in patients with post-stroke; workflow diagram in Figure 1. The diagram depicts three however, it did not establish a link between the biomarkers main phases with their associated tasks: data preparation, and the type, intensity, and duration of the rehabilitation modeling, and explanation. It also includes a field indicating program and the corresponding impact of the rehabilitation on the involvement of healthcare professionals across tasks at each cognitive improvement. phase. This section presents a brief description of the data and The automatic cognitive prognosis model by Serrà et al. methods used in this study. (32), implemented multiple methodologies, including decision tree learning, instance-based learning, probabilistic learning, 3.1. Data Sources and support vector machines, in order to build classifiers for This study relies on clinical data from ischemic stroke surviving cognitive improvement in the areas of attention, memory, patients admitted at the Acquired Brain Injury rehabilitation and executive functioning. This study identified predictors Department of Institut Guttmann (Barcelona, Spain). All such as patient age at the time of injury, the etiology, participants were anonymized and non-identifiable. For this and pre-treatment neuropsychological evaluation scores as study, no specific written consent was required from participants; essential for accurate cognitive prognosis. However, this cross- however, when admitted to Institut Guttmann’s rehabilitation sectional study focused on pre-treatment diagnosis data, center, all patients provided written informed consent to be without including cognitive rehabilitation treatment, discharge, included in research studies carried out by the Institut Guttmann. or compliance data that would aid therapists in creating The authors confirm that this study is compliant with the personalized rehabilitation programs. Helsinki Declaration of 1975, as revised in 2008, and it was In their research into the personalization of rehabilitation approved by the Ethics Committee of Clinical Research of programs, García-Rudolph et al., built various classifiers Guttmann Institut. based upon Knowledge Discovery in Databases (KDD) framework (33), focusing on pre-processing, patterns, and 3.1.1. Cognitive Records knowledge extraction. They implemented rehabilitation Patients that are admitted at the Institut Guttmann rehabilitation programs (cognitive computerized tasks) as sequences of center go through an interview and a set of neuropsychological sessions with pattern recognition utilizing methodologies assessments where demographic and clinical characteristics such as association rules, classification, clustering and shallow are entered into electronic health records from the hospital. neural models in order to assess patient rehabilitation task A qualified neuropsychologist is in charge of conducting execution patterns with positive and negative responses to the all neuropsychological assessments during admission. rehabilitation treatment (34). This framework allowed for a finer These standardized assessments aim to identify the level of representation of the design and configuration of successful cognitive impairment of the patient within different domains. rehabilitation programs for patients and the identification of Similarly, whenever a patient is to be discharged from the small variations within these configurations that can aid in rehabilitation center (i.e., the patient completed the therapy personalizing rehabilitation programs. at the center), the same set of assessments are conducted Finally, in the study by García-Rudolph (22), the authors and recorded to analyze the patient’s improvement. Table 1 evaluated and compared multiple predictive techniques and lists the cognitive assessments included in this study, the models to gain insight into the efficiency of rehabilitation cognitive domain they target, and the scale they use to report programs in terms of the largest gain in function with the their results. Cognitive assessments report their results on lowest duration of treatment. Traditional outcome prognosis different scales, as observed in the Scale column from Table 1. models with demographic and clinical variables were compared Except for TMT-A, TMT-B, and WCST Errors, cognitive to models augmented with additional variables describing assessments use scales where higher numbers are considered the execution and configuration of cognitive computerized good outcomes. Frontiers in Neurology | www.frontiersin.org 3 July 2022 | Volume 13 | Article 886477 Martinez et al. Understanding Cognitive Improvement in Stroke Rehab FIGURE 1 | Diagram of the methodology used in this study. 3.1.2. Therapy Records language, executive functioning, calculus, gnosias, and praxias. The implementation of computerized rehabilitation platforms After each task execution, the patient receives performance has gained increasing interest in recent years (10, 11). feedback ranging from 0 to 100 (0 being the lowest and 100 Studies have shown that these platforms can bring advantages the highest). The system organizes this information as temporal such as therapy monitoring and feedback for patients, easy entries and stores each task execution and its result with a set of access (e.g., home-based care settings), and cost benefits (10). relevant variables such as the date of execution and identifiers for For therapy purposes, Institut Gutmann uses the Guttmann, the patient, the therapist, and the executed task. Despite this type NeuroPersonalTrainer R (GNPT) (43) which is a rehabilitation of automation, therapists always have the capacity to select and platform for treatment systematization. The system contains 149 adjust the treatment whenever they consider it necessary. different cognitive rehabilitation web-based tasks assigned to the patient without a specific order. Therapy sessions are scheduled 3.2. Study Population 2–5 times per week during the entire therapy, lasting from This study included an initial cohort of 1,162 ischemic 2 to 6 months depending on the therapist’s recommendation. stroke patients admitted to the rehabilitation unit of the A single therapy session can take 45 min to 1 h, and it can Institut Guttmann (Barcelona, Spain), between March 2007 and include 4–10 cognitive rehabilitation tasks. Each task targets a September 2019. For this study, the inclusion criteria, applied particular cognitive domain like orientation, attention, memory, over the initial cohort, were based on the neuropsychologist’s Frontiers in Neurology | www.frontiersin.org 4 July 2022 | Volume 13 | Article 886477 Martinez et al. Understanding Cognitive Improvement in Stroke Rehab TABLE 1 | List of neuropsychological assessments administrated at the variables (around 26% of the current 475 patient entries). Guttmann Institut at admission and discharge. Removing all entries with missing assessments would have Assessment Cognitive domain Scale reduced the dataset to 81 patient entries and 77 accompanying variables. Although data imputation is usually applied to treat TB Personal Orientation (35) Orientation [0–7] missing values, this practice is not always recommended when TB Spatial Orientation (35) Orientation [0–5] dealing with clinical data (44, 45). Thus, to mitigate the TB Temporal Orientation (35) Orientation [0–23] impact of removing entries with missing values, the assessments Digits Span (36) Attention [0–9] with the lowest completeness levels were selected for removal TMT-A (37) Attention [0–Inf] (assessments with less than 30% of completeness rate). Eight Stroop-Words (38) Attention [0–Inf] variables were selected for removal: Stroop-words, Stroop-color, Stroop-Color (38) Attention [0–Inf] Stroop-words/color, TMT-B, WCST categories, WCST errors, Stroop-Words/Colors (38) Attention [0–Inf] Stroop-interference, and Visuospatial WAIS-III. This procedure TB Language Repetition (35) Language [0–10] resulted in a dataset containing 201 patient entries with 69 TB Language Denomination (35) Language [0–14] descriptive variables. Although the number of entries is reduced, TB Language Comprehension (35) Language [0–16] the approach aims to retain the data description capacity across Digit Span Backwards WAIS-III (36) Memory [0–8] all cognitive domains. Numbers and Letters WAIS-III (36) Memory [0–16] RAVLT Learning (39) Memory [0–75] 3.3. Feature Engineering RAVLT Free Recall (39) Memory [0–15] Through data aggregation, therapy variables like daily sessions and the number of tasks were constructed by summing the RAVLT Recognition (39) Memory [0–15] number of days and tasks registered in the GNPT therapy TMT-B (37) Executive Functions [0–Inf] records. Non executed tasks and non executed proportion are two WCST Categories (40) Executive Functions [0–6] variables that capture the number of tasks with performance WCST Errors (40) Executive Functions [0–Inf] scores of “0.” Obtaining a score of “0” after a task execution Stroop-Interference (38) Executive Functions [Inf] indicates that the patient could not fulfill the minimum required PMR (41) Executive Functions [0–Inf] by the rehabilitation task, which can be considered normal Visuospatial WAIS-III (36) Visual [0–Inf] during an early stage of the rehabilitation process. However, Images WAIS-III (36) Visual [0–20] from a neuropsychologist’s perspective, patients that accumulate Cubes WAIS-III (36) Visual [0–Inf] a large amount of “0” scores evidence difficulties completing (*) NIHSS (42) Overall impairment [0–42] their rehabilitation therapy. Therefore, variables that can capture TB, Test Barcelona; TMT, Trail Making Test; WAIS-III, Wechsler Adult Intelligence Scale this condition, such as non executed tasks and non executed 3rd version; RAVLT, Rey Auditory Verbal Learning Test; WCST, Wisconsin Card Sorting; proportion, might carry relevant information for cognitive NIHSS, National Institutes of Health Stroke Scale; (*), administrated only at admission. outcome and therapy compliance predictions. Admission compliance and discharge compliance are two variables derived from a neuropsychologist’s expertise. From input: 1) admission at the rehabilitation center during the first their analysis, it was observed that the level of completeness 6 months since the stroke event; 2) patients who went through of cognitive assessments varied from admission to discharge. therapy and had records in the GNPT platform, and 3) a Neuropsychologists perceive this as a common occurrence maximum length of therapy of 180 days. Applying these criteria since some patients tend to skip some assessments at admission resulted in the removal of 678 patients from the initial cohort because of their initial condition, especially when they are leaving 484 patients with 77 descriptive variables. admitted very close to the date of the stroke event. Once Clinicians recommended another criterion to avoid including they have completed their therapy and regained some of their misleading entries. They commented that, in some cases, patients cognitive capacities, they usually complete more assessments. who could not complete their cognitive assessments