CLINICAL RESEARCH e-ISSN 1643-3750 © Med Sci Monit, 2019; 25: 1994-2001 DOI: 10.12659/MSM.913283 Received: Accepted: 2018.09.19 2018.10.16 Sociodemographic Indicators of Health Status Published: 2019.03.17 Using a Machine Learning Approach and Data from the English Longitudinal Study of Aging (ELSA) Authors’ Contribution: ABCDEF 1 Worrawat Engchuan 1 The Centre for Applied Genomics, Genetics and Genome Biology, The Hospital for Study Design A ABDF 2,3 Alexandros C. Dimopoulos Sick Children, Toronto, ON, Canada Data Collection B 2 Department of Nutrition and Dietetics, School of Health Science and Education, Statistical Analysis C AEF 2,4,5,6 Stefanos Tyrovolas Harokopio University, Athens, Greece Data Interpretation D BDEF 7,8 Francisco Félix Caballero 3 Department of Informatics and Telematics, School of Digital Technology, Manuscript Preparation E DEF 4,5,6 Albert Sanchez-Niubo Harokopio University, Athens, Greece Literature Search F 4 Research, Innovation and Teaching Unit, Institut de Recerca Sant Joan de Déu, Funds Collection G ABCDEFG 9 Holger Arndt Carrer Dr. Antoni Pujadas, Barcelona, Spain EF 6,10,11 Jose Luis Ayuso-Mateos 5 Parc Sanitari Sant Joan de Déu, Universitat de Barcelona, Fundació Sant Joan de DEF 4,5,6 Josep Maria Haro Déu, Sant Boi de Llobregat, Barcelona, Spain 6 Instituto de Salud Carlos III, Centro de Investigación Biomédica en Red de Salud DEF 12 Somnath Chatterji Mental (CIBERSAM), Madrid, Spain ABDEFG 2,13 Demosthenes B. Panagiotakos 7 Department of Preventive Medicine and Public Health, and Microbiology, Universidad Autónoma de Madrid, Madrid, Spain 8 Instituto de Salud Carlos III, CIBER of Epidemiology and Public Health (CIBERESP), Madrid, Spain 9 Spring Techno Gmbh and Co. KG, Bremen, Germany 10 Hospital Universitario de La Princesa, Instituto de Investigación Sanitaria Princesa (IP), Madrid, Spain 11 Department of Psychiatry, Universidad Autónoma de Madrid, Madrid, Spain 12 Health Metrics and Measurement, World Health Organization, Geneva, Switzerland 13 Faculty of Health, University of Canberra, Canberra, Australia Corresponding Author: Demosthenes B. Panagiotakos, e-mail: dbpanag@hua.gr Source of support: The ATHLOS project has received funding from the European Union Horizon 2020 Research and Innovation Program under grant agreement No. 635316 (EU HORIZON2020-PHC-635316) Background: Studies on the effects of sociodemographic factors on health in aging now include the use of statistical mod- els and machine learning. The aim of this study was to evaluate the determinants of health in aging using ma- chine learning methods and to compare the accuracy with traditional methods. Material/Methods: The health status of 6,209 adults, age <65 years (n=1,585), 65–79 years (n=3,267), and >80 years (n=1,357) were measured using an established health metric (0–100) that incorporated physical function and activities of daily living (ADL). Data from the English Longitudinal Study of Ageing (ELSA) included socio-economic and sociodemographic characteristics and history of falls. Health-trend and personal-fitted variables were gener- ated as predictors of health metrics using three machine learning methods, random forest (RF), deep learning (DL) and the linear model (LM), with calculation of the percentage increase in mean square error (%IncMSE) as a measure of the importance of a given predictive variable, when the variable was removed from the model. Results: Health-trend, physical activity, and personal-fitted variables were the main predictors of health, with the%incMSE of 85.76%, 63.40%, and 46.71%, respectively. Age, employment status, alcohol consumption, and household income had the%incMSE of 20.40%, 20.10%, 16.94%, and 13.61%, respectively. Performance of the RF method was similar to the traditional LM (p=0.7), but RF significantly outperformed DL (p=0.006). Conclusions: Machine learning methods can be used to evaluate multidimensional longitudinal health data and may pro- vide accurate results with fewer requirements when compared with traditional statistical modeling. MeSH Keywords: Artificial Intelligence • Data Interpretation, Statistical • Decision Support Techniques • Socioeconomic Factors Full-text PDF: https://www.medscimonit.com/abstract/index/idArt/913283 3155   1   4   31 Indexed in: [Current Contents/Clinical Medicine] [SCI Expanded] [ISI Alerting System] This work is licensed under Creative Common Attribution- NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0) 1994 [ISI Journals Master List] [Index Medicus/MEDLINE] [EMBASE/Excerpta Medica] [Chemical Abstracts/CAS] Engchuan W. et al.: Health status using machine learning and ELSA © Med Sci Monit, 2019; 25: 1994-2001 CLINICAL RESEARCH Background Material and Methods As the global population is getting older, the study of health This study was conducted according to the Strengthening the and aging has become increasingly important, particularly in Reporting of Observational Studies in Epidemiology (STROBE) planning current and future healthcare resources. People of Statement [24]. 