Development of Prediction Models for Healthy Ageing in Community-Dwelling Middle-Aged and Older Adults: A Longitudinal Study Using Machine Learning

Brief summary

Analysing data on older adults in England, we found that 27.9% aged healthily, with physical activity, body composition, and internet use shown to influence healthy ageing in machine learning models.

Introduction

Life expectancy has increased in both developed and developing countries, and a significant number of older adults are living with health problems that require preventive care or specialised treatment. Besides the challenges it presents in the health sector, population ageing is also an economic issue, as there is an imbalance between the existing workforce and the increased demand for resources to ensure the health and well-being of older adults.¹

Specialists aim to promote healthy ageing in order to foster an active and socially engaged older population, which is a desirable goal.² However, this task is complex, as there is no consensus on the definition or determinants of healthy ageing.

The World Health Organization (WHO) defines healthy ageing as the development and maintenance of the functional ability that enables well-being in older age. This holistic definition is based on criteria such as functional ability, which includes functional status—i.e., the ability to perform activities of daily living and mobility—intrinsic capacity, encompassing vitality or muscle strength, cognition, and psychological capacities; and favourable environmental conditions, including access to health services and technology, as well as emotional support.³ Conversely, there is concern about identifying the probability of surviving and remaining free of diseases that may affect functioning, i.e., healthy survival, along with factors related to healthy ageing based on clinical data (chronic diseases, functionality, and intrinsic capacity domains). Additionally, healthy survival throughout the life course is important since social and behavioural factors modulate health span.⁴

Healthy ageing is linked to factors such as belonging to a younger age group, having more income or wealth,^4,5 engaging in physical activity,^5,6 and maintaining a normal body mass index (BMI).⁶ Identifying predictors of healthy ageing in middle-aged and older adults—based on their social, health, and psychosocial characteristics available in epidemiological datasets—is crucial for developing interdisciplinary strategies aimed at promoting health, preventing disease, and improving social conditions to enhance quality of life in this group. In this context, machine learning (ML), a branch of artificial intelligence, increasingly supports researchers and healthcare professionals in making accurate, data-driven decisions.⁷ While ML has been widely used to detect adverse health outcomes among older individuals,⁸ its potential in identifying predictors of healthy ageing remains relatively underexplored.⁹

The present study aimed to develop machine learning models to predict healthy ageing, which is defined by factors such as the absence of chronic diseases, functional ability, and intrinsic capacity among community-dwelling middle-aged and older adults. The study sought to create an easy-to-understand algorithm and assess the positive or negative influence of each sociodemographic, health, lifestyle, and psychosocial predictor on healthy ageing.

Methods

Study Design, Data Collection and Participants

This retrospective study extracted data from the English Longitudinal Study of Ageing (ELSA), an ongoing panel study involving community-dwelling individuals aged 50 or older in England. The first wave began in 2002, with new waves commencing every two years. Some participants received visits from nurses every four years to collect clinical data.⁹ For this analysis, we included participants of both sexes, as defined by sex assigned at birth, who were aged 50 years or older and had been assessed according to healthy ageing criteria over a four-year period, specifically during waves 2 (baseline) and 4 (follow-up). Only individuals alive at the follow-up assessment were included in this analysis. The participants signed a consent form, and the ELSA study was approved by the National Research Ethics Service (London Multicentre Research Ethics Committee [MREC/01/2/91]).

Healthy Ageing

At baseline and follow-up, the following clinical criteria were used to identify healthy ageing:^3,10,11,12 (a) Functional status (yes or no) in Instrumental Activities of Daily Living (IADL), including preparing meals, shopping, managing money and medicines, answering the telephone, cleaning the house, and maintaining the home. Participants who experienced no difficulty performing IADL were classified as having preserved functional status.¹³ (b) Preserved mobility (yes or no), assessed by the ability to perform the following activities: walking 100 yards, ascending a flight of stairs without stopping, lifting, or carrying weights over 10 pounds, and pulling or pushing large objects. Participants who faced no difficulty performing these activities were classified as having preserved mobility.¹⁴ (c) Preserved muscle strength (yes or no), determined by hand grip strength of the dominant arm. Participants were asked to perform three hand grip movements using an isokinetic dynamometer. The best value was divided into quintiles and stratified by age and sex. The highest hand grip strength quintile was used to classify individuals as having preserved muscle strength.¹⁵ (d) Absence of elevated depressive symptoms (yes or no), defined as a score of less than 4 on the Centre for Epidemiologic Studies Depression Scale (CES-D).^{16, 17} (e) Preserved cognitive function (yes or no), assessed based on a cognitive screening test. Each participant was asked to recall as many words as possible from a list of ten words the interviewer provided immediately and after five minutes. The test had a minimum score of 0 and a maximum score of 20. Scores were divided into quintiles and stratified by age, sex, and educational level. The quintile with the highest test score was used to classify individuals as having preserved cognitive function.^{18, 19} (f) Absence of self-reported doctor-diagnosed disabling diseases (yes or no), such as acute myocardial infarction, angina pectoris, heart failure, stroke, Parkinson's disease, chronic obstructive pulmonary disease, diabetes, and cancer.^{11, 12} Participants without any diseases were classified positively on this criterion.