in the first On that account, the level of assessment compliance, i.e., trial had to be assessed again on the same day. This procedure the number of cognitive assessments patients complete would erroneously record a therapy duration of 0 days. They at admission, can be a relevant indicator of the patient’s recommended adding a minimum length of therapy (e.g., 14 cognitive capacity. days) to avoid including registries with inconsistent therapy The global improvement was calculated by evaluating the lengths. The minimum length of therapy was established at 14 difference between discharge and admission assessment scores days, resulting in 475 patients with 77 descriptive variables. and the summation of assigned markers. For instance, if Finally, both admission and discharge assessments were the score difference between discharge and admission is required to measure the improvement of a patient; however, positive, then the improvement of that assessment is “1,” if as it is common when dealing with real-world data, the level it is negative, then improvement is “–1,” and if there is no of completeness was different for most cognitive variables. This difference, the improvement is “0.” The global improvement posed a huge limitation because, in order to have a set of is obtained by summing the improvement markers from all patients with all cognitive assessments, the dataset needed to cognitive assessments and dividing the result by the number be shrunk to the lowest completeness percentage of cognitive of assessments. The same process is carried out to generate Frontiers in Neurology | www.frontiersin.org 5 July 2022 | Volume 13 | Article 886477 Martinez et al. Understanding Cognitive Improvement in Stroke Rehab improvement values for all cognitive domains (e.g., attention 3.5. Modeling improvement, memory improvement, etc.). For cognitive 3.5.1. Cognitive Outcome assessments like TMT-A, TMT-B, and WCST Errors, which For this study, the global improvement was considered as have an inverse score scale, a reverse procedure was applied to the target variable to predict the cognitive outcome. The calculate the improvement. global improvement combines improvement indicators of all Finally, gain proportion is introduced as a variable that tries standardized cognitive assessments administered to the patient, to capture the relative gain after the execution of each task. which can help describe the patient’s overall status after This variable was derived from the weighted summation of therapy. This variable can rely on demographic, cognitive, all task results based on their performance ranges. For this and therapy variables to act as predictor features. For this study, different markers are assigned to each execution: a “0” study, the global improvement was adapted to fit a common marker is assigned to tasks with results of “0,” a “0.5” marker binary classification problem. A threshold was used to separate is assigned to tasks with results falling in “1–64” and “86–100” patients with positive improvement scores and negative (or zero) ranges, finally, a marker “1” is assigned to tasks with results improvement scores. (Class “0”: global improvement<=0, Class in the range of “65–85.” The assignment of these markers is “1”: global improvement>0). From a binary classification based on the assumption that “65–85” is the optimal range of perspective, the positive class is referred to as the class of interest performance for task execution (46). The intuition behind the for a defined problem. In this study, the main focus was to definition of this optimal therapeutic range is that it is best to identify patients “at risk,” that is, patients reporting negative or avoid results that are too low (“1–64”) from tasks that are too zero global improvement values. Therefore, Class “0” was selected difficult to get responses from a damaged area of the brain; or as the positive class. extremely high results (“86–100”) from tasks that are too simple to exercise such damaged areas. During their therapy, a patient 3.5.2. Pre-processing accumulates different markers for each task execution. These A pre-processing step is recommended to standardize and format markers are then summed to obtain the total gain after therapy. the features before training the ML model. Although some The gain proportion results from dividing the total gain by the families of ML algorithms are able to handle features with number of tasks. Equivalent variables are calculated for each different scales and formats, pre-processing before training the cognitive domain (attention execution gain, memory execution model is recommended to guarantee better performance (49). It gain, etc.). is also desired for features to maintain a normal distribution so they can be applied to different ML algorithms. For this study, input variables were scaled using the StandardScaler method from the Scikit-learn Python library. This method removes 3.4. Data Exploration the mean of features (i.e., set it to zero) and scale values to In addition to measures like mean, variance, minimum, and unit variance. For variables with a skewed distribution, the maximum, which can help describe the variables individually, Scikit-learn implementation of the PowerTransformer method correlation and clustering analyses are applied over the (50) was used to treat features to obtain a more “Gaussian- rehabilitation dataset to explore the relevant relationships among like” distribution. Finally, nominal features were encoded and the available features. transformed into numeric features to allow for performing For this study, correlation heatmaps were used to understand mathematical operations. how the numeric variables relate to each other within the rehabilitation dataset. This type of analysis is a good starting point to identify multicollinearity scenarios where two or 3.5.3. Classification Algorithms more variables present a strong correlation. From a data Twenty different classification algorithms from nine different scientist’s standpoint, evidence from this analysis can serve to families were considered for this study: five Linear Models, reduce overlap in input variables by removing variables that one Nearest Neighbor, two Decision Trees, two Support Vector do not add prediction power to the model or to assist in Machines, two Naive Bayes, five Ensemble Methods, one mitigating sparse data by complementing missing information Gaussian Process, one Linear Discriminant Analysis, and one from variables. For the cluster exploration, the analysis was Boosted Trees. The implementation of these algorithms, available carried out using the Principal Component Analysis (PCA) at the scikit-learn library and the xgboost package for Python, were and the t-Distributed Stochastic Neighbourhood Embedding (t- used for training and evaluation. SNE) techniques (47, 48). The objective was to identify possible patterns, unseen structures, or possible anomalies across the 3.5.4. Performance Evaluation available demographic, cognitive, and therapy data. Both PCA The algorithms’ performance was measured using standard and t-SNE are dimensionality reduction techniques that ease data classification metrics like F1 score, Recall, Precision, and the visualization and allow the discovery of veiled insights from it. Area Under the Receiver Operator Characteristic Curve (ROC- They differ in the sense that PCA is linear and deterministic, and AUC). The Recall score was selected as the main evaluation it tries to retain the global structure of the data; meanwhile, t-SNE criteria as it prioritizes the correct positive instances that is non-linear and non-deterministic (randomized), and it tries to are correctly classified, i.e., patients with zero or negative preserve the local structure of data. improvement scores (class “0”). To cope with a low number Frontiers in Neurology | www.frontiersin.org 6 July 2022 | Volume 13 | Article 886477 Martinez et al. Understanding Cognitive Improvement in Stroke Rehab of entries in the dataset, a k-fold (k = 5) cross-validation re- TABLE 2 | Sets of hyper-parameters used for the grid-search corresponding to sampling method with 5 repetitions was used to evaluate the the pre-selected classification algorithms. performance of different classification algorithms. For this task, Algorithm Hyper-paramets set the algorithms were assessed using default Scikit-learn settings with no customized parameters. ExtraTreesClassifier n_estimators = [10, 100, 1,000] max_depth = [3, 7, 9] 3.5.5. Feature Selection min_samples_split=[2,10,20] Before the optimization, feature selection was carried out criterion=[“gini,” “entropy”] to identify the most important features using the model min_weight_fraction_leaf = [0,0.2,0.3,0.5] performance as the main criteria. Different from the selection RandomForestClassifier criterion=[“gini,” “entropy”] made using the clinicians and data scientists as a reference, n_estimators = [10, 100, 1,000] this procedure uses the relationship between input and target max_features = [“sqrt,” “log2”] variables to rank the contribution of each input feature. The max_depth = [9, 15] idea is to verify the models’ performance using different amounts min_samples_split=[2,10,20] of input features. Using fewer features while maintaining good min_weight_fraction_leaf = [0,0.2,0.5] performance is a desirable characteristic as it reduces the number KNeighborsClassifier n_neighbors = [2, 10, 21] of assessments required by the patient (51). This also aligns with weights = [“uniform,” “distance”] the data minimization principle from the European General Data metric = [“euclidean,” “manhattan,” “minkowski”] Protection Regulations (GDPR) (52). The feature selection step XGBClassifier eta = [0.001, 0.01, 0.1, 0.2, 0.3] was carried out using the SelectKBest method implemented in gamma = [0.05, 0.5, 1, 1.5] Scikit-learn. For a classification problem, this method uses an min_child_weight = [5, 7, 9, 10] implementation of the ANOVA f-test to rank features and select subsample = [0.5, 0.8, 1] the ones with higher scores. For this case study, the entire set colsample_bytree = [0.6, 0.8, 1] of features, plus the best 20 and 10 features, were selected and lambda_par = [0.1, 0.5, 1] included in the hyper-parameter optimization setting. LogisticRegression solver = [“newton-cg,” “lbfgs,” “liblinear”] 3.5.6. Hyper-Parameters Optimization penalty = [“L2”] To find the best performing parameters, a grid-search was C = [100, 10, 1.0, 0.1, 0.01] carried out using the GridSearchCV method implemented in max_iter = [1,000] Scikit-learn along with a k-fold (k = 5) cross-validation method. Table 2 presents the sets of values selected to carry out the grid-search for each parameter and pre-selected algorithm. For the ExtraTreesClassifier and RandomForestClassifier, allows interpreting each feature’s average contribution to the similar parameters were set for optimization; meanwhile, the model’s prediction. For this study, the implementation of the KNeighborsClassifier, XGBClassifier, and LogisticRegression had SHAP library for Python was used to examine the trained models. specific tuning parameters. The models were optimized based on their Recall performance. Then, using the best parameters from the Recall optimization, F1, Precision, and ROC-AUC scores 4. RESULTS were evaluated via k-fold (k = 5) cross-validation. 