65 years of age and older now represent an increasing pro- portion of the population, particularly in Europe, Asia, and Working dataset the USA [1–6]. Changing age demographics poses a dramatic shift towards an increased health burden of non-communi- The English Longitudinal Study of Ageing (ELSA) dataset was cable diseases and disability [7,8]. Therefore, a current pub- used to test the research hypothesis, which included six waves lic health challenge is to identify health-related factors and of longitudinal data from 6,209 adults [21]. Only subjects who to understand how to maintain a healthy life with increasing had at least five waves of longitudinal data were used. The age. Sociodemographic factors, which include the employment number of individuals included in the study analysis was 6,209. status, household income, level of education, marital status, The health metric in the last wave of data collection for each and social support are recognized major determinants of many subject was treated as the target health metric to be predicted. health outcomes, including healthy aging [9,10]. The sociodemographic indicators used included; gender, Several analytical models have been proposed to evaluate men (n=2,638) and women (n=3,571); age group <65 years healthy aging in relation to lifestyle characteristics, as well (n=1,585), 65–79 years (n=3,267), >80 years (n=1,357); quin- as biological, genetic and clinical factors, based on classi- tiles of household income; formal education (yes or no); mari- cal statistical hypothesis testing [11–13]. However, residual tal status, (married, single, and other); smoking history (never confounding and unexplained health risks are a common prob- smoked, former smoker, current smoker); alcohol consumption lem in almost all these hypothesis-derived models. Recently, (non-drinker, drinking alcohol twice weekly or less, drinking the use of health informatics has received increasing atten- a regular glass of wine glass or equivalent of 12% alcohol more tion as it allows for collection and analysis of large amounts of than twice weekly); physical activity (inactive, moderately ac- data and can extract data on patterns of risk that are free from tive, and active); employment status (employed or unemployed); the strict methodological assumptions of traditional statisti- size of social network (number of relatives and close relation- cal modeling methods [14,15]. In particular, machine learning ships, small <5 people, moderate 5–9 people, large >9 people). offers a data-driven approach to allow for the analysis of pat- terns of health-associated variables and can provide insight In addition, to sociodemographic data, a history of falls was into data without forming an a priori defined hypothesis re- also incorporated as a possible confounder. More informa- garding the involved variables [16–19]. There are several ma- tion on discretization of each characteristics can be found in chine learning algorithms that are used to analyze health data, a previous paper [21]. The health metric was developed using including support vector machine, decision tree, random forest item response theory (IRT) using questionnaire data on indi- (RF), the linear model (LM), and more recently, deep learning vidual physical functionality, activities of daily living (ADL), (DL) [20]. Choosing a machine algorithm for a particular ana- and instrumental activities of daily living (IADL), as previously lytical problem is important, as these models have rarely been described [23]. The range of the metric was set to 0–100. compared before in terms of their efficiency and accuracy. Figure 1 shows the distribution of the health metric, strati- fied according to each of the studied sociodemographic char- The present study was part of the Ageing Trajectories of acteristics. The increasing trend of the health metric over the Health: Longitudinal Opportunities and Synergies (ATHLOS) increasing value of the sociodemographic characteristic, sug- project (http://athlosproject.eu/). The aims of this study were gests a positive relationship. to evaluate the sociodemographic determinants of healthy aging using three machine learning methods, RF, DL, and LR, Feature engineering to develop