Fernández‑Gonzalo et al. Crit Care (2020) 24:618 https://doi.org/10.1186/s13054-020-03334-2 RESEARCH Open Access Cognitive phenotypes 1 month after ICU discharge in mechanically ventilated patients: a prospective observational cohort study Sol Fernández‑Gonzalo1,2,3* , Guillem Navarra‑Ventura1,4, Neus Bacardit5, Gemma Gomà Fernández1, Candelaria de Haro1,4, Carles Subirà6, Josefina López‑Aguilar1,4, Rudys Magrans7, Leonardo Sarlabous1, Jose Aquino Esperanza1,4,8, Mercè Jodar2,3,9, Montse Rué10,11, Ana Ochagavía1,4, Diego J. Palao2,12,13, Rafael Fernández4,6 and Lluís Blanch1,4,8 Abstract Background: ICU patients undergoing invasive mechanical ventilation experience cognitive decline associated with their critical illness and its management. The early detection of different cognitive phenotypes might reveal the involvement of diverse pathophysiological mechanisms and help to clarify the role of the precipitating and predispos‑ ing factors. Our main objective is to identify cognitive phenotypes in critically ill survivors 1 month after ICU discharge using an unsupervised machine learning method, and to contrast them with the classical approach of cognitive impairment assessment. For descriptive purposes, precipitating and predisposing factors for cognitive impairment were explored. Methods: A total of 156 mechanically ventilated critically ill patients from two medical/surgical ICUs were prospec‑ tively studied. Patients with previous cognitive impairment, neurological or psychiatric diagnosis were excluded. Clini‑ cal variables were registered during ICU stay, and 100 patients were cognitively assessed 1 month after ICU discharge. The unsupervised machine learning K-means clustering algorithm was applied to detect cognitive phenotypes. Exploratory analyses were used to study precipitating and predisposing factors for cognitive impairment. Results: K-means testing identified three clusters (K) of patients with different cognitive phenotypes: K1 (n = 13), severe cognitive impairment in speed of processing (92%) and executive function (85%); K2 (n = 33), moderate-to- severe deficits in learning-memory (55%), memory retrieval (67%), speed of processing (36.4%) and executive func‑ tion (33.3%); and K3 (n = 46), normal cognitive profile in 89% of patients. Using the classical approach, moderate-to- severe cognitive decline was recorded in 47% of patients, while the K-means method accurately classified 85.9%. The descriptive analysis showed significant differences in days (p = 0.016) and doses (p = 0.039) with opioid treatment in K1 vs. K2 and K3. In K2, there were more women, patients were older and had more comorbidities (p = 0.001) than in K1 or K3. Cognitive reserve was significantly (p = 0.001) higher in K3 than in K1 or K2. Conclusion: One month after ICU discharge, three groups of patients with different cognitive phenotypes were iden‑ tified through an unsupervised machine learning method. This novel approach improved the classical classification *Correspondence: msfernandez@tauli.cat 1 Critical Care Center, Parc Taulí Hospital Universitari, Fundació‑ I3PT, UAB, Sabadell, Spain Full list of author information is available at the end of the article © The Author(s) 2020. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creat​iveco​mmons​.org/licen​ses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creat​iveco​ mmons​.org/publi​cdoma​in/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data. Fernández‑Gonzalo et al. Crit Care (2020) 24:618 Page 2 of 11 of cognitive impairment in ICU survivors. In the exploratory analysis, gender, age and the level of cognitive reserve emerged as relevant predisposing factors for cognitive impairment in ICU patients. Trial registration: ClinicalTrials.gov Identifier:NCT02390024; March 17,2015. Keywords: Cognition in ICU survivors, Neuropsychological profiles, Critical illness, Post-intensive care syndrome Introduction Most of the studies evaluating cognitive impairment Cognitive impairment in critically ill survivors who have in ICU survivors after MV have focused on global cog- undergone invasive mechanical ventilation (MV) during nitive impairment, and little is known about the charac- the Intensive Care Unit (ICU) stay is a well-established terization of different phenotypes of alteration [12]. The health problem [1–11]. ICU admission is associated early detection of distinct groups of patients regarding with a greater cognitive decline than general ward hos- cognitive deficits might reveal the involvement of diverse pitalization [12], and the rate of dementia diagnosis after pathophysiological mechanisms. Thus, the main objective 3 years of follow-up has been reported to be higher in of this study was to describe the cognitive phenotypes elderly ICU survivors than in the general population [13]. 1 month after ICU discharge in survivors of critical ill- This cognitive decline affects the functional and socio- ness who had undergone MV during their ICU stay, using economic status of ICU survivors and their families, and an unsupervised machine learning method. Two second- reduces their quality of life after ICU discharge [5, 14– ary objectives were also proposed. To warrant clinical 16]. Cognitive impairment may be as high as 64% in ICU interpretation of these results, we contrasted the cogni- survivors [17]. The domains most commonly affected are tive phenotypes with the classical definition of cognitive attention/concentration, memory and executive function impairment in critically ill patients established by Jackson [18]. et al. [7]. An exploratory analysis of the predisposing and The pathophysiological mechanisms that lead to cogni- precipitating factors (e.g., gender, medications, severity tive dysfunction after critical illness are not well under- of illness and days with MV) of the cognitive dysfunction stood. Many precipitating or modifiable factors and after critical care was also carried out. predisposing or non-modifiable factors have been related to the short- and long-term cognitive deficit observed in Material and methods ICU survivors [19]. Among precipitating factors such as Sample and procedure MV [2, 20], length of ICU stay [9, 20], hypoxemia, hypo- This prospective cohort