At baseline and follow-up, participants who scored positively on all the criteria were classified as experiencing healthy ageing. The outcome variable “healthy ageing” (yes or no) was defined. The "yes" category included participants who had aged healthily at both waves and those who were not experiencing healthy ageing at baseline but improved at follow-up. The “no” category consisted of participants without healthy ageing at both waves and those experiencing healthy ageing at baseline but not at follow-up.

Candidate predictor variables

Baseline variables deemed necessary for predicting healthy ageing were initially selected.^{3, 20} Sociodemographic predictors included age (50–59, 60–69, 70–79, and 80 years or older), sex (male and female), ethnicity (non-white and white), marital status (having or not having a partner), living arrangement (not living alone and living alone),²¹ level of education (>13 years, 12–13 years, and 0–11 years), social class based on occupation (managerial and professional occupations, intermediary occupations, semi-routine and routine occupations, other occupations/never worked), household wealth in quintiles, employment situation (in paid work and not in paid work), and housing situation (owner, buying, renting, free rent).⁹

Psychosocial predictors relate to emotional and supportive factors, as well as social and technological engagement. The variables included: desired age and self-perceived age (50–59, 60–69, 70–79, and 80 years or older); the experience of ageing (positive or negative);⁹ positive social support (high or low);²² contact frequency on social networks (high, average, or low);²³ self-reported loneliness (almost never/never, sometimes, or very often);²⁴ use of the internet or emails (yes or no); membership in a political party (yes or no); membership in a church or religious group (yes or no); and reported frequency of volunteering (twice a month or more, once a month or every few months, once or twice a year, less than once a year, never).⁹

Lifestyle and health predictors include physical performance based on each participant’s time in seconds in the chair-rise test, categorised as normal (<10.8 seconds), intermediate (10.8 to 12.7 seconds), or poor (≥12.8 seconds);²⁵ balance problems (never, sometimes, or very often/always); self-reported hearing and sight (excellent/very good, good, fair, or poor); alcohol consumption (rarely/never, frequently, or daily); smoking status (nonsmoker, former smoker or smoker); obesity based on BMI, categorised as obese (≥30.0 kg/m²), overweight (25.0 kg/m² to 29.9 kg/m²), and underweight/normal (≤24.9 kg/m²);²⁶ abdominal obesity (yes or no);²⁷ self-reported high blood pressure (yes or no); pain status (no pain, mild, moderate, or severe);²⁸ and physical activity (vigorous activity at least once a week), moderately active (moderate activity at least once a week), or inactive (no moderate or vigorous activity every week);²⁹ self-reported sleep quality (very good/good or bad/very bad). Total cholesterol levels (normal or elevated); high-density lipoprotein (HDL) levels (normal or low); triglyceride levels (normal or elevated);³⁰ and haemoglobin levels (normal or low - anaemia).³¹ A full description of candidate predictors can be found in Supplementary Material 1A.

Development of the machine learning models

The Decision Tree (DT), Logistic Regression (LR), Neural Network (NNET), and Random Forest (RF) algorithms utilised in the study are described in Supplementary Material 1B.

In the first stage, the dataset was prepared. The variable ethnicity was excluded because its variance was close to zero. The variables haemoglobin, C-reactive protein, HDL, and triglyceride levels were excluded because a significant percentage of data was missing, approximately 30% (Supplementary Figure 1). Although biomarker collection in ELSA is carried out among participants in good health status, which characterises a missing-at-random mechanism—i.e., missingness dependent on other variables in the dataset, thus allowing for imputation—some individuals may decline the nurse visit due to discomfort with certain procedures, privacy concerns, or simply the decision not to take part in that component of the study. None of the predictors exhibited high collinearity (Supplementary Figure 2). The original dataset was divided into a training dataset (70% of the sample) and a test dataset (30% of the sample). This division was balanced within each category of the outcome variable to maintain the data distribution.