the new predictors of the and to compare these methods in terms of their efficiency. personal-fitted variable and the health trend variable The working dataset was the English Longitudinal Study of Ageing (ELSA), which includes six waves of longitudinal data Feature engineering is the process of generating a new pre- from 6,209 adults, from between 2002–2012 [21,22]. A pre- dictor that is higher order and more meaningful than existing viously developed and validated health metric of aging was ones [25]. The sociodemographic characteristics and health used as the outcome, which was developed based on char- metrics from previous four waves of data were used to gener- acteristics including physical function, activities of daily living ate two new predictors by modeling the relationship between (ADL), and instrumental activities of daily living (IADL) [21,23]. predictors, and also the trend of the health metric. The first type of new predictor was a personal-fitted variable, which Indexed in: [Current Contents/Clinical Medicine] [SCI Expanded] [ISI Alerting System] This work is licensed under Creative Common Attribution- NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0) 1995 [ISI Journals Master List] [Index Medicus/MEDLINE] [EMBASE/Excerpta Medica] [Chemical Abstracts/CAS] Engchuan W. et al.: CLINICAL RESEARCH Health status using machine learning and ELSA © Med Sci Monit, 2019; 25: 1994-2001 100 100 100 100 75 75 75 75 Health metric Health metric Health metric Health metric 50 50 50 50 25 25 25 25 Female Male <65 65–79 >79 Married Never Divorce/Widowed Sex Age group Age group Education 100 100 100 100 75 75 75 75 Health metric Health metric Health metric Health metric 50 50 50 50 25 25 25 25 Not in work In work Inactive Moderate Active Q5 Q4 Q3 Q2 Q1 Employment Physical activity Household wealth Social network size 100 100 100 75 75 75 Health metric Health metric Health metric 50 50 50 25 25 25 No twice/week Inactive Moderate Active No Yes Employment Physical activity History of falls Figure 1. Global health metric distribution stratified by each value in each predictor. The boxplots show the distribution of the health metrics stratified by the unique value of the predictors. The difference in the distribution within the predictor suggests a relationship between the predictor and the health metric. captured the relationship between the sociodemographic data The second type of new predictor was the health trend variable, and the health metric for each subject. An individual health which captured the trend of the health metric for an individual metric prediction model was built using data from the previ- subject. A health metric prediction model was built by fitting ous waves, as shown in model 1. The variable was then gen- the health metric with the time from the previous four waves erated as a prediction from the model on the current wave of of data, as shown in model 2. The variable was generated as data. A linear model (LM) was used to generate the new vari- a prediction of the model at the current time point. Another able (model 1). linear model was used to generate the variable (model 2). health metric ~ gender + age group + quintile of household health metric ~ time (model 2) wealth + formal education + marital status + falls + smoking behavior + alcohol consumption + physical activity + employ- In total, there were 13 predictors, of which 11 were sociode- ment + size of social network (model 1) mographic determinants, one personal-fitted variable, and one trend variable. The notation of the linear model (model 1) is described as the regressing gender, age group, quintile of household wealth, formal education, marital status, fall, smoking behavior, alco- hol consumption, physical activity, employment and size of so- cial network on health metric. Indexed in: [Current Contents/Clinical Medicine] [SCI Expanded] [ISI Alerting System] This work is licensed under Creative Common Attribution- NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0) 1996 [ISI Journals Master List] [Index Medicus/MEDLINE] [EMBASE/Excerpta Medica] [Chemical Abstracts/CAS] Engchuan W. et al.: Health status using machine learning and ELSA © Med Sci Monit, 2019; 25: 1994-2001 CLINICAL RESEARCH NTree parameter optimization Mtry parameter optimization 70 100 90 65 80 60 MSE MSE 70 55 60 50 0 500 1000 1500 2000 10 20 30 ntree mtry Figure 2. Parameters optimization by grid search. The left panel represents the mean squared error (MSE) change over different values of the number of trees used to build a random forest model (ntree). The right panel represents the MSE change for the different