study enrolled patients from two glycemia, hyperglycemia, fluctuations in serum glucose medical/surgical ICUs from October 2015 until Decem- [20–22] and perceived stress levels during the ICU stay ber 2017. [23], the presence of delirium has shown the closest asso- The study sample comprised critically ill adult patients ciation with the cognitive impairment observed in ICU (≥ 18 years old) who had undergone invasive MV for at survivors [2, 3, 24–26]. Predisposing or non-modifiable least 24 h during the ICU stay and had been enrolled dur- factors have been less explored and in most studies have ing the first 48 h of MV. Exclusion criteria were: prior been considered as confounding factors. Nonetheless, cognitive impairment or dementia, diagnosis of neu- older age, previous cognitive impairment and higher ill- rological disease, psychiatric disorder, sensory deficits ness severity seem to increase the risk of developing (blindness or deafness), non-Spanish speaking and life cognitive dysfunction after ICU stay [18, 27]. The role of expectancy < 3 months. Previous neurological diagnosis other individual predisposing factors has only rarely been (specifically of prior cognitive impairment or dementia) studied in critical illness. One of these factors is cogni- was checked in patients’ clinical records and through tive reserve—that is, the ability of the brain to actively interviews with their relatives. In patients > 60 years old, address brain dysfunction by using pre-existing cogni- preexisting cognitive impairment was assessed at ICU tive processing approaches or by enlisting compensatory admission using the Spanish version of the Short Form approaches [28]. Cognitive reserve may confer a better of the Informant Questionnaire on Cognitive Decline in resilience to pathological brain changes; that is, people the Elderly (Short-IQCODE), filled in by relatives [29]. with higher cognitive reserve may be less vulnerable to The cutoff point for exclusion on the Short-IQCODE neurophysiological insults such as the impact of critical was > 3.56. Daily screening was performed by a critical illness and its management. Cognitive reserve may also care nurse. At enrollment, written informed consent was be a target for rehabilitation programs when the brain obtained from the patients or their authorized surro- insult has already occurred. gates; if consent was initially obtained from a surrogate, Fernández‑Gonzalo et al. Crit Care (2020) 24:618 Page 3 of 11 it was subsequently obtained from the patient once s/he and sedative and opioid administration were adjusted was deemed to be mentally competent. for days of ICU. This data correction was performed by consensus between authors, on the assumption that data Measurements adjusted to the days of ICU can be a better reflection of At inclusion, severity of illness was measured by the ICU patients’ true status. However, both variables (with Acute Physiology and Chronic Health Evaluation and without adjustment) were initially considered in the (APACHE II) and level of comorbidity by the Charlson univariate analysis. Total doses of opioids (morphine and Comorbidity Index. Level of consciousness and presence fentanyl) and sedatives (midazolam, propofol and loraz- of delirium were assessed daily using the Richmond Agi- epam) administered each day were recorded and con- tation Sedation Scale (RASS) and the Confusion Assess- verted to morphine and midazolam equivalents [30]. ment Method for the ICU (CAM-ICU), respectively. The cognitive reserve index was obtained for each Sequential Organ Failure Assessment (SOFA) was meas- patient based on two proxies: educational attainment, ured every day during ICU stay. Number of days receiv- defined as the number of years in full-time formal edu- ing sedatives and opioids as well as the accumulated cation, and literacy, based on the corresponding intelli- doses of sedation and analgesia were also recorded daily. gence quotient (IQ) score on the Spanish version of the All sequential data were recorded until ICU discharge or National Adult Reading Test (NART) [31]. for a maximum of 30 days. Raw scores of the cognitive tests were transformed into A complete and comprehensive neuropsychological z-scores (mean = 0; SD = ± 1) using the normative popu- assessment was administered 1 month after ICU dis- lation data offered by each test. Details of the cognitive charge. Details of the neuropsychological battery are index calculation are given in Additional file 1: Table S1. shown in Table 1. The neuropsychological assessment Two different approaches were used for assessing cog- was administered to all participants by an expert neu- nitive impairment in the sample: the first a more classi- ropsychologist during a single session lasting between 45 cal approach based on the clinical definition of cognitive and 60 min. Six cognitive indexes were calculated from impairment, and the second a method based on statisti- the neuropsychological test scores: attention, learning cal clustering algorithms. and memory storage, memory retrieval, speed of pro- In the classical approach, global cognitive impairment cessing, working memory and executive function. was considered when a patient scored ≥ 1.5 SD below the mean on two or more of the six cognitive indexes, Statistical analysis or else when a patient scored ≥ 2 SD below the mean on Baseline demographics and clinical characteristics during one or more of the indexes. This clinical definition has the ICU stay were summarized using medians (IQR) for been previously used in the literature [7] and is consid- continuous variables and percentages for categorical var- ered a strict, accurate description of cognitive deficit. The iables. For the purposes of analysis, the variable diagnosis z-scores were also used to determine the level of cogni- was operationalized in three categories (medical, surgical tive deficit in each domain for every patient. Likewise, and polytrauma). Daily SOFA scores were summarized by cognitive domain impairment was defined as moderate if a single value defined as the slope of the regression line the values of the index were ≥ 1.5 SD below the mean and on the first five SOFA scores. Length of delirium, MV, severe if the