4.1. Experiment Data Table 3 presents basic statistics of all demographic, cognitive, 3.6. Explanation Methods and therapy variables from the rehabilitation dataset to help In addition to the classification metrics that report the understand the characteristics of the available data. The TMT- performance of the model, explanatory material that allows A and TMT-B assessments showed higher standard deviation clinicians to interpret the inner workings of the model is needed values (48.43 and 82.51); meanwhile, TB Personal and TB (53–55). From several explanation classes, clinicians identified Language Repetition assessments reported the lowest variability that having a report of the most influential features on a model’s among assessments (0.00). For therapy variables, the total outcome was critical, as it allowed them to compare the model’s number of tasks showed a SD value of 85.22, which evidence the decision process with their clinical judgment process (53). The load of therapy differences among patients. The same variable SHapley Additive exPlanations (SHAP) method (56) was used reported a maximum value of 480, meaning that a single patient to analyze the trained models and describe the model’s inner executed this amount of rehabilitation tasks. Table 4 presents the rationale. SHAP is a unified framework based on six feature task performance (grouped by ranges) given by the GNPT system importance methods that facilitate the interpretation of ML after each task execution. Executed tasks with a performance models’ predictions. The underlying logic behind SHAP’s inner result of “0” represent 16.2% of all executed tasks. Results falling workings is based on evaluating the trained model using different in the “65–85” range represent 18.9% of the executed tasks. sets of feature permutations and calculating each feature’s average Within the GNPT platform, this range is considered optimal contribution to the model’s output; this average value is called since it balances the cognitive gain of the patient (46). The the SHAP value. Computing SHAP values for each input feature distribution of executed tasks shows that memory tasks are the Frontiers in Neurology | www.frontiersin.org 7 July 2022 | Volume 13 | Article 886477 Martinez et al. Understanding Cognitive Improvement in Stroke Rehab TABLE 3 | Basic statistics of demographic, cognitive and therapy variables from the rehabilitation dataset. Group Variable Mean SD Min Max Demographic Age 49.70 10.18 16.74 81.92 Age at injury (mc) 49.52 10.18 16.65 81.84 Time since injury in days 66.46 40.30 1.60 173.43 Length of therapy 65.35 30.79 15.00 173.00 Sex (c) Male N/A N/A N/A N/A Female N/A N/A N/A N/A Marital status (c) Married N/A N/A N/A N/A Single N/A N/A N/A N/A Divorce N/A N/A N/A N/A Separate N/A N/A N/A N/A Widow N/A N/A N/A N/A Cognitive Admission compliance 0.90 0.11 0.67 1.00 Discharge compliance (mc) 0.94 0.09 0.67 1.00 Global improvement (t) 0.18 0.21 -0.50 0.75 Attention improvement (mc) 0.02 0.17 -0.33 0.67 Orientation improvement (mc) 0.27 0.40 -0.80 1.00 Language improvement (mc) 0.04 0.15 -0.33 0.67 Memory improvement (mc) 0.27 0.44 -1.00 1.00 Ex. Functions improvement (mc) 0.23 0.39 -1.00 1.00 Visual improvement 0.31 0.41 -0.67 1.00 NIHSS 9.86 4.65 1.00 22.00 Variable Mean SD Min Max Adm Dis Adm Dis Adm Dis Adm Dis Orientation TB Personal Orientation 6.99 7.00 0.12 0.00 6 7 7 7 TB Spatial Orientation 4.97 4.99 0.17 0.10 4 4 5 5 TB Temporal Orientation 22.59 22.71 1.41 1.22 12 11 23 23 Attention Digits Span 5.91 6.03 1.10 1.07 3 4 9 9 TMT-A 67.44 53.42 48.43 31.70 4 6 289 240 Stroop - Words (md) 78.43 82.05 16.37 15.51 37 40 123 125 Stroop - Color (md) 55.51 57.70 12.45 12.54 23 27 89 90 Stroop - Words/Colors (md) 32.04 34.03 10.71 11.04 6 8 85 73 Language TB Language Repetition 9.99 10.00 0.10 0.00 9 10 10 10 TB Language Denomination 13.96 14.00 0.27 0.00 11 14 14 14 TB Language Comprehension 15.78 15.91 0.76 0.50 9 12 16 16 Memory Digit Span Backwards WAIS-III 4.19 4.35 0.98 0.98 2 2 7 8 Numbers and Letters WAIS-III 8.14 8.72 2.61 2.52 1 3 14 15 RAVLT Learning 42.14 46.23 10.66 11.49 21 9 70 70 RAVLT Free Recall 8.26 9.35 3.51 3.41 0 0 15 15 RAVLT Recognition 11.44 12.19 3.91 3.20 0 1 15 15 Executive Functions TMT-B (md) 141.13 112.07 82.51 48.63 30 30 565 300 WCST Categories (md) 4.11 4.24 2.11 2.13 0 0 6 6 WCST Errors (md) 18.36 16.25 15.07 14.64 0 0 63 72 Stroop - Interference (md) -0.21 0.41 7.20 7.48 -21 -22 35 25 PMR 31.93 35.10 13.25 13.30 3 5 72 84 (Continued) Frontiers in Neurology | www.frontiersin.org 8 July 2022 | Volume 13 | Article 886477 Martinez et al. Understanding Cognitive Improvement in Stroke Rehab TABLE 3 | Continued Group Variable Mean SD Min Max Visual Visuospatial WAIS-III (md) 41.21 46.23 15.59 16.04 10 13 92 92 Images WAIS-III 19.30 19.67 1.49 0.94 11 14 20 20 Cubes WAIS-III 26.96 30.31 11.97 11.94 2 6 66 66 Variable Mean SD Min Max Therapy Daily sessions 12.15 7.91 1.00 52.00 Total number of tasks (mc) 111.89 85.22 2.00 480.00 Total non executed tasks (mc) 15.84 18.93 0.00 117.00 Non executed proportion 0.14 0.10 0.00 0.53 Total gain proportion (mc) 0.53 0.08 0.25 0.75 Attention Number of tasks (mc) 20.93 18.36 1.00 123.00 Task proportion 0.19 0.10 0.01 0.67 Non executed tasks 2.01 3.85 0.00 40.00 Execution gain 11.25 10.22 0.00 72.00 Memory Number of tasks (mc) 45.65 40.40 2.00 294.00 Task proportion 0.40 0.15 0.07 1.00 Non executed tasks 5.22 8.08 0.00 53.00 Execution gain 25.57 23.53 0.50 181.00 Ex. Functions Number of tasks (mc) 38.16 30.58 1.00 170.00 Task proportion 0.35 0.13 0.05 0.75 Non executed tasks 8.12 9.13 0.00 62.00 Execution gain 18.08 15.17 0.00 96.50 Language Number of tasks (mc) 8.86 13.35 1.00 55.00 Task proportion 0.16 0.28 0.01 1.00 Non executed tasks 0.71 1.10 0.00 3.00 Execution gain 5.04 9.10 0.50 37.00 Orientation Number of tasks (mc) 4.31 5.27 1.00 31.00 Task proportion 0.03 0.04 0.00 0.20 Non executed tasks 0.39 1.30 0.00 10.00 Execution gain 2.06 2.58 0.00 15.00 Number of tasks (mc) 11.81 11.45 1.00 62.00 Calculus Task proportion 0.09 0.07 0.01 0.31 Non executed tasks 1.42 2.32 0.00 11.00 Execution gain 6.25 6.29 0.00 36.50 Number of tasks (mc) 8.50 14.68 1.00 81.00 Gnosias Task proportion 0.06 0.09 0.01 0.49 Non executed tasks 0.47 1.47 0.00 8.00 Execution gain 4.81 8.11 0.00 40.00 Praxias Number of tasks (mc) 3.71 2.99 2.00 12.00 Task proportion 0.02 0.01 0.01 0.07 Non executed tasks 0.24 0.64 0.00 2.00 Execution gain 1.82 1.74 0.00 6.50 N = 201; Adm, admission; Dis, discharge; c, categorical variable; mc, removed to prevent multicollinearity issues; md, removed to prevent missing data issues; t, target variable. most repeated, as they represent 40.2% of all performed tasks. visualization purposes, correlation outcomes are presented in Meanwhile, attention and executive functions cover 19.5% and separated plots: Figure 2 (demographic + cognitive) and Figure 3 29.4% of all entries. (demographic + therapy). Color bars were used to a show strong positive correlation (red) and a strong negative correlation (blue). 4.2. Exploratory Analysis Figure 2 (demographic + cognitive) shows several clusters For this study, two correlation heatmaps were generated of low-to-moderate positive correlation coefficients (0.2–0.6), using the demographic, cognitive, and therapy variables. For corresponding to standardized cognitive assessment variables. Frontiers in Neurology | www.frontiersin.org 9 July 2022 | Volume 13 | Article 886477 Martinez et al. Understanding Cognitive Improvement in Stroke Rehab Smaller clusters of strong correlation coefficients (0.7–0.8) TABLE 4 | Summary report from temporal records of the GNPT platform. belonging to variables from the same group of standardized Variable n = 72,002 assessments are observed near the diagonal. This trait was also observed for variables admission compliance and discharge Task performance by result ranges, n (%) compliance. These results show that some assessments carry very [0] 11,669 (16.2%) similar information, especially when they belong to the same [1–64] 23,501 (32.6%) cognitive domain. Other variables like age, time since injury, [65–85] 13,573 (18.9%) NIHSS, and global improvement reported low-to-moderate [86–100] 23,259 (32.3%) negative correlation coefficients (−0.2 to −0.6). For some Number of tasks per cognitive function, n (%) variables, a strong correlation is expected, and it can be easily Attention 14,015 (19.5%) explained like for the pair of features age and age at injury. In a Memory 28,963 (40.2%) stroke rehabilitation context, it is expected that patients start their Ex. Functions 21,172 (29.4%) therapy within the first 6 months of a stroke occurring; thus, it is Language 1,923 (2.7%) likely that these two variables have close values. On that account, Orientation 979 (1.4%) including only one of these variables in the descriptor set might Calculus 2,377 (3.3%) be sufficient for modeling purposes. Gnosias 2,379 (3.3%) Figure 3 (demographic + therapy) shows a more sparse Praxias 158 (0.2%) distribution of correlation outcomes. For this analysis, strong correlation clusters are more abundant, and they are related, in most cases, to three variables: daily sessions, number of tasks, and non executed tasks. The heatmap shows that these The PCA analysis presented in Figure 4A reported an clusters are formed across each cognitive domain, and that they accumulated explained variance of 0.42 for the two components are stronger among the ones with a higher number of tasks (PCA1 and PCA2). Figure 4A shows four clusters plus two executions (attention, memory and executive functions). The smaller sets of data points slightly separated from two of pair of variables daily sessions and number of tasks reported the main clusters. For the t-SNE method, different perplexity strong correlation coefficients. Since they both depend on the and rate parameters were tested at different runs of the length of the therapy (more therapy sessions will result in more method. The scatter plot in Figure 4B shows the results from executed tasks), a strong correlation between these two variables setting the perplexity parameter at 10 and the learning rate is expected. Again, keeping only one of these variables for at 200. The scatter plot reported three main clusters plus a modeling purposes seems to be a valid suggestion. few smaller sets of data points. Both methods were able to Including groups of highly correlated variables as input to produce relatively dense clusters and separate classes at some train an ML model might not affect the model’s performance, level. Yet, the formation of these clusters does not seem to but it can reduce the capacity to interpret the effect of variables be linked to the level of stroke severity (NIHSS score) or the on the model’s outcome. Thus, to develop explainable ML tools, outcome of patients (global improvement), as no patterns are data scientists recommend handling multicollinearity