values of the number of predictors randomly picked at each branch of the tree (mtry). The optimal parameters (ntree=500) (mtry=15) were used in the final model. Health metric prediction using random forest (RF), linear activation function, 5% neurons dropout, 1 output neuron). The model (LM) and deep learning (DL) models parameters used to build the model were optimized using the Adam optimization algorithm using a batch size of 128 (de- Random forest (RF), linear model (LM), and deep learning (DL) fault). The size of the validation set was 20% of the training set. were applied to evaluate sociodemographic determinants of the health metric. RF is an ensemble learner of decision tree The performance assessment was undertaken using 10-fold (DT), which identifies a predictor by majority data. RF was im- cross-validation to robustly estimate the performance by using plemented in this study by using the RF library of R [26]. Two ten equal random subsets, or folds, of the whole dataset. One RF parameters were ntree, the number of trees used to build of the folds was used as a test set and the remainder were used a random forest model, and mtry, the number of predictors as a training set. The process was repeated ten times until all randomly picked at each branch of the tree. Both ntree and folds were used as a test set, then the final result was reported. mtry were optimized in the training data by undertaking grid search (Figure 2). The optimal parameters (ntree=500, and Assessment of predictor performance mtry=15) were used in the final model. Understanding which predictors in the model are driving the LM is a statistical approach that builds a learning model by fit- performance and how they are doing so can help to adjust ting beta coefficients of the predictors on linear relationships the model, so that the performance can be improved. The ef- between them and a target class. LM is a simple and fast ap- fect of each predictor was assessed by obtaining the stan- proach to build a model. However, as its name suggests, this dardized coefficient of the LM model by scaling each predic- approach only performs well when the problem has a linear re- tor using the scale function of the base library in R, and the lationship. The R statistical function, lm, of the R Stats Package variable importance of RF was assessed using the importance was used to perform linear regression in this study [27]. function of the randomForest library in R [26]. The magnitude of the standardized coefficient of the LM model represented DL is an established algorithm that mimics a biological neural the effect size of a predictor on the health metric, while the network. Multiple cascading layers of neurons are connected sign represented the direction of the relationship. The vari- and pass information from one layer to the next by transforming able importance of RF assessed the effect of a predictor by re- and extracting new predictors. DL has been shown to be supe- moving the predictor from every single tree in the forest and rior to other algorithms in many applications [20]. This study measuring how the accuracy changed. The effect was reported implemented DL using keras library in R [28]. The structure of as the percentage increase in mean square error (%IncMSE), the DL model consisted of five layers (13 neurons with relu ac- where more important predictors have higher%incMSE [26]. tivation function, 10% neurons dropout, 5 neurons with linear Indexed in: [Current Contents/Clinical Medicine] [SCI Expanded] [ISI Alerting System] This work is licensed under Creative Common Attribution- NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0) 1997 [ISI Journals Master List] [Index Medicus/MEDLINE] [EMBASE/Excerpta Medica] [Chemical Abstracts/CAS] Engchuan W. et al.: CLINICAL RESEARCH Health status using machine learning and ELSA © Med Sci Monit, 2019; 25: 1994-2001 A Personal-fifted vs. health metric, R=0.62 B Personal-fifted vs. health metric, R=0.81 100 100 75 75 Personal-fifted Health trend 50 50 25 25 0 0 0 25 50 75 100 0 25 50 75 100 Health metric Health metric C D 2000 200 Variable 11 predictors 1500 150 All predictors 11 predictors 1000 100 SE 500 50 0 0 0 500 1000 1500 2000 11 predictors All predictors All predictors Predictors Figure 3. Contribution of historical, personal-fitted, and health trend features. The scatter plots (A–C) illustrate the relationship between the personal-fitted predictor (A), health trend predictor (B), the prediction from 11 predictors (C), and the health metric. The boxplot (D) shows the squared errors (SE) from 11 predictors and all predictors. Results 40±17.34 in the non-active, 59.21±14.95 in the moderately ac- tive, and 68±10.01 in the active population (p<0.001). The so- Sociodemographic factors and health metrics cial health metrics were 62.41±15.53 for people with a small social network to 66.77±11.69 for people with a large social The health metric varied from 65.95±13.56 in men to network (p<0.001). The health metrics were 65.70±13.68 for 62.83±5.01 in women (p<0.001), 68.92±11.31 in people <65 people without a history of falls and 59.84±15.79 with a his- years, 65.75±12.79 in people between 65–75 years, and tory of falls (p<0.001). 