values were ≥ 2 SD below the mean. Table 1 Cognitive tests used in the neuropsychological assessment battery Cognitive domain Tests Premorbid intelligence quotient (IQ) estimation The National Adult Reading Test (NART)—Spanish version- S.1 Verbal attention and working memory Subtest of Digits from the Wechsler Adult Intelligence Scale version III (WAIS III) S.2 Visual attention and working memory Subtest of Spatial Span from the Wechsler Memory Scale version III (WMS III) S.3 Learning, short- and long- term verbal memory Rey Auditory Verbal Learning test S.4 Visual memory Benton Visual Retention test S.5 Speed of processing Subtest of Symbol Search (WAIS III) S.2 Speed of processing and Executive function (Automatic response Stroop Color and Word test S.6 inhibition) Speed of processing and Executive function (Flexibility) Trail Making Test S.7 Executive function (phonetic verbal fluency) FAS S.8 See references in the Supplemental Material Fernández‑Gonzalo et al. Crit Care (2020) 24:618 Page 4 of 11 The second approach aimed to find characteristic 66.67% of participants showed alterations in memory phenotypes and clusters of patients’ cognitive decline retrieval, 54.5% in learning and memory storage, 36.36% through an unsupervised machine learning technique, in speed of processing and 33.33% in executive functions. the K-means clustering algorithm. K-means clustering is In the K3 cluster, only 11% of patients showed global cog- widely used in the data mining field to reveal naturally nitive impairment: 15.22% presented impaired executive occurring patterns or groupings, without targeting a spe- function and 13.04% reduced speed of processing. Fur- cific outcome [32, 33]. The clustering algorithm splits ther details are given in Fig. 2 and in Additional file 1: the observations (i.e., patients) into distinct and non- Table S2. overlapping clusters of cognitive phenotypes based on their cognitive score, without the need to establish an ‘a Prevalence of global cognitive impairment and cognitive priori’ definition of cognitive deficit. The procedure was domain deficits using the classical approach: comparison as follows: at each iteration, the algorithm computed the with the K‑means clustering algorithm approach cluster centroids and each observation was assigned to Using the classical approach, 43 patients (47%) presented the closest centroid. The Euclidean distance was used as global cognitive impairment. When the six cognitive the dissimilarity measure to minimize the within-cluster indexes were explored separately, the most commonly variation. affected domain was speed of processing (33 patients, The Kruskal–Wallis test was used to analyze the dif- 36%), followed by executive function (29 patients, 31%), ferences between the K groups in the demographic and memory retrieval (23 patients, 25%) and learning and clinical data, and the Mann–Whitney U test for group memory storage ability (20 patients, 22%). Attention comparisons. and working memory were marginally affected, with one Only for exploratory purposes, the association between patient (1%) showing a deficit in each domain. Note that predisposing and/or precipitating factors and cognitive the same patient could present deficits in different cogni- impairment was studied by binomial logistic regression tive domains at the same time. analyses. Factors with a p value < 0.2 in the univariate The results obtained with the two approaches were analysis were included in the multivariable analysis. Mod- compared in order to warrant a feasible clinical interpre- eling adjustments focused on obtaining a simplified tation of the groups obtained through the unsupervised model included all variables with p values < 0.05 or with machine learning approach. Compared to the classi- clinical relevance for explaining the dependent variable. cal definition of cognitive impairment, the K-means test classified most of the participants with cognitive impair- Results ment between the K1 and K2 groups; however, 89% of Four hundred and forty-two ICU patients were screened the participants in K3 did not present cognitive impair- for inclusion. One hundred and fifty-six met the inclu- ment. Thus, the K-means test failed to classify 13 patients sion criteria, of whom 100 were cognitively assessed (14.1%) but correctly classified 79 (85.9%) (Fig. 3). 1 month after ICU discharge. The final analysis was performed in 92 patients (Fig. 1). Mean age of the sam- Analysis of predisposing and precipitating factors ple was 63.03 years (SD 12.78), and 57 participants were The descriptive analysis of the differences in the demo- male (62%). Further demographic and clinical data of the graphic and clinical variables between the three phe- sample are shown in Table 2. notypes showed that in K2 there were more women (p = 0.001), patients were older (p < 0.001) and had Cognitive clusters of patients with the unsupervised more comorbidities (p < 0.001) than in K1 or K3. Sig- machine learning K‑means clustering algorithm approach nificant differences were also found in days (p = 0.016) The K-means test results showed three differential clus- and doses (p = 0.039) of opioid treatment in K1 than in ters of patients grouped regarding their characteristic K2 or K3. Cognitive reserve was significantly higher in cognitive phenotypes (13 patients in K1, 33 in K2 and 46 K3 (p = 0.001) than in K1 or K2. Further details of the in K3) (Additional file 1: Fig. S1). The correlation analy- demographic and clinical data according to cluster group sis of the composite cognitive variables showed that all and percentage of diagnosis are shown in Table 3 and in cognitive variables could be maintained in the K-means Additional file 1: S3. analysis (Additional file 1: Fig. S2). Since the K1 cluster contained only 13 patients, it was In the K1 cluster, all patients showed global cognitive merged in a single group with the K2 cluster in order to impairment: the most altered functions were speed of explore the influence of predisposing and precipitat- processing, which affected 93.23% of patients, and execu- ing factors on global cognitive impairment by means of tive function, which affected 84.62%. In the K2 cluster, the binomial logistic regression