instances perceived among the shapes and colors of the data points. These before training the model (57). Based on the findings from the findings highlight that both severity and improvement labels correlation analysis, a few variables were selected for removal were not strongly correlated with the information carried by the to mitigate possible overlapping effects. Thirteen variables were input features. removed from the dataset: age at injury, discharge compliance, number of tasks, non executed tasks, gain proportion, attention 4.3. Cognitive Improvement Prediction tasks, memory tasks, F. executive tasks, language tasks, orientation For this study, the global improvement was adapted to fit a tasks, calculus tasks, gnosias tasks, and praxias tasks. After this common binary classification problem. Figure 5 shows the procedure, the number of available variables is 56 (54 numeric distribution of the improvement scores and the threshold and 2 nominal). used to separate patients with positive improvement Following the exploratory stage, a cluster analysis was carried scores and negative (or zero) improvement scores (Class out using the Principal Component Analysis (PCA) and the “0”: global improvement<=0, Class “1”: global improvement>0). t-Distributed Stochastic Neighbourhood Embedding (t-SNE) The histogram shows that there were fewer cases with methods (47, 48). The objective was to identify possible patterns, negative or no improvement scores (58 out of 201 unseen structures, or possible anomalies across the available registries), which evidence the imbalanced distribution of demographic, cognitive, and therapy data. Figure 4 presents 2D the target. visualizations generated using the PCA and t-SNE methods. Twenty different ML classification algorithms were trained Each data point, representing a patient, was labeled according and evaluated using a k-fold (k = 5) cross-validation re- to the severity of the condition of the patient (from the NIHSS sampling method with 5 repetitions. Table 5 list the ten score) and the improvement after the therapy (from the global best-performing algorithms and their corresponding results for improvement variable). This type of visualization can help display Recall, F1, Precision, and AUC-ROC scores. The algorithms homogeneous groups of patients and verify if predictor variables reported Recall scores around 0.70; meanwhile, the reported correlate with the target labels of the model. F1 scores reached 0.63. Based on these results, five algorithms Frontiers in Neurology | www.frontiersin.org 10 July 2022 | Volume 13 | Article 886477 Martinez et al. Understanding Cognitive Improvement in Stroke Rehab FIGURE 2 | Correlation outcomes from demographic + cognitive variables. Red: strong positive correlation (+1), blue: strong negative correlation (–1). were pre-selected to be considered for hyper-parameter input features; meanwhile, the performance results from the optimization: RandomForestClassifier, ExtraTreesClassifier, Recall, F1, Precision, and AUC-ROC scores are presented KNeighborsClassifier, XGBClassifier, and LogisticRegression. in Table 7. These results showed no significant improvement All five pre-selected algorithms were trained using three for any algorithm compared to the results presented in different sets of input features: “All”: no feature selection, Table 5. Recall and F1 scores floated around 0.7 and 0.6, “20”: best 20 features, and “10”: best 10 features. Table 6 respectively, for most classification algorithms. In addition, no presents the optimized values for each parameter for all considerable performance differences were reported regarding five classification algorithms and the corresponding sets of the number of features used to train the models. These Frontiers in Neurology | www.frontiersin.org 11 July 2022 | Volume 13 | Article 886477 Martinez et al. Understanding Cognitive Improvement in Stroke Rehab FIGURE 3 | Correlation outcomes from demographic + therapy variables. Red: strong positive correlation (+1), blue: strong negative correlation (–1). results imply that it is possible to train a model with fewer behavior and feature relevance across the entire dataset and an features at a low cost in performance. Despite the room individual interpretation exploring the cases of four particular for performance improvement, these results show the model’s subjects from the dataset. The analysis was made using the best ability to identify patients with poor cognitive improvement performing hyper-parameters found during the optimization after therapy. process (Table 6) and the entire set of features since no significant difference was observed among the input sets of features. 4.4. Explanation Reports Due to the absence of an outstanding algorithm in terms of For this study, the feature importance analysis was organized performance, the XGBClassifier_All model was selected for the at two levels: a general interpretation covering the model’s explanatory analysis. Frontiers in Neurology | www.frontiersin.org 12 July 2022 | Volume 13 | Article 886477 Martinez et al. Understanding Cognitive Improvement in Stroke Rehab FIGURE 4 | Cluster analysis using PCA and t-SNE over demographic, cognitive, and therapy variables. Data points labeled according to severity (NIHSS score) and improvement (global improvement). PCA: accumulated explained variance 0.42 (A). t-SNE: perplexity 10, learning rate 200 (B). 4.4.1. General Interpretation Interpretation at a general level seeks to provide information for patient populations and the key parameters guiding the model’s outcome. Through this type of analysis, relevant trends of outcomes and the features that influenced them can be identified. The SHAP method assigns different values to each feature based on their contribution to the model’s outcome (positive or negative); these values are denoted as SHAP values. For the general interpretation analysis, the mean absolute SHAP values (for each feature) are computed over all entries in the dataset, and they are presented in Figure 6. The features were sorted based on their absolute impact on the model’s outcome, and they are presented over two cohorts: sex (male and female) and age (below and above 50 years old). This visualization reveals if different features have a different effect depending on the sex of the patient or their age group. Figures 6A,B show that the time since injury was the most determining feature for predicting the global improvement of a patient. This feature seemed to have a slightly stronger effect on men under 50 years of age. Admission compliance was identified as the second most important feature with no significant difference in terms of sex or age group. Out of the FIGURE 5 | Scores distribution of global improvement. Class group of standardized cognitive assessments, TMT-A (attention), “0”: global improvement<=0, Class “1”: global improvement>0. Cubes (visual), and Ravlt learning (memory) were among the most important features for the global improvement prediction. From this group, only Ravlt learning exhibits a noticeable effect across sex and age, orientation task proportion was the only difference within the sex cohort. As for therapy features, memory variable showing a clear difference over these cohorts. task proportion and orientation task proportion were the most The isolated effect of features on the model’s prediction was important features for the model prediction. With respect to its represented using the dependence scatter plots presented in Frontiers in Neurology | www.frontiersin.org 13 July 2022 | Volume 13 | Article 886477 Martinez et al. Understanding Cognitive Improvement in Stroke Rehab Figure 7. These plots show the normalized range of values a the first row; meanwhile, memory gain, memory non executed, feature can take vs. the SHAP value assigned to each instance and memory task proportion are presented in the second row. in the dataset. The plots include a histogram at the bottom Being memory the cognitive domain with the highest proportion showing the distribution of the feature values. For this case study, of assigned tasks (40.2%), the three features related to this negative SHAP values contribute to the patient being classified cognitive domain were selected for analysis. as class “0” (no cognitive improvement). Meanwhile, positive For the time since injury, the dependence plot in Figure 7A SHAP values push the model toward classifying the patient as shows that the trained model considered that higher values for class “1” (cognitive improvement). The dependence scatter plots the feature corresponded to lower SHAP values. Therefore, the presented in Figure 7 show how the trained model reacts to model interprets that delays in the start of the rehabilitation different feature instances and identifies what values the model (high time since injury values) increase the chances of a patient considers relevant for the final decision. being placed in the no improvement class (class “0”). For the Dependence scatter plots were generated for six input features, length of therapy presented in Figure 7B, higher SHAP values are as observed in Figure 7. Scatter plots for the time since injury, assigned to therapy lengths up to a certain point (approximately length of therapy, and non executed proportion are presented in 1 on the normalized x-axis), after this point, SHAP values start decreasing. For this feature, it looks like the model identified an optimal length of therapy which pushes the decision toward a TABLE 5 | Evaluation of the classification algorithms without hyper-parameter positive cognitive outcome (class “1”). The plot for non executed tuning. proportion in Figure 7C shows that higher SHAP values were Algorithm F1 Recall Precision AUC given to feature instances falling in the [–1, 1] range on the x- axis. In this case, the model seemed to penalize instances where RandomForestClassifier 0.638 (0.08) 0.701 (0.05) 0.652 (0.08) 0.526 (0.05) the proportion of non executed tasks was too low or too high. ExtraTreesClassifier 0.630 (0.08) 0.697 (0.07) 0.640 (0.11) 0.528 (0.05) Dependence plots generated for the memory features showed KNeighborsClassifier 0.614 (0.10) 0.671 (0.08) 0.606 (0.13) 0.512 (0.06) similarities with the previous cases. For memory gain, the plot XGBClassifier 0.636 (0.08) 0.664 (0.07) 0.642 (0.08) 0.543 (0.07) presented in Figure 7D shows that the model penalized extreme LogisticRegression 0.643 (0.06) 0.658 (0.06) 0.646 (0.08) 0.549 (0.08) feature instances as in Figure 7C. The model assumes that RidgeClassifier 0.633 (0.05) 0.651 (0.05) 0.637 (0.07) 0.534 (0.07) memory gain values (i.e., memory task executions falling in the BaggingClassifier 0.610 (0.08) 0.642 (0.06) 0.632 (0.06) 0.533 (0.07) optimal result range) increased the chances of a patient being LinearSVC 0.634 (0.06) 0.641 (0.07) 0.639 (0.07) 0.544 (0.07) classified as having cognitive improvement only at a defined LinearDiscriminant 0.614 (0.06) 0.625 (0.06) 0.616 (0.07) 0.515 (0.07) range of values (between 0 and 1 in the normalized x-axis). The Analysis memory non executed feature, presented in Figure 7E, shows the