Preprocessing was carried out on the training dataset to prevent information leakage between datasets. During the initial preprocessing phase, we used the missForest imputation method, which relies on the RF algorithm. The observed values in the dataset are used to train RF to predict the missing values. To assess imputation error, we employed the Proportion of Falsely Classified (PFC) metric. A PFC value close to 0 indicates good imputation accuracy, while a value near 1 reflects poor imputation accuracy.³² The RF algorithm's performance resulted in a PFC of 0.082473, indicating that only 8% of the imputed values were misclassified, demonstrating satisfactory imputation performance.

Subsequently, the numerical variables were normalised to have a mean of 0 and a standard deviation of 1. Coding and transformation techniques were not applied to the models trained with DT and RF, as they are unnecessary for these algorithms. We identified the most important variables for predicting healthy ageing using the Boruta algorithm.³³ Among these variables, 21 were considered significant and incorporated into the ML models (Supplementary Figure 3).

The training dataset was divided into ten folds for the cross-validation test. Nine folds were used to train the models, while one was set aside for testing. This procedure was repeated until all the folds had been utilised for testing. In the training dataset, parameter tuning was carried out using 10 tuning grids to select the best hyperparameter for each model, specifically the hyperparameter that enhanced the model’s performance in the cross-validation test. The performance value from each fold was recorded, and the ML models with the highest mean performance across all the folds were selected. The receiver operating characteristic curves (ROC) and the mean values of the metrics and parameters for the chosen models during the cross-validation test in the training phase can be found in Supplementary Figure 4 and Supplementary Table 1.

Testing of the models

The models selected in the previous stage were tested for internal validity using the test dataset. The performance of the algorithms on the test dataset was evaluated using the following metrics: sensitivity, specificity, F1-score, positive predictive values (PPV), negative predictive values (NPV), ROC, area under the ROC (ROCAUC), and the area under the precision-recall curves (PRAUC). It is important to recognise that PPV and NPV are key complementary metrics in health research, as they allow for the quantification of how many observations were correctly classified as having or not having healthy ageing, respectively. Additionally, the F1-score acts as a vital complementary metric that provides an accurate assessment of ML model performance in imbalanced datasets, such as that of the current study. This is because it considers correctly classified cases of healthy ageing, even when this category contains fewer observations. Unlike accuracy, the F1-score is not biased towards the majority class, which makes it especially useful in such situations.

The models' residuals were estimated based on their performance on the test dataset. The lower the median residual value, the greater the accuracy of the model. Finally, the calibration curve was used to determine whether the models were calibrated.

Explanation of how healthy ageing was predicted

We first applied the Variable Importance technique using permutation tests. This method provides a global explanation of the model, showing which predictors affect its performance. If modifying a variable’s values affects the model's overall performance after the permutation test has been run, that variable can be considered significant for predicting healthy ageing.

The SHapley Additive exPlanations (SHAP) algorithm was utilised because it explains how each predictor contributes to the prediction of the outcome (local explanation). The algorithm considers the effect of each predictor on the outcome and adjusts that predictor’s effect considering the other predictors. In other words, the influence is cumulative rather than isolated. Predictors with positive values contribute positively to the outcome prediction, while those with negative values contribute negatively. We calculated SHAP values using observations selected from the test dataset for an accurate and local interpretation of the model's prediction.³⁴

Finally, we employed Friedman’s H statistic ³⁵ to verify interaction effects among all predictors and between chronological and perceived age with sociodemographic and clinical predictors. The H-statistic produces values ranging from 0 to 1, quantifying the strength of interactions among predictors, with higher values indicating stronger interaction effects. The H-statistic calculates the percentage of explained variance in the model that comes from the interaction between predictors. Indeed, the effect of sociodemographic and clinical predictors on healthy ageing may depend on chronological or self-perceived age. For example, younger participants or those who feel younger may score higher on the chair-rise test, be less likely to encounter health problems, and consequently, be healthier.

Statistical analysis

Baseline categorical and numerical variables were compared using the chi-squared Pearson and Mann-Whitney tests based on healthy ageing status at follow-up. A significance level of p< 0^.05 was adopted. Odds ratios for healthy ageing were calculated through univariate analysis. At this stage, the unprocessed dataset was utilised. RStudio version 2023^.12^.0+369 was used for all data analyses.