model. The exploratory 76% of patients showed global cognitive impairment: analysis was run in a group (K1 + K2) in which most Fernández‑Gonzalo et al. Crit Care (2020) 24:618 Page 5 of 11 Patients screened for inclusion (n=442) 01.07.2015 – 14.12.2017 Excluded (n=286) •Life expectancy <2 months (n=23) •Psychiatric diagnosis (n=28) •Neurological diagnosis (n=54) •Cardiorespiratory arrest (n=4) •Therapeutic effort limitation (n=11) •Toxic abuse (n=2) •MV >48h (n=54) •MV <24h (n=65) •Previous cognitive impairment-IQCODE >57 (n=14) •Without informed consent (n=13) •Non-Spanish speaker (n=3) •Social issues (n=1) •ICU admission from other center (n=8) •Living outside of the hospital’s geographical area (n=6) Participants (n=156) Excluded from neuropsychological assessment (n=56) •ICU readmission (n=3) •ICU mortality (n=27) •Hospital mortality (n=4) •Palliative care (n=1) •No show at appointment/impossible to locate (n=16) •Voluntary reject (n=5) Participants assessed 1 month after ICU discharge (n=100) Excluded (n=8) •Delirium (n=2) •Some cognitive scores could not be calculated due to illiteracy (n=5) •Outlier (Cognitive Index Z =< -15) (n=1) Participants included in the analysis (n=92) Fig. 1 Flowchart representing the distribution of the sample during the different phases of the study participants were cognitively impaired versus the group multivariable model and ‘age’ as a confounding factor. (K3) in which most participants had a normal cognitive Age (OR 1.05; 95% CI 1.00–1.01; p = 0.048), gender (OR status. All variables from the univariate analyses with a 2.81; 95% CI 1.01–7.84; p = 0.048) and cognitive reserve p value < 0.2 (Additional file 1: Table S4) were included (OR 0.37; 95% CI 0.16–0.83; p = 0.016) resulted stati- in a multivariable model. A reduced final multivariable cally significant, while days with opioids ratio presented model was run including the significant factors (gender, a trend toward significance (OR 0.17; 95% CI 0.03–1.08; cognitive reserve and days with opioids) from the initial p = 0.061). Further details of the multivariable regression Fernández‑Gonzalo et al. Crit Care (2020) 24:618 Page 6 of 11 Table 2 Demographic and clinical characteristics domains, instead of the two categories (impaired vs. of the sample non-impaired) differentiated in the classical method by Demographic and clinical variables Total sample (N = 92) Jackson et al. [7]. Furthermore, in two of the three pheno- types, different levels and types of cognitive impairment Age, yr 64 (56–71) were observed in the participants: while the K1 pheno- Female gender 35 (38%) type was characterized by severe alterations in speed of Cognitive Reserve, z-scores 0.04 (− 0.32 to 0.37) processing and executive function, the K2 phenotype was Diagnosis: distinguished by moderate-to-severe deficits in learning Medical 74 (80.4%) and memory retrieval, and impaired speed of process- Surgical 10 (10.9%) ing and executive function. The last phenotype (K3) was Polytrauma 8 (8.7%) characterized by an almost normal performance in most APACHE II at ICU admission, points 17 (13–21) participants on most of the cognitive domains assessed. SOFA at ICU admission, points 7 (5–9) These results are in line with other studies [5, 8, 9], except SOFA slope − 0.9 (− 1.5 to − 0.2) for the low impairment in attention. Charlson Index at ICU admission 3 (2–5) The presence of cognitive alterations in different Length of MV, days 6 (4–10) domains in ICU survivors might suggest different pat- MV days ratio 0.64 (0.44–0.77) terns of brain dysfunction which probably involve vari- Length of delirium, days 1 (0–2) ous pathophysiological mechanisms or pathways. More Delirium ratio 0.037 (0–0.18) specifically, while speed of processing, executive func- Days with sedatives 4 (2–7) tions and memory retrieval impairments are related to Days with sedatives ratio 0.37 (0.18–0.58) dysfunctions in subcortical areas and the frontal lobe, Accumulated dose of sedatives, Eq 3.31 (0.95–8.79) the presence of learning and memory storage problems Days with opioids 4 (2–7.75) points to alterations in the hippocampus and the medial Days with opioids ratio 0.37 (0.20–0.62) temporal lobe. Accumulated dose of opioids, Eq 1.02 (0.23–2.44) Using the classical approach for cognitive impairment, Length of ICU stay, days 10.50 (8–16) almost half of the participants (47%) showed moderate- Length of hospital stay, days 16.00 (10–34.5) to-severe cognitive impairment 1 month after ICU dis- Data are expressed as n (%) or median (IQR), as appropriate charge. This low incidence in comparison with other IQR interquartile range, APACHE Acute Physiology and Chronic Health studies [2, 3, 7] may be related to the patients’ clinical Evaluation, SOFA Sequential Organ Failure Assessment, MV mechanical characteristics, given that our sample presented lower ventilation, Eq equivalents severity of illness and shorter duration of delirium dur- ing ICU stay than samples in other reports. Moreover, models are shown in Additional file 1: Table S5. The while pre-existing cognitive impairment was ruled out in receiver operating characteristic (ROC) curve analysis our patients, in other studies it may have contributed to of the final multivariable regression model is shown in the overall cognitive impairment observed. Interestingly, Additional file 1: Figure S3. when the results of the unsupervised learning machine method were compared with the classical approach, 86% Discussion of the participants were accurately classified between The main finding of this study is the characterization of the two categories of impaired and non-impaired cogni- three different cognitive phenotypes in critically ill-ven- tion. Thus, the unsupervised learning machine method tilated patients 1 month after ICU discharge using the not only allowed detection of cognitive decline but also unsupervised machine learning K-means clustering algo- improved the classification of patients by identifying rithm. The descriptive and exploratory analysis of factors different patterns of cognitive impairment among ICU revealed female gender and older age as potential risk survivors. factors for specific cognitive impairment, while cognitive The three cognitive phenotypes differed