BernoulliNB 0.611 (0.06) 0.619 (0.07) 0.626 (0.07) 0.516 (0.07) same behavior as the time since injury feature. For this case, the model interprets that patients with low therapy compliance (i.e., Results sorted by the mean Recall scores accompanied by their corresponding SD. Re- sampling at k-fold (k = 5) cross-validation with 5 repetitions. The bold values indicated to large amounts of task executions with “0” results) are more likely highlight the recall columns. to be classified as having no cognitive improvement. For memory TABLE 6 | Best performing sets of hyper-parameters gathered during the grid-search optimization process. Algorithm Hyper-parameters ExtraTreesClassifier_All criterion: gini, max_depth: 9, min_samples_split: 2, min_weight_fraction_leaf: 0, n_estimators: 1,000 ExtraTreesClassifier_20 criterion: gini, max_depth: 3, min_samples_split: 2, min_weight_fraction_leaf: 0.2, n_estimators: 10 ExtraTreesClassifier_10 criterion: gini, max_depth: 3, min_samples_split: 2, min_weight_fraction_leaf: 0.2, n_estimators: 10 RandomForestClassifier_All criterion: entropy, max_depth: 15, max_features: log2, min_samples_split: 10, min_weight_fraction_leaf: 0, n_estimators: 1,000 RandomForestClassifier_20 criterion: entropy, max_depth: 15, max_features: sqrt, min_samples_split: 20, min_weight_fraction_leaf: 0.2, n_estimators: 10 RandomForestClassifier_10 criterion: gini, max_depth: 15, max_features: log2, min_samples_split: 2, min_weight_fraction_leaf: 0.2, n_estimators: 10 KNeighborsClassifier_All metric: euclidean, n_neighbors: 17, weights: distance KNeighborsClassifier_20 metric: manhattan, n_neighbors: 19, weights: uniform KNeighborsClassifier_10 metric: euclidean, n_neighbors: 19, weights: uniform XGBClassifier_All colsample_bytree: 0.6, eta: 0.01, gamma: 1, min_child_weight: 5, reg_lambda: 0.5, subsample: 0.8 XGBClassifier_20 colsample_bytree: 0.8, eta: 0.1, gamma: 0.5, min_child_weight: 9, reg_lambda: 0.1, subsample: 0.5 XGBClassifier_10 colsample_bytree: 0.6, eta: 0.001, gamma: 0.05, min_child_weight: 5, reg_lambda: 0.1, subsample: 0.5 LogisticRegression_All C: 0.01, max_iter: 300, penalty: l2, solver: newton-cg LogisticRegression_20 C: 0.01, max_iter: 300, penalty: l2, solver: newton-cg LogisticRegression_10 C: 0.01, max_iter: 300, penalty: l2, solver: newton-cg Recall scores as evaluation criteria. Results reported for the different number of input features. “All”: no feature selection, “20”: best 20 features, “10”: best 10 features. Re-sampling at k-fold (k = 5) cross-validation with five repetitions. Frontiers in Neurology | www.frontiersin.org 14 July 2022 | Volume 13 | Article 886477 Martinez et al. Understanding Cognitive Improvement in Stroke Rehab task proportion, the dependence scatter plot showed three clusters TABLE 7 | Evaluation of the pre-selected algorithms with hyper-parameters corresponding to different ranges of the feature values. Although tuning. the model interprets that patients with high memory task Algorithm F1 Recall Precision AUC proportion values have more chances of being classified as an improvement case; mid values seemed to push the patient toward ExtraTreesClassifier_All 0.637 (0.09) 0.714 (0.06) 0.640 (0.15) 0.536 (0.05) a no improvement outcome, even more than small values for the ExtraTreesClassifier_20 0.593 (0.08) 0.711 (0.06) 0.510 (0.08) 0.500 (0.00) same feature. ExtraTreesClassifier_10 0.593 (0.08) 0.711 (0.06) 0.510 (0.08) 0.500 (0.00) RandomForestClassifier_All 0.612 (0.08) 0.713 (0.06) 0.610 (0.13) 0.512 (0.03) 4.4.2. Individual Interpretation RandomForestClassifier_20 0.597 (0.08) 0.713 (0.06) 0.515 (0.09) 0.502 (0.01) The AI model’s predictions need to be presented to clinicians with RandomForestClassifier_10 0.593 (0.08) 0.714 (0.06) 0.508 (0.08) 0.500 (0.01) enough information to contextualize and justify their outcomes. KNeighborsClassifier_All 0.606 (0.07) 0.714 (0.06) 0.596 (0.13) 0.508 (0.02) At an individual level, a report of feature importance allows KNeighborsClassifier_20 0.624 (0.08) 0.708 (0.06) 0.654 (0.12) 0.522 (0.04) clinicians to compare the rationale of the model and their own KNeighborsClassifier_10 0.609 (0.09) 0.692 (0.07) 0.591 (0.13) 0.507 (0.04) medical criteria, especially when they do not align together (53). XGBClassifier_All 0.617 (0.09) 0.713 (0.07) 0.624 (0.16) 0.517 (0.04) The individual interpretation analysis was focused on specific XGBClassifier_20 0.608 (0.10) 0.716 (0.07) 0.556 (0.15) 0.516 (0.04) patients and their particular conditions to observe the model’s XGBClassifier_10 0.593 (0.08) 0.712 (0.06) 0.510 (0.08) 0.500 (0.00) behavior at an individual level. For this analysis, four cases were LogisticRegression_All 0.607 (0.07) 0.708 (0.06) 0.595 (0.13) 0.507 (0.03) selected to present the model’s inner decision workings: two LogisticRegression_20 0.607 (0.07) 0.709 (0.06) 0.595 (0.13) 0.507 (0.03) patients reporting the highest global improvement scores and two LogisticRegression_10 0.604 (0.08) 0.712 (0.06) 0.567 (0.13) 0.507 (0.02) reporting the lowest scores. Recall scores as evaluation criteria. Results reported for the different number of input The individual analysis was made using waterfall plots features. “All”: no feature selection, “20”: best 20 features, “10”: best 10 features. Re- generated using the SHAP method; such plots are presented in sampling at k-fold (k = 5) cross-validation with 5 repetitions The bold values indicated to Figures 8, 9. The plots show how each feature affects positively highlight the recall columns. (red row) or negatively (blue row) the final decision of the trained model in each individual case. The plot should be read from the bottom to the top as it starts with the initially expected outcome (E[f (x)]) on the x-axis, which changes according to considering that the original record shows a feature value of “0.” the effect of different features until reaching the final outcome Based on this observation, it appears that the model lowered (f (x)). SHAP reports the model’s outcome in log-odds units the chances of patients with specific amounts of orientation task (before the logistic link function). So, positive outcomes in executions. Finally, features like Cubes, TMT-A, and admission the x-axis should be interpreted as probabilities above 0.5 compliance positively affected both cases. of a patient having a positive cognitive improvement (class Two other patients reporting the lowest global improvement “1”). scores were selected for the individual analysis of the model. The two patients with the highest global improvement scores From this pair of patients, only one of them was correctly were picked out and they were identified using their entry indexes classified as a no improvement case (class “0”). For Patient_27, in the dataset. These two patients, identified as Patient_123 and the SHAP analysis reported a positive outcome of 0.118, and it Patient_133, were correctly classified by the XGBClassifier_All was erroneously considered an improvement case. Meanwhile, model as having a positive cognitive improvement (class “1”). Patient_33 reported a negative outcome of –0.182, and it was The positive outcomes for Patient_123 and Patient_133 from correctly classified as a no improvement case. Despite the the SHAP analysis (0.341 and 1.061, respectively) are presented erroneous classification of Patient_27, it is noticed that the in Figure 8. Although Patient_123 was classified as a cognitive reported outcome (0.118) was very close to reaching a negative improvement case (positive outcome of 0.341), the waterfall value which would have resulted in the correct classification of plot shows that two features negatively affected the model’s the patient. outcome (time since injury and orientation task proportion) For these two cases, the prediction paths are presented in pushing the model toward a no improvement prediction (class Figure 9. From the group of features that reported a strong “0”). contribution to the model’s outcome, the cognitive assessment Figure 8 shows that the time since injury had a different Cubes WAIS-III was the only one showing a completely inverse impact on both patients. This can be explained by observing the effect for both patients. As for the time since injury, although values corresponding to that feature. Patient_133 reported a time the effect was negative for both cases, the difference in the since injury value of –0.556 (39 days in the original record) and magnitude of the impact seemed to have been decisive for the an effect of +0.14; meanwhile, Patient_123 reported a value of model’s outcome. For this feature, a difference in the start of the 1.55 (138 days in the original record) with an effect of –0.23. This rehabilitation (105 days for Patient_27 and 155 for Patient_33) analysis indicates that the model will lower the patient’s chances resulted in almost 0.12, which is close to the 0.118 reported in of cognitive improvement if the time since injury is too high. the SHAP analysis. As observed in the previous cases, the model Regarding the orientation task proportion effect, it is rather odd considers the time since injury as a determinant factor for the to observe a positive effect on the outcome of Patient_133 (+0.02) cognitive outcome of a patient. Frontiers in Neurology | www.frontiersin.org 15 July 2022 | Volume 13 | Article 886477 Martinez et al. Understanding Cognitive Improvement in Stroke Rehab FIGURE 6 | Global feature importance plot based on mean SHAP values for sex (A) and age (B) cohorts of patients. Results from the XGBClassifier_All optimized model. 