54.73±17.24 in people >79 years (p<0.001). The health met- rics were 56.36±17.05, 61.41±15.22, 64.97±13.02, 66.84±11.98, Sociodemographics, personal-fitted, and health trend as and 69.60±9.89, respectively in the 1st to 5th quintiles of house- health predictors hold wealth, (p<0.001); 66.24±13.06 in educated people, to 59.05±16.37 in uneducated people (p<0.001); 66.68±12.65 in Before comparing different prediction models, validation that married people, 63.50±14.91 in single people, and 59.30±16.46 the newly extracted variables, personal-fitted, and health in previously married people (p<0.001). The health metrics were trend, added information to the model. The scatter plots pre- 65.40±14.10 in non-smokers, 63.72±14.56 in former smokers, sented in Figure 3A and 3B show the correlation between and 62.23±15.13 in current smokers (p<0.001). The health met- personal-fitted, health trend, and health metric, respectively. rics were 55.91±18.73 in non-alcohol drinkers, 66.18±11.93 in The health trend variable was more closely correlated (0.81) people who drank twice or less each week, and 68.86±9.56 in with the health metric than the personal-fitted variable (0.62). those who drank more than twice a week of a regular glass of Figure 3C shows the square error (SE) of the model with so- wine or the equivalent of 12% alcohol (p<0.001). The health ciodemographic and extracted variables. Having more vari- metrics were 72.21±5.52 in those who were employed and ables changed the performance in most cases, which also 62.56 in the unemployed (p<0.001). The health metrics were showed improvement in performance (more values in the Indexed in: [Current Contents/Clinical Medicine] [SCI Expanded] [ISI Alerting System] This work is licensed under Creative Common Attribution- NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0) 1998 [ISI Journals Master List] [Index Medicus/MEDLINE] [EMBASE/Excerpta Medica] [Chemical Abstracts/CAS] Engchuan W. et al.: Health status using machine learning and ELSA © Med Sci Monit, 2019; 25: 1994-2001 CLINICAL RESEARCH A p-value=2.2e-16 B 4000 50 p-value=0.006 40 3000 p-value=0.7 30 2000 SE SE 20 1000 10 0 0 Random DL LM RF Random DL LM RF Model Model Figure 4. Performance comparison between three prediction models and random prediction. (A) Box plots show the distribution of squared errors of the random prediction model, deep learning (DL), the linear model (LM), and the random forest (RF) model. (B) A magnified version of (A) shows the difference between the DL, LM, and RF. Student’s t-test was used to calculate the P-values. upper part of the diagonal line) (Figure 3D). From the distribu- by%incMSE and the standardized coefficient were aligned, ex- tion of SEs, the model with the new variables was significantly cept for social network size characteristics which were less im- lower than the model with only sociodemographic character- portant than age, employment status, alcohol consumption, istics (p<0.001) [19]. and household income. Comparison of the machine learning prediction methods The results of the predictor importance assessment confirmed the hypothesis that having health trend and personal-fitted To determine whether one model was better than another, the variables would help improve the performance of the model. mean square error (MSE) was calculated [29]. The higher the Also, the effects of most of the sociodemographic characteristics MSE the model had, the worse was the performance of the found in a previous study [21], were confirmed by these findings. model. Also random prediction was done as a baseline. The random prediction was done by label permutation to main- tain the health metric distribution. From 10-fold cross-valida- Discussion tion, RF, LM and DL were much better than random prediction (Figure 4A, 4B). The best model was RF, with MSE of 51.11, This study aimed to build a predictive model to accurately while LM, DL and random prediction had a higher MSE of 52.07, estimate health status based on sociodemographic charac- 59.08 and 418.40, respectively. Figure 4 shows