in terms of reserve emerged as a protective factor against cognitive several demographic and clinical factors, a circumstance deficit. that may have had an impact on how the clusters were This is the first study to report patterns of cognitive configured. Patients with the K1 phenotype had signifi- impairment after ICU discharge in mechanically venti- cantly fewer days of opioid treatment than patients in K2 lated patients based on an unsupervised machine learn- and K3, and lower accumulated doses of opioids than K3. ing algorithm clustering method. This novel approach Participants in the K2 phenotype were mostly women, allowed us to detect three different phenotypes in the older and had more comorbidities than those in K1 and ICU survivors based on the exploration of six cognitive K3. Moreover, they presented lower accumulated doses Fernández‑Gonzalo et al. Crit Care (2020) 24:618 Page 7 of 11 Fig. 2 Cognitive function distribution in the three patient clusters. The six cognitive indexes are represented according to each patient cluster. Z-scores between 0 and − 1.5 are considered normal for the cognitive function in question, between − 1.5 and − 2 a moderate deficit, and below − 2 a severe deficit of sedative than K3. Finally, participants with the K3 phe- However, we cannot rule out a relation between age and notype showed higher levels of cognitive reserve than K1 gender in our sample, since patients in phenotype K2 and K2. (deficits in memory, speed of processing and executive When the most cognitively affected phenotypes were dysfunction) tended to be female, older and presented combined (K1 and K2), the exploratory analysis of the more comorbidities. If female patients are commonly predisposing and precipitating factors suggested that cer- older, and if aging affects cognitive status in women dif- tain factors may play a more important role than others ferently than in men, it may be that the impact of ICU for the development of cognitive decline after ICU. Spe- stay on cognition in older critically ill patients also differs cifically, female gender, older age and a lower level of cog- between genders. nitive reserve were significantly associated with cognitive In this study, patients in the phenotype with the best impairment. cognitive performance (K3) presented the highest level of Looking at these factors individually, we found that cognitive reserve. Although the analysis is only explora- women were more likely to present cognitive impairment tory, cognitive reserve was found to be a protective factor 1 month after ICU discharge. The role of gender in the against cognitive alterations 1 month after ICU survival. cognitive impairment after ICU has not been specifically Only one previous study [10] has included this concept addressed, and the occasional references in the literature as a predisposing factor for cognitive decline in ICU are contradictory [5, 34, 35]. Our results coincide with patients. Interestingly, cognitive reserve has also been Habib et al. [35] in suggesting that female gender may be identified as a protective factor for cognitive decline in a risk factor for developing post-ICU cognitive impair- healthy older adults [39, 40] and in a wide range of medi- ment, at least early after ICU discharge. However, these cal populations [41–43], including older patients with conclusions should be interpreted with caution: although postsurgical delirium [44]. in healthy populations older women usually perform bet- One of the phenotypes with cognitive impairment ter in verbal memory tests than older men [36], normal (K1) presented significantly fewer days with opioid aging itself entails cognitive deficits, especially memory treatment than the others. Although the ratio between and speed of processing [37]. Previous results for the days with opioids and days of ICU stay only reached impact of age on the cognitive status of ICU survivors are a trend towards significance in the exploratory analy- controversial [3, 4, 9, 25, 34, 35, 38]. Our results corrobo- sis, opioid treatment was the only precipitating factor rate the notion that older critical care patients are more that could be related to cognitive decline in our sam- likely to present cognitive decline after critical illness. ple. It should be borne in mind that opioid treatment Fernández‑Gonzalo et al. Crit Care (2020) 24:618 Page 8 of 11 Fig. 3 Comparison of patients with the two approaches used to analyze global cognitive impairment. The classical approach identifies patients with and without global cognitive impairment, while the unsupervised learning machine method also determines specific cognitive profiles. With the classical definition of global cognitive impairment, the k-means test classified five patients with global cognitive deficit in the K3 cluster, a group in which most patients presented a non-cognitive deficit profile. Moreover, eight patients without global cognitive impairment were grouped in K2, a cluster with a greater presence of patients with cognitive impairment. *Patients incorrectly classified by the unsupervised learning machine approach according to the classical definition of global cognitive impairment improves the welfare and comfort of critically ill Although the analysis was only exploratory, our results patients, enhancing their emotional status. This emo- suggested a higher burden of predisposing factors (such tional well-being related to the management of the as gender, age and cognitive reserve) than precipitat- analgesia and sedation during ICU stay may impact the ing factors in the specific cognitive impairment detected cognitive status of survivors. early after ICU discharge. Duration of delirium was not related to cognitive The current results should be confirmed in future stud- impairment. This may have been due to the notably ies with a higher number of participants. Nevertheless, short duration of delirium in our patients and the low our preliminary findings may serve as a starting-point for inter-subject variability; furthermore, our patients were further research. Of particular interest is the evolution of assessed 1 month after ICU discharge, while in the the cognitive sequelae in the two phenotypes with cog- other