Results

A total of 9171 individuals were assessed based on healthy ageing criteria at baseline; of these, 6519 (71.1%) were reassessed after four years of follow-up. After applying exclusion criteria, a final sample of 6332 individuals was included in the analysis. A flowchart outlining the stages of developing the ML models with sample selection is provided in Supplementary Figure 5.

In the dataset without preprocessing, 2212 individuals (34.9%) were between 50 and 59 years old, and 3580 (56.5%) were women at baseline. At follow-up, 1768 (27.9%) were experiencing healthy ageing. At baseline, most participants who were experiencing healthy ageing at follow-up were between 50 and 59 years old (51.5% vs. 28.5% of those who were not experiencing healthy ageing), reported feeling between 50 and 59 years old (70.2% vs. 40.3%), and had greater household wealth (35.4% vs. 17.5%). Similarly, at baseline, the majority of those who were ageing healthily at follow-up had normal physical performance (60.1% vs. 32.0% of those who were not ageing healthily), and were physically active (25.6% vs. 12.6%) (Table 1). The distribution of the variables in the training and test datasets is shown in the Supplementary Table 2.

Table 1.

Baseline sociodemographic and health characteristics by healthy-ageing status at follow-up

	Total	Healthy Ageing=No	Healthy Ageing=Yes		Healthy Ageing=Yes
	(n=6332)	(n=4564)	(n=1768)	p value†	Odds ratio (CI)††
Age, median (IQR), y	64.0 (57.0–72.0)	66.0 (58.0 –74.0)	59.0 (55.0 –65.0)	<0.0001	0·95 (0.94–0.96)
Age group				<0.0001
≥80	537 (8.5)	523 (11.5)	10 (0·7%)		Ref
70–79	1448 (22.9)	1261 (27.6)	163 (10·9%)		3.01 (2.22–4.14)
60–69	2135 (33.7)	1479 (32.4)	562 (37·5%)		5.37 (4.02–7.32)
50–59	2212 (34.9)	1301 (28.5)	763 (50·9%)		6.63 (4.97–9.03)
Self-perceived age				<0.0001
≥80	177 (2.8)	174 (3.8)	3 (0.1)		Ref
70–79	664 (10.5)	598 (13.1)	66 (3.7)		3.53 (1.95–7.09)
60–69	1418 (22.4)	1098 (24.0)	320 (18.1)		6.24 (3.52–12.3)
50–59	3173 (50.1)	1931 (42.3)	1242 (70.2)		11.90 (6.75–23·30)
Self-perceived age - missing	900 (14.2)	763 (16.7)	137 (7.7)	<0.0001	0·42 (0·37–0·51)
Sex *				0.0109
Male	2752 (43.5)	1960 (42.9)	792 (44.8)		Ref
Female	3580 (56.5)	2604 (57.1)	976 (55.2)		0.80 (0.72–0.89)
Household wealth				<0.0001
First quintile (lowest quintile)	920 (14.5)	824 (18.0)	96 (5.4)		Ref
Second quintile	1118 (17.6)	908 (19.8)	210 (11.8)		2.02 (1.62–2.54)
Third quintile	1202 (19.0)	896 (19.6)	306 (17.3)		2.95 (2·15–3·43)
Fourth quintile	1313 (20.7)	871 (19.0)	442 (25.0)		3.80 (2·93–4·64)
Fifth quintile (highest quintile)	1429 (22.6)	803 (17.5)	626 (35.4)		5.12 (3·82–6·00)
Household wealth - missing	350 (5.5)	262 (5.7)	88 (4.9)	0.16	0.90 (0.69–1.15)
Level of education				<0.0001
0–11 years	3006 (47.5)	2466 (54.0)	540 (30.5)		Ref
12–13 years	1602 (25.3)	1044 (22.8)	558 (31.5)		1.92 (1.69–2.19)
>13 years	1704 (26.9)	1038 (22.7)	666 (37.6)		2.19 (1.93–2.48)
Level of education - missing	20 (0.3)	16 (0.3)	4 (0.2)	0.72
Physical performance				<0.0001
Poor	2869 (45.3)	2468 (54.0)	401 (22.6)		Ref
Intermediate	986 (15.5)	690 (15.1)	296 (16.7)		4·24 (3·54–5·08)
Normal	2444 (38.6)	1381 (30.2)	1063 (60.1)		6·59 (5.72–7·62)
Physical performance - missing	33 (0.5)	25 (0.5)	8 (0.4)	0.50	0.86 (0.36–1.85)
Balance problems				<0.0001
Very often/always	358 (5.6)	347 (7.6)	11 (0.6)		Ref
Sometimes	899 (14.2)	800 (17.5)	99 (5.6)		3.52 (2.09–6.36)
Never	4973 (78.5)	3324 (72.8)	1649 (93.2)		14.8 (9.21–26.3)
Balance problems - missing	102 (1.6)	93 (2.0)	9 (0.5)	0.002
Level of physical activity
Inactive	3731 (58.9)	2980 (65.4)	751 (42.4)		Ref
Moderately active	1523 (24.0)	965 (21.1)	558 (31.5)		2.15 (1.89–2.43)
Active	1032 (16.3)	578 (12.6)	454 (25.6)		2.83 (2.46–3.27)
Level of physical activity- missing	46 (0.7)	41 (0.9)	5 (0.2)		0.29 (0.17–1.20)
Abdominal obesity				0.007
Yes	3584 (56.6)	2703 (59.3)	881 (49.8)		Ref
No	2748 (43.4)	1861 (40.8)	887 (50.2)		1.69 (1.54–1.92)
Use of the internet or e-mails				<0.0001
No	3301 (52.1)	2631 (57.7)	670 (37.9)		Ref
Yes	2416 (38.1)	1414 (31.0)	1002 (56.6)		2.24 (2.01–2.50)
Use of the internet or e-mails- missing	615 (9.7)	519 (11.4)	96 (5.4)		0.83 (0.61–1.11)