5. DISCUSSION clinicians was key to building and discovering new variables, exploring the relationship among the feature set, identifying This study presented an ML model for cognitive improvement inconsistencies in the dataset, applying appropriate filters, and prediction of patients with ischemic stroke after therapy. To establishing strategies to deal with missing data entries. do so, several classifier algorithms were evaluated using varying In this phase, hand-crafted features derived from the therapy input features from demographic, cognitive, and therapy health records were built using common aggregation functions like records. The study uses feature importance reports to analyze count, sum, or mean. Yet, other key variables that were not predictions from the best-performing model at general and implicitly represented in these records were revealed only individual levels. The analysis presented in this study helped to through the input of clinical experts. The analysis then showed understand the underlying effect of different variables on the that these features (e.g., memory task proportion, orientation patients’ cognitive outcomes. Moreover, it was possible to identify task proportion) were strong indicators of the patient’s cognitive contributing therapeutic variables that resulted in a favorable outcome. In this study, we observed how a close interaction outcome. Throughout the development process, the article brings with domain experts leads to discovering important descriptive attention to the role of data scientists and clinicians and the value features. Although we relied on a specific platform (GNPT) to of their input at each stage of the development process. In this collect therapy information and build these descriptive features, section, reflections on each step of the process are presented, and the same approach can be adapted over similar computerized a discussion is presented on how some of the findings of this study can be generalized and applied across other healthcare- settings to derive similar features. related projects. This study relied on correlation and clustering analyses to examine the relationship among the dataset features. Similar 5.1. On the Data Exploration Process to what clinicians reported regarding model explainability (53), The development process presented in this study begins with well-designed visualization tools are essential to facilitate the the data exploration phase. At this phase, the involvement of understanding and interpretation of data exploration findings. Frontiers in Neurology | www.frontiersin.org 16 July 2022 | Volume 13 | Article 886477 Martinez et al. Understanding Cognitive Improvement in Stroke Rehab FIGURE 7 | Dependence scatter plots showing the effect of six input features in the model predictions. SHAP values (y-axis) pushes the outcome toward a class “0” (no cognitive improvement, negative SHAP values) or a class “1” (cognitive improvement, positive SHAP values). Normalized instances of features with corresponding histograms depicted in the x-axis. (A) Time since injury, (B) Length of therapy, (C) Non executed proportion, (D) memory gain, (E) memory non executed, and (F) memory task proportion. FIGURE 8 | SHAP waterfall plots for explanations of individual predictions. Illustrative cases of two patients with the highest global improvement scores. Positive feature effects are represented in red and negative effects in blue. The plot should be read from E[f(x)], the expected value of the model output, toward f(x), the model output. Positive outcomes in the x-axis (in log-odds units) are probabilities above 0.5 of classifying a patient as a cognitive improvement case (class “1”). (A) Patient_123 and (B) Patient_133. Frontiers in Neurology | www.frontiersin.org 17 July 2022 | Volume 13 | Article 886477 Martinez et al. Understanding Cognitive Improvement in Stroke Rehab FIGURE 9 | SHAP waterfall plots for explanations of individual predictions. Illustrative cases of two patients with the lowest global improvement scores. Positive feature effects are represented in red and negative effects in blue. Plot should be read from E[f(x)], expected value of the model output, toward f(x), the model output. Negative outcomes in the x-axis (in log-odds units) are probabilities above 0.5 of classifying a patient as a no cognitive improvement case (class “0”). (A) Patient_27 and (B) Patient_33. Using good visual reports was key to easing clinicians’ particular condition for which neuropsychologists recommend a interpretation of the findings. This article employed heatmaps specific set of cognitive assessments. Capturing all clinical cases is and cluster scatter plots to expose relationships among the a challenging task and requires a combined effort from clinicians predictor variables and the target features. Evidence gathered and data scientists. from this analysis served to arrive at a collaborative decision Similar to previous studies (53, 58), this article confirms the on removing correlated variables to reduce the overlap in input necessity of having a higher level of involvement from clinicians variables (multicollinearity) or keeping them to mitigate data in the development process of ML tools. More specifically, sparsity by complementing missing information. stakeholders involved in the data collection process should be Findings at this phase can be used to identify inconsistent aware of the data quality requirements in the context of ML; thus, entries and apply proper filters to assure the quality of the assuring less biased models due to data inconsistency. data. As advised in other studies, clinicians are encouraged to make sure that data from patients is valid and reliable, especially 5.2. On the Modeling Process in a context where healthcare AI is gaining more relevance Following the data exploration, the modeling phase receives a (58). Biased outcomes can be avoided if the proper filters are curated dataset to train an ML model targeting an outcome applied to the dataset before starting the modeling phase. This variable from the dataset. Several methods, commonly applied aspect is especially relevant when establishing procedures to in ML development projects, were used along with this phase. treat missing data. This study relied on recommendations from Data standardization and one-hot encoding for data pre- both clinicians and data scientists to balance the size of the processing, ANOVA f-tests for feature selection, cross-validation dataset and its descriptive capacity. Yet, several patients are for performance and accuracy estimation, and a grid-search for still left out of the final training set due to their low number parameter optimization. The adoption of structured processes of cognitive assessments. As commented by neuropsychologists, and validated methods guarantees an efficient development there are several reasons why a patient completes few cognitive process and more reliable models (59, 60). assessments at admission. In this study, the admission compliance Compared to other development phases (data exploration feature tried to capture this information and use it to predict and model explanation), clinical involvement might not be the patient’s outcome. However, there could be patients with a especially crucial during the modeling stage. Yet, as mentioned Frontiers in Neurology | www.frontiersin.org 18 July 2022 | Volume 13 | Article 886477 Martinez et al. Understanding Cognitive Improvement in Stroke Rehab in several studies (53, 58, 61), healthcare personnel (e.g., 5.4. On the Cognitive Improvement clinicians, students, attending physicians) is encouraged to Prediction engage with ML key concepts, and standard methods as AI- The model presented in this article targeted the global based systems continue to gain relevance in the clinical domain. improvement variable to predict the cognitive outcome after Better knowledge of the broad concepts in ML would increase therapy. This variable was derived from the improvement the clinicians’ trust and, ultimately, their adoption of the indicators from admission and discharge cognitive assessments. model. In addition, knowing how models function can help As in other studies, this variable was adapted to fit a binary clinicians understand their limitations and, thus, guide their classification problem by establishing an improvement own expectations. This type of engagement seeks to bridge the threshold: class “0,” global improvement <=0; class “1,” collaboration between clinicians and data scientists, putting both global improvement>0. Although this variable is not a on common ground. standardized measure, it can easily be adapted to fit different stroke rehabilitation settings, where cognitive assessments are 5.3. On the Explanatory Analysis administered. The performance of the models was evaluated Finally, the explanatory stage covered the tasks and techniques based on the Recall score to prioritize the correct identification used to deliver the model’s outcomes to clinicians. Model of patients in the class “0.” The study showed the importance of explainability is a critical field of research as it was understood informed decisions regarding the positive class, which directly that translating raw model predictions to clinical practice was affects the metric used for the learning and optimization of the not suitable since the poor justification of predictions affected the model. As for AUC scores, we observed values above 0.5 for the trust in the model and reduced the chances of it being adopted initial evaluation of algorithms and after optimization. Similar in a healthcare setting (53, 54). Clinicians identified feature to the other reported metrics, no significant improvement was importance reports as a desirable tool for rationalizing models’ noticed. These results showed that changing the algorithm predictions and using them to compare their decision-making or optimizing its parameters did not improve the model’s process (53). This study applied feature importance charts to performance. One alternative would be to deal with some of the explain the model’s inner decision workings. The analysis was data limitations, such as the class imbalance reported in Figure 5. made at a general and an individual level which allowed us to Over-sampling techniques can be applied to observe how the observe the impact of different features over stratified patient model’s performance responds. Another possible strategy would populations and individual clinical cases. be adding more variables with higher discriminatory capacity From this analysis, it was interesting to observe how the (e.g., comorbidity indicators, additional standardized scales). model replicated some of the experts’ reasoning regarding This study has paid special attention to the data preparation certain variables. For instance, for the time since injury, the phase. The techniques included in this phase are intended to results confirmed the importance of this feature and how larger facilitate the inclusion of new patients’ information. It is expected values were determinant in predicting no improvement cases. that as new patients are included, adjustments can be made, and Similarly, it was observed that extremely high values for the performance will improve. Alongside performance, the study non executed proportion feature increased the probabilities of a seeks to set the foundation for a robust ML tool at different patient being classified as a no improvement case. This behavior phases during the cognitive rehabilitation workflow. Moreover, aligned with the comments from neuropsychologists in the sense future work will explore the application of oversampling that poor compliance during therapy was an indicator of a techniques or data imputations to improve the performance of poor cognitive outcome. These results show that the trained the model. model is, in fact, following similar decision criteria used by a Very few studies have explored the use of therapeutic variables clinician. Similarly, other features that have a critical role in as predictor factors for cognitive outcomes after rehabilitation. the model’s prediction should be analyzed to find a coherent The study presented in (22) included variables representing behavior with the clinician’s decision-making process. Although the load and type of rehabilitation activities (e.g., number of the model’s prediction may not replicate the clinician’s rationale, it can provide a perspective to support clinical decisions, even sessions, number of executed activities and mean number of if it does by providing data supporting a counterargument executed activities per session) and used the Disability Ranking or