studies the impact of delirium on cognition was nitive impairment. Establishing how patients in K1 and assessed at 3 or 12 months. K2 resolve (or maintain) their cognitive deficits might Fernández‑Gonzalo et al. Crit Care (2020) 24:618 Page 9 of 11 Table 3 Demographic and clinical characteristics of the sample according to cluster group Demographic and clinical variables K1 (N = 13) K2 (N = 33) K3 (N = 46) P(< 0.05) Age, yearsa,c 59 (51.5–63.5) 72 (66–78) 60 (50–69) < 0.001 Gender (%) Female gender 3 (23.1) 21 (63.6) 11 (23.9) 0.001 Cognitive reserveb,c − 0.07 (− 0.94 to − 0.05) − 0.11 (− 0.45 to − 0.23) 0.31 (− 0.08 to 0.68) < 0.001 Diagnosis (%) 0.507 Medical 12 (92.3) 27 (81.8) 35 (76.1) Surgical 1 (7.7) 5 (15.2) 4 (8.7) Polytrauma 0 (0.00) 1 (3.0) 7 (15.2) APACHE II at ICU admission, points 15 (12–18) 18 (15.5–22.5) 17 (12–20.25) 0.082 SOFA at ICU admission, points 8 (5.5–10.5) 7 (4–10) 7 (5–9) 0.719 SOFA slope − 1.1 (− 1.8 to − 0.35) − 0.9 (− 1.5 to − 0.25) − 0.75 (− 1.4 to − 0.12) 0.536 Charlson Index at ICU admissiona,c 3 (2.5–4) 5 (3–6) 3 (1–4) < 0.001 Length of MV, days 5 (4.5–10) 5 (3–8) 7 (4–11) 0.222 MV days ratio 0.67 (0.45–0.73) 0.64 (0.38–0.71) 0.63 (0.44–0.8) 0.492 Length of delirium, days 0 (0–1) 1 (0–2) 0 (0–2) 0.451 Delirium ratio 0 (0–0.12) 0.08 (0–0.22) 0 (0–0.10) 0.367 Accumulated dose of sedativesc 4.29 (0.29–16.38) 1.35 (0.35–6.20) 3.85 (1.92–10.62) 0.050 Accumulated dose of opioidsb 0.00 (0.00–2.33) 0.83 (0.29–2.10) 1.45 (0.63–2.52) 0.039 Days with sedatives 4 (1–7) 3 (1–6.5) 4 (3–7.25) 0.178 Days with sedatives ratio 0.44 (0.13–0.64) 0.28 (0.12–0.47) 0.40 (0.22–0.60) 0.163 Days with opioidsa,b 0 (0–5) 4 (2–7) 5 (3–9) 0.016 Days with opioids ratioa,b 0.00 (0.00–0.37) 0.35 (0.22–0.61) 0.5 (0.33–0.64) 0.007 Length of ICU stay, days 11 (7–12.5) 9 (7–18.5) 11 (8–16.25) 0.604 Length of hospital stay after ICU discharge, days 15 (9–30) 17 (11–49) 17.50 (9–31.5) 0.591 Data are expressed as n (%) or median (IQR), as appropriate. IQR interquartile range, APACHE Acute Physiology and Chronic Health Evaluation, SOFA Sequential Organ Failure Assessment, MV mechanical ventilation a Significant difference between K1 versus K2 b Significant difference between K1 versus K3 c Significant difference between K2 versus K3 help to clarify the burden of predisposing factors in long- for performing comparisons with other studies, although term cognitive decline in ICU survivors. Determining it was appropriate for detecting the cognitive phenotype how brain changes associated with aging in both genders in the early stages of the recovery phase. may be impacted by critical illness, and how cognitive reserve may decrease this impact, also deserves further Conclusions investigation. Three cognitive phenotypes were identified in the critically The main limitation of this study is the size of the ill mechanically ventilated survivors 1 month after ICU dis- sample obtained in one of the clusters generated by the charge using the unsupervised machine learning K-means K-means clustering algorithm. This small size hampered clustering algorithm. This approach improved the classical the analysis of the role of the predisposing and precipitat- classification of patients by identifying different patterns ing factors in the three cognitive phenotypes in the ICU of cognitive impairment among ICU survivors. Despite MV survivors. The two phenotypes that included most the exploratory nature of the analysis, female gender, older of their participants with cognitive impairment had to age and low cognitive reserve seemed to play relevant be combined in a single group in order to run the analy- roles as predisposing factors for severe cognitive impair- sis. Nonetheless, the analysis was underpowered and it ment in these patients. It is worth noting that the findings must be considered as exploratory. However, the strict of the factors’ analysis are statistically underpowered, and selection of participants (with control of any previous that further studies should be conducting to obtain defi- cognitive impairment), and the careful statistical analysis nite conclusions. Thus, critical illness, together with the vouch for the accuracy of our conclusions. The optimal predisposing characteristics of each patient, might trigger interval for cognitive assessment may also be a limitation Fernández‑Gonzalo et al. Crit Care (2020) 24:618 Page 10 of 11 different brain dysfunctions the mechanically ventilated References 1. Ehlenbach WJ, Hough CL, Crane PK, Haneuse SJPA, Carson SS, Curtis R, patients at early stages of the ICU recovery. et al. Association between acute care and critical illness hospitalization and cognitive function in older adults. JAMA. 2010;303:763. https​://doi. Supplementary information org/10.1001/jama.2010.167. 2. Girard TD, Jackson JC, Pandharipande PP, Pun BT, Thompson J, Ayumi Supplementary information accompanies this paper at https​://doi. SK, et al. Delirium as a predictor of long-term cognitive impairment org/10.1186/s1305​4-020-03334​-2. in survivors of critical illness. Crit Care Med. 2010;38:1513. https​://doi. org/10.1097/CCM.0b013​e3181​e47be​1. Additional file 1: This additional file contains Tables S1–S5, supplemental 3. Pandharipande PP, Girard TD, Jackson JC, Morandi A, Thompson JL, Pun references and Figures S1–S3. BT, et al. Long-term cognitive impairment after critical illness. N Engl J Med. 2013;369:1306. https​://doi.org/10.1056/NEJMo​a1301​372. 4. Torgersen J, Hole JF, Kvåle R, Wentzel-Larsen T, Flaatten H. Cognitive Acknowledgements impairments after critical illness. Acta Anaesthesiol Scand. 2011;55:1044. The authors thank Michael Maudsley for his invaluable support in editing the https​://doi.org/10.1111/j.1399-6576.2011.02500​.x. manuscript, and Carles Forné for his contribution to the statistical analysis. 