The data are represented as n (%). The results of the descriptive statistics in the table were estimated with the original dataset without ML preprocessing.

^*As no information on participants’ sex was available during data collection, only birth-assigned sex was recorded. There is currently no information on participants’ gender.

^†Statistically significant difference between the groups (p < 0.05).

^††Odds ratios were calculated based on univariate analysis using the original dataset without ML preprocessing.

The model trained with RF demonstrated superior performance in predicting healthy ageing, achieving a ROCAUC of 0.78 (95% CI 0.76–0.80) and a PRAUC of 0.63 (95% CI 0.59–0.67). The sensitivity values for all models ranged from 0.69 (95% CI 0.65–0.78) to 0.75 (95% CI 0.71–0.80), while the PPV values varied between 0.53 (95% CI 0.46–0.59) and 0.62 (95% CI 0.58–0.64) (Table 2).

Table 2.

Performance of the Machine Learning algorithms on the test dataset.

	ROCAUC (95% CI)	PRAUC (95% CI)	Sensitivity (95% CI)	Specificity (95% CI)	PPV (95% CI)	NPV (95% CI)	F1-score*	Threshold†
DT	0.70 (0.68–0.74)	0.55 (0.53–0.63)	0.69 (0.65–0.78)	0.70 (0.65–0.76)	0.53 (0.46–0.59)	0.77 (0.72–0.85)	0.60 (0.58–0.65)	0.391
NNET	0.77 (0.74–0.82)	0.58 (0.54–0.60)	0.73 (0.65–0.76)	0.71 (0.68–0.77)	0.54 (0.47–0.56)	0·84 (0.77–0.89)	0.62 (0.60–0.69)	0·401
RF	0.78 (0.76–0.80)	0.63 (0.59–0.67)	0.75 (0.71–0.80)	0.72 (0.65–0.76)	0.62 (0.58–0.64)	0.78 (0.75–0.83)	0.67 (0.63–0.70)	0·360
LR	0.77 (0.72–0.84)	0.59 (0.56–0.63)	0.73 (0.66–0.84)	0.72 (0.63–0.75)	0.56 (0.52–0.60)	0.83 (0.80–0.89)	0.63 (0.59–0.67)	0·371

The data are shown as the mean and confidence interval. DT=decision tree; LR=logistic regression; NNET= neural network; RF=random forest; ROCAUC=area under the receiver-operating characteristic curve; PRCAUC=precision recall area under the curve; PPV=positive predictive value; NPV=negative predictive value.

^*The balance between PPV (precision) and sensitivity (recall) is given by F1-score = 2 x precision x recall/ (precision + recall).

^†The Youden index was used to determine the ideal threshold for predicting healthy ageing by optimizing specificity, sensitivity, PPV and NPV.