perspective. Scale (DRS) as the target variable. That study reflected on the The feature importance charts generated using the SHAP importance of providing actionable information to therapists and method served well to interpret the outcomes from the use how this can lead to the design of better intervention programs. case presented in this study. However, this approach might Our study extended the granularity level of therapy variables by not apply to more complex ML methods. Depending on decomposing the therapy information across cognitive domains. the context, alternative solutions like the local interpretable This enables the delivery of detailed information on the specific model-agnostic explanations (LIME), which target individual cognitive domain impacting the overall cognitive outcome (e.g., level explanations (62), and new explanation techniques memory task proportion and orientation task proportion). In for Deep Neural Network (DNN) models (63, 64) can addition, we introduce new variables that have shown promising be utilized. results acting as predictor variables (e.g., admission compliance). Frontiers in Neurology | www.frontiersin.org 19 July 2022 | Volume 13 | Article 886477 Martinez et al. Understanding Cognitive Improvement in Stroke Rehab Although performance can be improved, possibly by 6. CONCLUSION including comorbidity or demographic information and using more sophisticated ML algorithms, the model showed its Stroke rehabilitation poses several challenges to clinicians, capacity to identify patients “at risk” (i.e., patients belonging to especially in building effective therapeutic programs that the class “0,” global improvement<=0). This information can be maximize patients’ cognitive recovery. ML-based tools can assist helpful to alert clinicians of cases where the therapy program clinicians in this task, but they need to produce actionable requires adjustment. Using the trained model, outcomes for information and well-justified outcomes to increase their chances different therapy settings can be tested and reported using scatter of adoption into a clinical setting. This article presented charts to identify the configuration that maximizes the patient’s an ML prediction model that targets cognitive improvement cognitive improvement. after therapy for ischemic stroke surviving patients. Besides demographic and cognitive assessments variables, which are commonly used in literature, the model included therapeutic 5.5. Limitations variables as predictor features. A large number of classifier The sparsity of the data was the main limitation of this study. algorithms were trained, optimized, and evaluated, varying As reported, features from the cognitive assessment records had the number of features received as input. A set of feature very different completeness levels, forcing the removal of some importance visual reports allowed the interpretation of the variables to maintain an acceptable number of entries in the inner decision workings of the best-performing model. The dataset. In terms of descriptive features, the study didn’t consider study showed how certain therapeutic variables impacted the some comorbidity indicators and some relevant demographic rehabilitation outcome for individual cases and across the entire information like the level of education. For this study, global study population. Potentially, clinicians can use this type of improvement was chosen as the model’s target variable with a evidence to tune a therapeutic setting that maximizes the defined threshold used to frame the model output as a binary cognitive outcome. classification problem. Both the variable and the threshold used in this study still need further exploration to support DATA AVAILABILITY STATEMENT their clinical significance. The threshold intends to separate patients with cognitive improvement from those that showed no The datasets generated for this study will be made available upon improvement or got worse. Therefore, they must be considered reasonable request to the corresponding author of the article. just as referential indicators of a patient’s cognitive status. This study relied on specific cognitive assessments to act as ETHICS STATEMENT cognitive outcome predictors; however, standardized scales for functional independence (e.g., FIM, BI) or degree of disability A specific written informed consent was not required for (e.g., mRS) in activities of daily living were not considered. participants to be included in this study, in accordance with the Cognitive functioning has been reported to be strongly involved local legislation and institutional requirements. Nevertheless at in the execution of activities of daily living. For example, Mori admission participants provide written informed consent to be et al. (65) concluded independence in eating is strongly associated included in research studies addressed by the Institut Guttmann with cognitive improvement in subacute stroke. Functional hospital. The authors confirm that this study is compliant with independence has been extensively assessed in rehabilitation the Helsinki Declaration of 1975, as revised in 2008 and it settings after stroke using standardized instruments such as the was approved by the Ethics Committee of Clinical Research of FIM or the Barthel index, which have not been considered Institut Guttmann. in this study. This leaves room for future research, e.g., considering individual FIM items and their relationship to AUTHOR CONTRIBUTIONS specific cognitive functioning assessments such as the TMT or RAVLT. Nevertheless, we have included stroke severity HM and AH conceived the study. HM, KC, and AG-R collected, in our study, assessed using the NIHSS, the FIM has been selected, and cleaned the data. HM, KC, JK, AG-R, and AH reported in previous research to be strongly associated with statistically analyzed the data and received funding for this study. the NIHSS. For example, according to Roth et al. (66) NIHSS HM drafted the manuscript. JK, AG-R, KC, and AH revised correlated significantly with motor and cognitive FIM subscores the manuscript critically for important intellectual content and for admission, discharge, and change measures. approved the final manuscript. All authors contributed to the article and approved the submitted version. 5.6. Future Study and Direction FUNDING This study presented a model targeting the cognitive outcome prediction task. An extended model will be implemented adding This research was partially funded by PRECISE4Q Personalized more features to increase the accuracy of the model. Following Medicine by Predictive Modelling in Stroke for better Quality a similar development methodology, other target features like of Life-European Union’s Horizon 2020 research and innovation non executed proportion will be considered to explore the level program under grant agreement No. 777107, and by the Science of therapy compliance of rehabilitation. Foundation Ireland (SFI) under Grant Number 17/RC/2289_P2. Frontiers in Neurology | www.frontiersin.org 20 July 2022 | Volume 13 | Article 886477 Martinez et al. Understanding Cognitive Improvement in Stroke Rehab REFERENCES algorithm in human subject that underwent a rehabilitation treatment in the early stage of stroke. J Stroke Cerebrovasc Dis. (2018) 27:2962–72. 1. Krishnamurthi RV, Feigin VL, Forouzanfar MH, Mensah GA, Connor M, doi: 10.1016/j.jstrokecerebrovasdis.2018.06.021 Bennett DA, et al. Global and regional burden of first-ever ischaemic 20. Dacosta-Aguayo R, Stephan-Otto C, Auer T, Clemente I, Davalos A, and haemorrhagic stroke during 1990-2010: findings from the global Bargalló N, et al. Predicting cognitive recovery of stroke patients from the burden of disease study 2010. Lancet Glob Health. (2013) 1:e259-e281. structural MRI Connectome Using a Naïve tree classifier. In: 2017 IEEE doi: 10.1016/S2214-109X(13)70089-5 30th International Symposium on Computer-Based Medical Systems (CBMS). 2. Ramirez L, Kim-Tenser MA, Sanossian N, Cen S, Wen G, He S, et al. Trends in Thessaloniki: IEEE (2017). p. 413–8. acute ischemic stroke hospitalizations in the United States. J Am Heart Assoc. 21. Kogan E, Twyman K, Heap J, Milentijevic D, Lin JH, Alberts M. (2016) 5:e003233. doi: 10.1161/JAHA.116.003233 Assessing stroke severity using electronic health record data: a machine 3. Tibæk M, Dehlendorff C, Jørgensen HS, Forchhammer HB, Johnsen SP, learning approach. BMC Med Inform Dec Making. (2020) 20:1–8. Kammersgaard LP. Increasing incidence of hospitalization for stroke and doi: 10.1186/s12911-019-1010-x transient ischemic attack in young adults: a registry-based study. J Am Heart 22. García-Rudolph A, Garcıa-Molina A, Tormos Muñoz JM. Predictive models Assoc. (2016) 5:e003158. doi: 10.1161/JAHA.115.003158 for cognitive rehabilitation of patients with traumatic brain injury. Intell Data 4. Béjot Y, Bailly H, Durier J, Giroud M. Epidemiology of stroke in Europe Anal. (2019) 23:895–915. doi: 10.3233/IDA-184154 and trends for the 21st century. La Presse Médicale. (2016) 45:e391–8. 23. Tang EYH, Robinson L, Stephan BCM. Risk prediction models for post-stroke doi: 10.1016/j.lpm.2016.10.003 dementia. Geriatrics. (2017) 2:19. doi: 10.3390/geriatrics2030019 5. Hankey GJ. Long-term outcome after ischaemic stroke/transient ischaemic 24. Lai S, Billot A, Varkanitsa M, Braun E, Rapp B, Parrish T, et al. An exploration attack. Cerebrovasc Dis. (2003) 16(Suppl. 1):14–19. doi: 10.1159/000069936 of machine learning methods for predicting post-stroke aphasia recovery. In: 6. Rabinowitz AR, Levin HS. Cognitive sequelae of traumatic brain injury. The 14th PErvasive Technologies Related to Assistive Environments Conference. Psychiatr Clin. (2014) 37:1–11. doi: 10.1016/j.psc.2013.11.004 Corfu (2021). p. 556–64. 7. Wade DT. Measurement in neurological rehabilitation. Curr Opin Neurol 25. Challen R, Denny J, Pitt M, Gompels L, Edwards T, Tsaneva-Atanasova Neurosurg. (1992) 5:682–6. K. Artificial intelligence, bias and clinical safety. BMJ Qual Safety. (2019) 8. Kreutzer JS, DeLuca J, Caplan B. Encyclopedia of Clinical Neuropsychology. 28:231–7. doi: 10.1136/bmjqs-2018-008370 New York, NY: Springer (2011). 26. Parikh RB, Teeple S, Navathe AS. Addressing bias in artificial intelligence in 9. Quinn TJ, Paolucci S, Sunnerhagen KS, Sivenius J, Walker MF, Toni D, health care. JAMA. (2019) 322:2377–8. doi: 10.1001/jama.2019.18058 et al. Evidence-based stroke rehabilitation: an expanded guidance document 27. Stinear CM, Smith MC, Byblow WD. Prediction tools for stroke rehabilitation. from the European Stroke Organisation (ESO) guidelines for management of Stroke. (2019) 50:3314–22. doi: 10.1161/STROKEAHA.119.025696 ischaemic stroke and transient ischaemic attack 2008. J Rehabil Med. (2009) 28. Miller EL, Murray L, Richards L, Zorowitz RD, Bakas T, Clark P, et al. 41:99–111. doi: 10.2340/16501977-0301 Comprehensive overview of nursing and interdisciplinary rehabilitation 10. O’Neil-Pirozzi TM, Hsu H. Feasibility and benefits of computerized cognitive care of the stroke patient: a scientific statement from the American Heart exercise to adults with chronic moderate-to-severe cognitive impairments Association. Stroke. (2010) 41:2402–48. doi: 10.1161/STR.0b013e3181e7512b following an acquired brain injury: a pilot study. Brain Injury. (2016) 29. Prvu Bettger J, Alexander KP, Dolor RJ, Olson DM, Kendrick AS, Wing L, 30:1617–25. doi: 10.1080/02699052.2016.1199906 et al. Transitional care after hospitalization for acute stroke or myocardial 11. Ye M, Zhao B, Liu Z, Weng Y, Zhou L. Effectiveness of computer- infarction: a systematic review. Ann Internal Med. (2012) 157:407–16. based training on post-stroke cognitive rehabilitation: a systematic doi: 10.7326/0003-4819-157-6-201209180-00004 review and meta-analysis. Neuropsychol Rehabil. (2020) 32:481–97. 