the performance teristics in an aging population using data from the English of each model by the distribution of their SEs. The SEs of RF Longitudinal Study of Ageing (ELSA), which included socio- were significantly lower when compared with the random economic and sociodemographic characteristics and history predictions (P<0.001), and also significantly lower when com- of falls. Because the dataset analyzed was quite large, con- pared with DL (P=0.006), but were comparable with LM (P=0.7). sisting of more than 6,000 samples, this allowed the analysis of many potential predictive factors to obtain the best health Predictor importance assessment metric prediction model. However, although the sample size was quite large, the number of time points was very limited. To understand the model, variable importance from RF and the As a result, the model was unable to recognize the changes standardized coefficient from LM were required. The health in health patterns over time, and also many time series data trend, physical activity, and personal-fitted variables were the analytical approaches could not be applied. main predictors of health metrics with%incMSE of 85.76%, 63.40%, and 46.71%, respectively. Age, employment status, This study included three machine learning methods, random alcohol consumption, and household income were also the forest (RF), deep learning (DL) and the linear model (LM), with main characteristics that help determine the health metric calculation of the percentage increase in mean square error with%incMSE of 20.40%, 20.10%, 16.94%, and 13.61%, respec- (%IncMSE) as a measure of the importance of a given predic- tively. Table 1 shows all the%incMSE and the standardized co- tive variable, when the variable was removed from the model. efficients of the predictors. The overall ranking of predictors An advantage of using RF was that it was transparent and the Indexed in: [Current Contents/Clinical Medicine] [SCI Expanded] [ISI Alerting System] This work is licensed under Creative Common Attribution- NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0) 1999 [ISI Journals Master List] [Index Medicus/MEDLINE] [EMBASE/Excerpta Medica] [Chemical Abstracts/CAS] Engchuan W. et al.: CLINICAL RESEARCH Health status using machine learning and ELSA © Med Sci Monit, 2019; 25: 1994-2001 Table 1. Summarized %incMSE and coefficients by predictors. Predictors %incMSE Standardized coefficients Health trend (health metric estimated by 4 previous health metric) 85.76 8.13 Physical activity (active vs. moderate vs. inactive) 63.40 3.30 Personal-fitted variable (health metric estimated by 11 socio-demographics and 46.71 1.54 fall history of previous 4 waves) Age groups (<65 vs. 65–79 vs. >79) 20.40 -0.61 Employment (in work vs. not in work) 20.10 0.51 Alcohol consumption 16.94 0.74 Quantiles Household Wealth (Q1–Q5) 13.61 0.46 Social network size (<5 vs. 5–9 vs. >9 people) 7.16 1.22 Falls (have fall history vs. no fall history) 5.30 –0.60 Marital status (married vs. never married vs. others) 5.29 –0.03 Smoke (never vs. former smoker vs. current smoker) 4.54 –0.19 Sex (males vs. females) 4.35 0.32 Education (no qualification vs. some formal education) 2.86 0.05 importance of the variables could be assessed. DL has previ- relationship between the changes of the sociodemographic and ously been reported to outperform state-of-the-art algorithms, health metrics might also help to reduce the error rate in cases but this was not the case for this dataset, which might have where the current health metric suddenly changed. However, been due to the simplicity of its implementation, as the opti- the change in the current health metric needs to be the prod- mization was performed as recommended by Keras [28]. The uct of a change in some of the sociodemographic characteris- choice of optimizers is also very important, and there are dif- tics to be able to be captured by the personal-fitted variable. ferent optimizers that may help to improve the networks, such If there is a change in the health metric from anything other as stochastic gradient descent (SGD), AdaGrad, and RMSProp, than factors in the captured sociodemographic characteristics, so further studies are needed to identify the best optimizer to the health trend variable might be helpful. adjust network structure and parameters [30]. A further rea- son for the poor performance of DL might be due to the lim- However, in this study, because the number of time points was ited number of predictors used, as DL is a powerful learning small, four time points for the training set were used, while at method but that requires a significant