5. Hopkins RO, Weaver LK, Collingridge D, Parkinson RB, Chan KJ, Orme JF. Two-year cognitive, emotional, and quality-of-life outcomes in acute Authors’ contributions respiratory distress syndrome. Am J Respir Crit Care Med. 2005;171:340. SFG, CDH, CS, JLA, ARO, RF and LB designed the study; RM, LS and MRM https​://doi.org/10.1164/rccm.20040​6-763OC​. performed the statistical analysis; GGF, CDH, SFG and GNV run the database 6. Iwashyna TJ, Ely EW, Smith DM, Langa KM. Long-term cognitive impair‑ and the eCRF; SFG, GNV, NB and MJ performed the neuropsychological ment and functional disability among survivors of severe sepsis. JAMA. assessments. All authors participated in the interpretation of the results. SFG, 2010;304:1787. https​://doi.org/10.1001/jama.2010.1553. GNV and JFL drafted the manuscript; DJP and MJ provided psychological and 7. Jackson JC, Girard TD, Gordon SM, Thompson JL, Shintani AK, Thomason neuropsychological expertise in designing the protocol. All authors read and JWW, et al. Long-term cognitive and psychological outcomes in the approved the final manuscript. awakening and breathing controlled trial. Am J Respir Crit Care Med. 2010;182:183. https​://doi.org/10.1164/rccm.20090​3-0442O​C. Funding 8. Jackson JC, Hart RP, Gordon SM, Shintani A, Truman B, May L, et al. This work was funded by projects PI13/02204 and PI16/01606, integrated in Six-month neuropsychological outcome of medical intensive care the Plan Nacional de R + D + I, and co-funded by the ISCIII Subdirección Gen‑ unit patients. Crit Care Med. 2003;31:1226. https​://doi.org/10.1097/01. eral de Evaluación y el Fondo Europeo de Desarrollo Regional (FEDER). CIBER CCM.00000​59996​.30263​.94. Enfermedades Respiratorias, Fundació Parc Taulí-I3PT, (Spain). 9. Jones C, Griffiths RD, Slater T, Benjamin KS, Wilson S. Significant cognitive dysfunction in non-delirious patients identified during and persisting Availability of data and materials following critical illness. Intensive Care Med. 2006;32:923. https​://doi. The datasets used and analyzed during this study are available from the cor‑ org/10.1007/s0013​4-006-0112-y. responding author on reasonable request. 10. Larson MJ, Weaver LK, Hopkins RO. Cognitive sequelae in acute respira‑ tory distress syndrome patients with and without recall of the intensive Ethics approval care unit. J Int Neuropsychol Soc. 2007;13:595. https​://doi.org/10.1017/ The institutional review boards of each participating ICU approved the proto‑ S1355​61770​70707​49. col and the informed consent documentation. 11. Mikkelsen ME, Christie JD, Lanken PN, Biester RC, Thompson BT, Bellamy SL, et al. The adult respiratory distress syndrome cognitive outcomes Consent for publication study. Am J Respir Crit Care Med. 2012;185:1307. https​://doi.org/10.1164/ All participants consent for the publication of data generated in this study. rccm.20111​1-2025O​C. 12. Schulte PJ, Warner DO, Martin DP, Deljou A, Mielke M, Knopman D, et al. Competing interests Association between critical care admissions and cognitive trajectories The authors declare that they have no competing interests. in older adults. Crit Care Med. 2019;47:1116. https​://doi.org/10.1097/ CCM.00000​00000​00382​9. Author details 13. Guerra C, Linde-Zwirble WT, Wunsch H. Risk factors for dementia after 1 Critical Care Center, Parc Taulí Hospital Universitari, Fundació‑ I3PT, UAB, critical illness in elderly Medicare beneficiaries. Crit Care. 2012;16:R233. Sabadell, Spain. 2 Centro de Investigación Biomédica En Red en Salud https​://doi.org/10.1186/cc119​01. Mental (CIBERSAM), Instituto de Salud Carlos III, Madrid, Spain. 3 Depart‑ 14. Iwashyna TJ, Netzer G. The burdens of survivorship: an approach to think‑ ment of Clinical and Health Psychology, Universitat Autònoma de Barcelona, ing about long-term outcomes after critical illness. Semin Respir Crit Care Bellaterra, Barcelona, Spain. 4 Centro de Investigación Biomédica En Red en Med. 2012;33:327. https​://doi.org/10.1055/s-0032-13219​82. Enfermedades Respiratorias (CIBERES), Instituto de Salud Carlos III, Madrid, 15. Langa KM, Chernew ME, Kabeto MU, Herzog AR, Ofstedal MB, Willis RJ, Spain. 5 Mental Health Department, Fundació Althaia - Xarxa Assistencial I et al. National estimates of the quantity and cost of informal caregiving Universitaria, Manresa, Spain. 6 Critical Care Center, Fundació Althai, Universitat for the elderly with dementia. J Gen Intern Med. 2001;16:770. https​://doi. Internacional de Catalunya, Manresa, Spain. 7 Better Care S.L., Barcelona, org/10.1111/j.1525-1497.2001.10123​.x. Spain. 8 Department of Medicine, Universitat de Barcelona, Barcelona, Spain. 16. Rockwood K, Brown M, Merry H, Sketris I, Fisk J. Societal costs of vascular 9 Neurology Department, Parc Taulí Hospital Universitari, I3PT, UAB, Sabadell, cognitive impairment in older adults. Stroke. 2002;33:1605. https​://doi. Spain. 10 Departament of Basic Medical Sciences, Universitat de Lleida, Lleida, org/10.1161/01.str.00000​17878​.85274​.44. Spain. 11 Health Services Research Network in Chronic Diseases (REDISSEC), 17. Wolters AE, Slooter AJC, van der Kooi AW, van Dijk D. Cognitive impair‑ Barcelona, Spain. 12 Mental Health Department, Parc Taulí Hospital Universitari, ment after intensive care unit admission: a systematic review. Intensive I3PT, UAB, Sabadel, Spain. 13 Department of Psychiatry and Forensic Medicine, Care Med. 2013;39:376. https​://doi.org/10.1007/s0013​4-012-2784-9. Universitat Autònoma de Barcelona, Bellaterra, Barcelona, Spain. 18. Wilcox ME, Brummel NE, Archer K, Wesley EE, James J, Hopkins R. Cogni‑ tive dysfunction in ICU patients: risk factors, predictors, and rehabilita‑ Received: 22 April 2020 Accepted: 6 October 2020 tion interventions. Crit Care Med. 2013;41:S81. https​://doi.org/10.1097/ CCM.0b013​e3182​a1694​6. 19. Fernandez-Gonzalo S, Turon M, De Haro C, Lopez-Aguilar J, Jodar M, Blanch L. Do sedation and analgesia contribute to long-term cognitive dysfunction in critical care survivors? Med intensiva. 