30. Ben-David A, Frank E. Accuracy of machine learning models versus “hand doi: 10.1080/09602011.2020.1831555 crafted” expert systems-a credit scoring case study. Expert Syst Appl. (2009) 12. Smith DM, Brown SL, Ubel PA. Mispredictions and misrecollections: 36:5264–71. doi: 10.1016/j.eswa.2008.06.071 challenges for subjective outcome measurement. Disabil Rehabil. (2008) 31. Lee MH, Siewiorek DP, Smailagic A, Bernardino A, Bermúdez i Badia S. 30:418–24. doi: 10.1080/09638280701625237 Interactive hybrid approach to combine machine and human intelligence for 13. Hendricks HT, Van Limbeek J, Geurts AC, Zwarts MJ. Motor recovery after personalized rehabilitation assessment. In: Proceedings of the ACM Conference stroke: a systematic review of the literature. Arch Phys Med Rehabil. (2002) on Health, Inference, and Learning. New York, NY (2020). p. 160–9. 83:1629–37. doi: 10.1053/apmr.2002.35473 32. Cognitive. Cognitive prognosis of acquired brain injury patients using 14. Obermeyer Z, Emanuel EJ. Predicting the future–big data, machine machine learning techniques. In: The Fifth International Conference on learning, and clinical medicine. New Engl J Med. (2016) 375:1216. Advanced Cognitive Technologies and Applications, Valencia (2013). p. 108– doi: 10.1056/NEJMp1606181 113. Available online at: http://personales.upv.es/thinkmind/COGNITIVE/ 15. Lee MH, Siewiorek DP, Smailagic A, Bernardino A, Bermúdez i Badia COGNITIVE_2013/cognitive_2013_5_50_40058.html SB. A human-AI collaborative approach for clinical decision making on 33. Rudolph AG, Molina AG, Opisso E, Tormos JM. Patterns in cognitive rehabilitation assessment. In: Proceedings of the 2021 CHI Conference on rehabilitation of traumatic brain injury patients: a text mining approach. In: Human Factors in Computing Systems. Yokohama (2021). p. 1–14. 2016 IEEE 16th International Conference on Data Mining (ICDM). Barcelona: 16. Venkataraman V, Turaga P, Lehrer N, Baran M, Rikakis T, Wolf SL. Decision IEEE (2016). p. 1185–90. support for stroke rehabilitation therapy via describable attribute-based 34. García-Rudolph A, Gibert K. Understanding effects of cognitive rehabilitation decision trees. Annu Int Conf IEEE Eng Med Biol Soc. (2014) 2014:3154–9. under a knowledge discovery approach. Eng Appl Artif Intell. (2016) doi: 10.1109/EMBC.2014.6944292 55:165–85. doi: 10.1016/j.engappai.2016.06.007 17. Solana J, Cáceres C, García-Molina A, Chausa P, Opisso E, Roig-Rovira 35. Peña-Casanova J, Guardia J, Bertran-Serra I, Manero R, Jarne A. Shortened T, et al., Intelligent Therapy Assistant (ITA) for cognitive rehabilitation in version of the Barcelona test (I): subtests and normal profiles. Neurología patients with acquired brain injury. BMC Med Inform Dec Making. (2014) (Barcelona, Spain). (1997) 12:99–111. 14:1–13. doi: 10.1186/1472-6947-14-58 36. Scale WDWAI. (WAIS-III). San Antonio, TX: The Psychological Corporation 18. Luvizutto GJ, Silva GF, Nascimento MR, Sousa Santos KC, Appelt (1997). PA, de Moura Neto E, et al. Use of artificial intelligence as an 37. Reitan RM, Wolfson D. The Halstead-Reitan neuropsychological test battery: instrument of evaluation after stroke: a scoping review based on theory and clinical interpretation. Vol. 4. In: Reitan Neuropsychology. New international classification of functioning, disability and health concept: York, NY (1985). AI applications for stroke evaluation. Topics Stroke Rehabil. (2021) 29:331–46. 38. Scarpina F, Tagini S. The stroop color and word test. Front Psychol. (2017) doi: 10.1080/10749357.2021.1926149 8:557. doi: 10.3389/fpsyg.2017.00557 19. Sale P, Ferriero G, Ciabattoni L, Cortese AM, Ferracuti F, Romeo L, et al. 39. Schmidt M. Rey Auditory Verbal Learning Test: A Handbook. Los Angeles, CA: Predicting motor and cognitive improvement through machine learning Western Psychological Services (1996). Frontiers in Neurology | www.frontiersin.org 21 July 2022 | Volume 13 | Article 886477 Martinez et al. Understanding Cognitive Improvement in Stroke Rehab 40. Heaton RK, Staff P. Wisconsin card sorting test: computer version 2. Odessa. 56. Lundberg SM, Lee SI. A unified approach to interpreting model predictions. (1993) 4:1–4. In: Proceedings of the 31st International Conference on Neural Information 41. Artiola L, Hermosillo D, Heaton R, Pardee R. Manual de normas y Processing Systems. Long Beach, NY (2017). p. 4768–77. procedimientos para la batería neuropsicológica en español. Tucson, AZ: M 57. Garg A, Tai K. Comparison of statistical and machine learning methods in Press Tucson (1999). modelling of data with multicollinearity. Int J Model Identification Control. 42. Brott T, Adams Jr HP, Olinger CP, Marler JR, Barsan WG, Biller J, et al. (2013) 18:295–312. doi: 10.1504/IJMIC.2013.053535 Measurements of acute cerebral infarction: a clinical examination scale. 58. Rowe M. An introduction to machine learning for clinicians. Acad Med. Stroke. (1989) 20:864–70. doi: 10.1161/01.STR.20.7.864 (2019) 94:1433–6. doi: 10.1097/ACM.0000000000002792 43. García-Molina A, Rajo PRR, Sánchez-Carrión R, Pulido AGG, Ensenyat A, 59. Serban A, van der Blom K, Hoos H, Visser J. Adoption and effects of software Rudolph AGG, et al. Clinical program of cognitive tele-rehabilitation for engineering best practices in machine learning. In: Proceedings of the 14th traumatic brain injury. In: eChallenges e-2010 Conference. Warsaw: IEEE ACM/IEEE International Symposium on Empirical Software Engineering and (2010). p. 1–10. Measurement (ESEM). Bari (2020). p. 1–12. 44. Norris CM, Ghali WA, Knudtson ML, Naylor CD, Saunders LD. Dealing with 60. Chen PHC, Liu Y, Peng L. How to develop machine learning models for missing data in observational health care outcome analyses. J Clin Epidemiol. healthcare. Nat Mater. (2019) 18:410–4. doi: 10.1038/s41563-019-0345-0 (2000) 53:377–83. doi: 10.1016/S0895-4356(99)00181-X 61. Vercio LL, Amador K, Bannister JJ, Crites S, Gutierrez A, MacDonald ME, 45. Kelleher JD, Mac Namee B, D’arcy A. Fundamentals of Machine Learning for et al. Supervised machine learning tools: a tutorial for clinicians. J Neural Eng. Predictive Data Analytics: Algorithms, Worked Examples, and Case Studies. (2020) 17:062001. doi: 10.1088/1741-2552/abbff2 Cambridge, MA: MIT Press (2020). 62. Ribeiro MT, Singh S, Guestrin C. “Why should i trust you?” Explaining 46. García-Rudolph A, Gibert K. A data mining approach to identify cognitive the predictions of any classifier. In: Proceedings of the 22nd ACM SIGKDD NeuroRehabilitation range in traumatic brain injury patients. Expert Syst International Conference on Knowledge Discovery and Data Mining. San Appl. (2014) 41:5238–51. doi: 10.1016/j.eswa.2014.03.001 Francisco, CA (2016). p. 1135–44. 47. Van der Maaten L, Hinton G. Visualizing data using t-SNE. J Mach Learn 63. Montavon G, Lapuschkin S, Binder A, Samek W, Müller KR. Explaining Res. (2008) 9:2579–605. Available online at: https://www.jmlr.org/papers/v9/ nonlinear classification decisions with deep taylor decomposition. Pattern vandermaaten08a.html Recogn. (2017) 65:211–22. doi: 10.1016/j.patcog.2016.11.008 48. Tipping ME, Bishop CM. Mixtures of probabilistic principal 64. Liu D, Zhang L, Luo T, Tao L, Wu Y. Towards interpretable and robust component analyzers. Neural Comput. (1999) 11:443–82. hand detection via pixel-wise prediction. Pattern Recogn. (2020) 105:107202. doi: 10.1162/089976699300016728 doi: 10.1016/j.patcog.2020.107202 49. Singh D, Singh B. Investigating the impact of data normalization 65. Mori N, Otaka Y, Honaga K, Matsuura D, Kondo K, Liu M, et al. Factors on classification performance. Appl Soft Comput. (2020) 97:105524. associated with cognitive improvement in subacute stroke survivors. J Rehabil doi: 10.1016/j.asoc.2019.105524 Med. (2021) 53:jrm00220. doi: 10.2340/16501977-2859 50. Yeo IK, Johnson RA. A new family of power transformations to 66. Roth EJ, Heinemann AW, Lovell LL, Harvey RL, McGuire JR, Diaz S. improve normality or symmetry. Biometrika. (2000) 87:954–959. Impairment and disability: their relation during stroke rehabilitation. Arch doi: 10.1093/biomet/87.4.954 Phys Med Rehabil. (1998) 79:329–35. doi: 10.1016/S0003-9993(98)90015-6 51. Álvarez JD, Matias-Guiu JA, Cabrera-Martín MN, Risco-Martín JL, Ayala JL. An application of machine learning with feature selection to improve Conflict of Interest: The authors declare that the research was conducted in the diagnosis and classification of neurodegenerative disorders. BMC Bioinform. absence of any commercial or financial relationships that could be construed as a (2019) 20:1–12. doi: 10.1186/s12859-019-3027-7 potential conflict of interest. 52. Voigt P, Von dem Bussche A. The eu General Data Protection Regulation (gdpr). A Practical Guide, 1st Edn. Cham: Springer International Publishing Publisher’s Note: All claims expressed in this article are solely those of the authors (2017). 53. Tonekaboni S, Joshi S, McCradden MD, Goldenberg A. What clinicians want: and do not necessarily represent those of their affiliated organizations, or those of contextualizing explainable machine learning for clinical end use. In: Machine the publisher, the editors and the reviewers. Any product that may be evaluated in Learning for Healthcare Conference. Ann Arbor, MI: PMLR (2019). p. 359–80. this article, or claim that may be made by its manufacturer, is not guaranteed or 54. Ahmad MA, Eckert C, Teredesai A. Interpretable machine endorsed by the publisher. learning in healthcare. In: Proceedings of the 2018 ACM International Conference on Bioinformatics, Computational Copyright © 2022 Martinez, Cisek, García-Rudolph, Kelleher and Hines. This is an Biology, and Health Informatics. Washington, DC (2018). open-access article distributed under the terms of the Creative Commons Attribution p. 559–60. License (CC BY). The use, distribution or reproduction in other forums is permitted, 55. Antoniadi AM, Du Y, Guendouz Y, Wei L, Mazo C, Becker BA, et al. provided the original author(s) and the copyright owner(s) are credited and that the Current challenges and future opportunities for XAI in machine learning- original publication in this journal is cited, in accordance with accepted academic based clinical decision support systems: a systematic review. Appl Sci. (2021) practice. No use, distribution or reproduction is permitted which does not comply 11:5088. doi: 10.3390/app11115088 with these terms. Frontiers in Neurology | www.frontiersin.org 22 July 2022 | Volume 13 | Article 886477