amount of data [20]. least 50 time points have been recommended for the autore- A disadvantage of the use of DL is its complexity, which can gressive integrated moving average (ARIMA) statistical analy- make it difficult to interpret. LM and RF are more transparent sis model to reliable recognize patterns [31]. As the results of and may be preferred. LM is commonly used in this field of this study showed, increasing the number of time points might study, and its performance is comparable to RF, while its stan- have helped to reduce the error in this study. Therefore, in fu- dardized coefficient is more informative. The standardized co- ture, a more efficient way to capture the sudden change in cur- efficient not only can indicate how important the predictor is rent health metrics needs to be applied. Either more data points but also can indicate the direction of the relationship between for each subject need to be collected so that a health trend can the predictor and the outcome variable. be recognized, or additional features that can better capture the changes in the health metric should be acquired. Adding newly extracted variables improved the performance of the model. Using health trend as a predictor helped to estimate the health metric for each subject, provided that the health met- Conclusions ric of the subject was constant. However, if the current health metric significantly deviated from the previously, predicting the This study investigated the application of machine learning health metric based only on the health trend variable would algorithms to accurately predict a health metric related to be difficult. Using a personal-fitted variable that captured the health status using sociodemographic in an aging population. Indexed in: [Current Contents/Clinical Medicine] [SCI Expanded] [ISI Alerting System] This work is licensed under Creative Common Attribution- NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0) 2000 [ISI Journals Master List] [Index Medicus/MEDLINE] [EMBASE/Excerpta Medica] [Chemical Abstracts/CAS] Engchuan W. et al.: Health status using machine learning and ELSA © Med Sci Monit, 2019; 25: 1994-2001 CLINICAL RESEARCH Data from the English Longitudinal Study of Ageing (ELSA) were Acknowledgments used and the study was part of the Ageing Trajectories of Health: Longitudinal Opportunities and Synergies (ATHLOS) project. This study was conducted within the Ageing Trajectories of Personal-fitted and health trend variables were incorporated in Health: Longitudinal Opportunities and Synergies (ATHLOS) the model, which were shown to be beneficial. Three prediction project. methods, random forest (RF), deep learning (DL) and the linear model (LM), with the calculation of the percentage increase in Statement mean square error (%IncMSE) were applied, and the best re- sults were achieved and tested from RF. DL may be superior in The views expressed in this manuscript are those of the au- other studies, but a significant amount of data and expertise thors and do not necessarily represent the views or policies are required. Different parameter optimization techniques can of the World Health Organization. be applied and more predictors can be added to improve the current DL model. The recommended setting can be applied to Conflict of interest other datasets in the ATHLOS project, in case those datasets have similar characteristics as the ELSA dataset. For the non- The authors declare that they have no conflict of interest. similar datasets, the testing procedure used in this study can be performed to find the most suitable setting for the dataset. References: 1. Population structure and ageing – Statistics Explained. Available 16. Pogorelc B, Bosnić Z, Gams M: Automatic recognition of gait-related health from: http://ec.europa.eu/eurostat/statistics-explained/index.php/ problems in the elderly using machine learning. Multimed Tools Appl, 2012; Population_structure_and_ageing#Further_Eurostat_information 58: 333 2. Mather M: Fact Sheet: Aging in the United States (2015). URL:http://www. 17. Song X, Mitnitski A, Cox J, Rockwood K: Comparison of machine learning prb.org/Publications/Media-Guides/2016/aging-unitedstates-fact-sheet.aspx techniques with classical statistical models in predicting health outcomes. 