2018;42:114. https​:// doi.org/10.1016/j.medin​.2017.06.010. Fernández‑Gonzalo et al. Crit Care (2020) 24:618 Page 11 of 11 20. Hopkins RO, Suchyta MR, Snow GL, Jephson AI, Weaver LK, Orme JF. 33. Livon D, Xie S, Eppig J, Wicas G, Lamar M, Lippa C, et al. The heterogeneity Blood glucose dysregulation and cognitive outcome in ARDS survivors. of mild cognitive impairment: A neuropsychological analysis. J Int Neu‑ Brain Inj. 2010;24:1478. https​://doi.org/10.3109/02699​052.2010.50686​1. ropsych Soc. 2010;16:84–93. https​://doi.org/10.1017/S1355​61770​99909​ 21. Duning T, van den Heuvel I, Dickmann A, Volker T, Wempe C, Reinholz 93. J, et al. Hypoglycemia aggravates critical illness-induced neurocogni‑ 34. Karnatovskaia LV, Schulte PJ, Philbrick KL, Johnson MM, Anderson BK, tive dysfunction. Diabetes Care. 2010;33:639. https​://doi.org/10.2337/ Gajic O, et al. Psychocognitive sequelae of critical illness and correla‑ dc09-1740. tion with 3 months follow up. J Crit Care. 2019;52:166. https​://doi. 22. Sakusic A, Gajic O, Singh TD, O’Horo J, Jenkins G, Wilson G, et al. Risk org/10.1016/j.jcrc.2019.04.028. factors for persistent cognitive impairment after critical illness. Nested 35. Habib S, Khan AR, Afridi MI, Saeed A, Jan AF, Amjad N. Frequency and Case-Control Study Crit Care Med. 2018;46:1977. https​://doi.org/10.1097/ predictors of cognitive decline in patients undergoing coronary artery CCM.00000​00000​00339​5. bypass graft surgery. J Coll Physicians Surg Pak. 2014;24:543. 23. Davydow DS, Zatzick D, Hough CL, Katon WJ. In-hospital acute stress 36. Munro CA, Winicki JM, Schretlen DJ, Gower EW, Turano KA, Muñoz B, et al. symptoms are associated with impairment in cognition 1 year after inten‑ Sex differences in cognition in healthy elderly individuals. Neuropsy‑ sive care unit admission. Ann Am Thorac Soc. 2013;10:450. https​://doi. chol Dev Cogn B Aging Neuropsychol Cogn. 2012;19:759. https​://doi. org/10.1513/Annal​sATS.20130​3-060OC​. org/10.1080/13825​585.2012.69036​6. 24. Gunther ML, Morandi A, Krauskopf E, Pandharipande P, Timothy GD, 37. Caselli RJ, Dueck AC, Locke DEC, Baxter LC. Sex-based memory James JC, et al. The association between brain volumes, delirium dura‑ advantages and cognitive aging: a challenge to the cognitive reserve tion, and cognitive outcomes in intensive care unit survivors: the VISIONS construct? J Int Neuropsychol Soc. 2015;21:95. https​://doi.org/10.1017/ cohort magnetic resonance imaging study*. Crit Care Med. 2012;40:2022. S1355​61771​50000​16. https​://doi.org/10.1097/CCM.0b013​e3182​50acc​0. 38. Sakusic A, O’Horo JC, Dziadzko M, Volha D, Ali R, Singh TD, et al. Potentially 25. Hope AA, Morrison RS, Du Q, Wallenstein S, Nelson JE. Risk factors for modifiable risk factors for long-term cognitive impairment after critical long-term brain dysfunction after chronic critical illness. Ann Am Thorac illness: a systematic review. Mayo Clin Proc. 2018;93:68. https​://doi. Soc. 2013;10:315. https​://doi.org/10.1513/Annal​sATS.20121​1-099OC​. org/10.1016/j.mayoc​p.2017.11.005. 26. van den Boogaard M, Schoonhoven L, Evers AWM, van der Hoeve JG, 39. Clare L, Wu Y-T, Teale JC, MacLeod C, Matthews F, Brayne C, et al. Poten‑ van Achterberg T, Peter P, et al. Delirium in critically ill patients: impact tially modifiable lifestyle factors, cognitive reserve, and cognitive function on long-term health-related quality of life and cognitive functioning. Crit in later life: a cross-sectional study. PLOS Med. 2017;14:1–14. https​://doi. Care Med. 2012;40:112. https​://doi.org/10.1097/CCM.0b013​e3182​2e9fc​9. org/10.1371/journ​al.pmed.10022​59. 27. Jackson JC, Archer KR, Bauer R, Abraham CM, Song Y, Greevey R, Guil‑ 40. Roldán-Tapia MD, Cánovas R, León I, García-Garcia J. Cognitive vulner‑ lamondegui O, et al. A prospective investigation of long-term cognitive ability in aging may be modulated by education and reserve in healthy impairment and psychological distress in moderately versus severely people. Front Aging Neurosci. 2017;9:1. https​://doi.org/10.3389/fnagi​ injured trauma intensive care unit survivors without intracranial hemor‑ .2017.00340​. rhage. J Trauma. 2011;71:860. https​://doi.org/10.1097/TA.0b013​e3182​ 41. Amoretti S, Cabrera B, Torrent C, Mezquida G, Lobo A, González-Pinto A, 15196​1. et al. Cognitive reserve as an outcome predictor: first-episode affective 28. Stern Y. Cognitive reserve in ageing and Alzheimer’s disease. Lancet versus non-affective psychosis. Acta Psychiatr Scand. 2018;138:441. https​ Neurol. 2012;11:1006. https​://doi.org/10.1016/S1474​-4422(12)70191​-6. ://doi.org/10.1111/acps.12949​. 29. Morales Gonzalez J, Gónzalez-Montalvo J, Del Ser QT, Bermejo PF. Valida‑ 42. Lucero C, Campbell MC, Flores H, Maiti B, Perlmutter JS, Foster ER. Cogni‑ tion of the S-IQCODE: the Spanish version of the informant questionnaire tive reserve and β-amyloid pathology in Parkinson disease. Parkinsonism on cognitive decline in the elderly. Arch Neurobiol (Madr). 1992;55:262. Relat Disord. 2015;21:899. https​://doi.org/10.1016/j.parkr​eldis​.2015.05.020. 30. de Haro C, Magrans R, López-Aguilar J, Montanyà J, Lena E, Subirpa C, 43. Soldan A, Pettigrew C, Cai Q, Wang J, Wang MC, Moghekar A, et al. et al. Effects of sedatives and opioids on trigger and cycling asynchronies Cognitive reserve and long-term change in cognition in aging and throughout mechanical ventilation: an observational study in a large preclinical Alzheimer’s disease. Neurobiol Aging. 2017;60:164. https​://doi. dataset from critically ill patients. Crit Care. 2019;23:245. https​://doi. org/10.1016/j.neuro​biola​ging.2017.09.002. org/10.1186/s1305​4-019-2531-5. 44. Tow A, Holtzer R, Wang C, Sharan A, Kim SJ, Gladstein A, et al. Cognitive 31. MacPherson SE, Healy C, Allerhand M, Spanò B, Tudor-Sfetea C, White M, Reserve and Postoperative Delirium in Older Adults. J Am Geriatr Soc. et al. Cognitive reserve and cognitive performance of patients with focal 2016;64:1341. https​://doi.org/10.1111/jgs.14130​. frontal lesions. Neuropsychologia. 2017;96:19. https​://doi.org/10.1016/j. neuro​psych​ologi​a.2016.12.028. 32. Sanchez-Pinto LN, Luo Y, Churpek MM. Big data and data science in Publisher’s Note critical care. Chest. 2018;154:1239. https​://doi.org/10.1016/j.chest​ Springer Nature remains neutral with regard to jurisdictional claims in pub‑ .2018.04.037. lished maps and institutional affiliations. Ready to submit your research ? Choose BMC and benefit from: • fast, convenient online submission • thorough peer review by experienced researchers in your field • rapid publication on acceptance • support for research data, including large and complex data types • gold Open Access which fosters wider collaboration and increased citations • maximum visibility for your research: over 100M website views per year At BMC, research is always in progress. Learn more biomedcentral.com/submissions