3. Asian Development Bank, Population and Aging in Asia: The Growing Stud Health Technol Inform, 2004; 107(Pt 1): 736–40 Elderly Population. Article, 18 Jan 2017. URL:https://www.adb.org/features/ 18. Smith JP, Kington R: Demographic and economic correlates of health in old asia-s-growing-elderly-population-adb-s-take age, Demography, 1997; 34(1): 159–70 4. World Health Organization (WHO). Men, ageing and health: Achieving health 19. Kotsiantis SB: Supervised machine learning: A review of classification tech- across the life span, 2008. URL:http://www.who.int/ageing/publications/ niques. Informatica, 2007; 31: 249–68 men/en/ 20. LeCun Y, Bengio Y, Hinton G: Deep learning. Nature, 2015; 521(7553): 436–44 5. United Nations, Department of Economic and Social Affairs, Population 21. Caballero FF, Soulis G, Engchuan W et al: Advanced analytical methodolo- Division (2015). World Population Ageing 2015 (ST/ESA/SER.A/390). gies for measuring healthy ageing and its determinants, using factor anal- URL:http://www.un.org/en/development/desa/population/publications/pdf/ ysis and machine learning techniques: The ATHLOS project. Sci Rep, 2017; ageing/WPA2015_Report.pdf 7: 43955 6. World Health Organization (WHO). Global status report on noncom- 22. Steptoe A, Breeze E, Banks J, Nazroo J: Cohort profile: The English longitu- municable diseases (2010). URL:http://apps.who.int/iris/bitstre dinal study of ageing. Int J Epidemiol, 2013; 42(6): 1640–48 am/10665/44579/1/9789240686458_eng.pdf 23. de la Fuente J, Caballero FF, Sánchez-Niubó A et al: Determinants of health 7. The Institute for Health Metrics and Evaluation (IHME). Washington, D.C. trajectories in England and the US: An approach to identify different pat- USA. Global Health Data. URL:http://www.healthdata.org/ terns of healthy aging, J Gerontol A Biol Sci Med Sci, 2018; 73(11): 1512–18 8. Salminen A, Hyttinen JM, Kaarniranta K: AMP-activated protein kinase in- 24. von Elm E, Altman DG, Egger M et al.; STROBE Initiative: The Strengthening hibits NF-kB signaling and inflammation: Impact on healthspan and lifes- the Reporting of Observational Studies in Epidemiology (STROBE) Statement: pan. J Mol Med (Berl), 2011; 89(7): 667–76 Guidelines for reporting observational studies. Int J Surg, 2017; 12(12): 9. Seeman TE, Crimmins E, Huang MH et al: Cumulative biological risk and so- 1495–99 cioeconomic differences in mortality: MacArthur studies of successful ag- 25. Yu H-F, Lo H-Y, Hsieh J-K et al: Feature engineering and classifier ensem- ing. Soc Sci Med, 2004; 58(10): 1985–97 ble for KDD Cup 2010. URL:http://pslcdatashop.org/KDDCup/workshop/pa- 10. Wu MS, Lan TH, Chen CM et al: Sociodemographic and health-related fac- pers/kdd2010ntu.pdf tors associated with cognitive impairment in the elderly in Taiwan. BMC 26. Liaw A, Wiener M: Classification and regression by randomForest. R News, Public Health, 2011; 11: 22 2002; 2/3: 18–22. URL:https://www.r-project.org/doc/Rnews/Rnews_2002-3. 11. Ravussin E, Redman LM, Rochon J et al: A 2-year randomized controlled trial pdf of human caloric restriction: feasibility and effects on predictors of health 27. The R Foundation. The R Project for Statistical Computing. URL:https:// span and longevity. J Gerontol A Biol Sci Med Sci, 2015; 70(9): 1097–104 www.r-project.org 12. McGeer PL, Schulzer M, McGeer EG: Arthritis and anti-inflammatory agents 28. Keras FC: Deep learning library for the ano and tensorflow 2015. URL:https:// as possible protective factors for Alzheimer’s disease A review of 17 epi- keras.io demiologic studies. Neurology, 1996; 47(2): 425–32 29. Draper NR, Smith H: Applied regression analysis. John Wiley & Sons, 2014 13. Holzinger A: Interactive machine learning for health informatics: When do we need the human-in-the-loop? Brain Inform, 2016; 3(2): 119–31 30. Wilson AC, Roelofs R, Stern M et al: The marginal value of adaptive gradient methods in machine learning. The 31st Conference on Neural Information 14. Lee KS, Lee BS, Semnani S et al: Curcumin extends life span, improves Processing Systems (NIPS 2017), Long Beach, CA, USA. MIT Press. 2017. health span, and modulates the expression of age-associated aging genes URL:https://papers.nips.cc/paper/7003-the-marginal-value-of-adaptive-gra- in Drosophila melanogaster. Rejuvenation Res, 2010; 13(5): 561–70 dient-methods-in-machine-learning.pdf 15. Mathias JS, Agrawal A, Feinglass J et al: Development of a 5-year life ex- 31. Box GEP, Tiao GC: Intervention analysis with applications to economic and pectancy index in older adults using predictive mining of electronic health environmental problems. J Am Stat Assoc, 1975; 70(349): 70–79 record data. J Am Med Inform Assoc, 2013; 20(e1): e118–24 Indexed in: [Current Contents/Clinical Medicine] [SCI Expanded] [ISI Alerting System] This work is licensed under Creative Common Attribution- NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0) 2001 [ISI Journals Master List] [Index Medicus/MEDLINE] [EMBASE/Excerpta Medica] [Chemical Abstracts/CAS]