The model trained with DT had the worst performance, as indicated by the lowest ROC and the highest absolute median residual value. In addition, this model displayed points above or below the diagonal line, suggesting poorer calibration (Figure 1 A–B). Conversely, the models trained with NNET, RF, and LR algorithms performed better in terms of ROC and exhibited improved calibration; the calibration curves were close to the diagonal on the graph. This indicates that the predicted probabilities aligned with the observed frequencies in the sample (Figure 1 A–B). The model trained with the RF algorithm had the lowest absolute median residual value, suggesting superior performance (Figure 1C).

Figure 1. ROC curves, calibration plots and residuals for machine learning models.

NB: A. Receiver operating characteristic curves; B. Calibration plots; C. Residuals plot. The receiver operating characteristic curves and residuals plot demonstrate the performance of the machine learning models with the test dataset. The calibration plots illustrate the observed versus predicted outcomes using the test dataset. A perfectly calibrated model would be a straight line where the predicted probabilities always equal the observed ones; the smooth lines are plotted using the locally estimated scatterplot smoothing (LOESS) method.

The most significant predictors - those that, when altered after the permutation test, can lower the AUC metric of the models- were household wealth, chronological age, internet use, physical performance, life satisfaction, physical activity, balance, pain status, and abdominal obesity (Figure 2A). Among the SHAP values, higher household wealth and perceived and chronological ages between 50 and 59 years, normal physical performance, absence of pain, and extreme life satisfaction contributed positively to the prediction of healthy ageing. Conversely, physical inactivity, lack of internet or email use, and abdominal obesity contributed negatively to healthy ageing (Figure 2 B).

Figure 2. Most important predictors and SHAP values for healthy ageing prediction.

NB: A. Variable importance plot; B. SHAP values plot; C. Interaction effect plot. All graphs are based on the model trained using the Random Forest algorithm, which achieved the best performance. In all graphs, the most important predictors are displayed at the top. In the SHAP values plot, the green bars indicate a positive contribution to healthy ageing, while the red bars signify a negative contribution. The contribution of each predictor is calculated by averaging all possible orderings of predictors; the SHAP algorithm evaluates the negative or positive contribution of a given predictor by assessing the change in the outcome prediction when this predictor is included in the model. The contribution value of each predictor is represented on the x-axis of each graph. Boxplots illustrate the distribution of a variable's assignment across all possible combinations of variable layouts. Chair = Physical performance based on the chair rise test; Age = Chronological age; Internet = Use of internet or e-mail; Wealth = Household wealth; Pain = Pain status; Balance = Balance problems; Life = Life satisfaction; Self-age = Self-perceived age; PA = Adequate level of physical activity; Waist = Abdominal obesity.

The interaction strength between each predictor and all other predictors was weak (H statistic > 0.025), explaining less than 2.5% of the variance in healthy ageing prediction (Figure 3 A). Similarly, the interaction strength of each sociodemographic and clinical predictor with age and self-perceived age was weak (H statistic ≤ 0.030), explaining 3.0% of the variance in healthy ageing prediction (Figure 3 B).

Figure 3. Interaction strength for healthy ageing prediction.

NB: A. The interaction strength of each predictor with all other predictors; B. The interaction strength of each predictor with chronological and self-perceived age. All graphs are based on the model trained using the Random Forest algorithm, which achieved the best performance. In each graph, the most significant interactions are presented at the top. The x-axis displays the interaction strength standardised using the H statistic.

Discussion

This study has developed ML models using various algorithms to predict healthy ageing over time in community-dwelling middle-aged and older English adults. The models were based on sociodemographic, health, lifestyle, and psychosocial predictors, with RF demonstrating better ROCAUC and PRAUC values on the test dataset. The contribution of each predictor to the study outcome prediction, along with interactions between predictors, was also presented.

The definition of healthy ageing used here does not cover all the WHO criteria. However, we evaluated functional ability by employing the criteria of functional status and preserved mobility, as well as intrinsic capacity using the criteria of vitality or muscle strength, preserved cognitive function, and the absence of elevated depressive symptoms. The absence of chronic diseases was considered in the present study because this criterion is part of the clinical model used to define healthy ageing.^4,10,12 Including this criterion makes the models presented here useful in clinical practice³⁶ and allows for comparisons with other studies since these criteria are objective and widely available in most epidemiological data.^{1,3,4,10,11,12}

In this study, 27.9 % of participants were experiencing healthy ageing after four years. In a cohort study with individuals in the same age group, 20% of the sample were experiencing healthy ageing, defined as the absence of chronic diseases, cognitive impairment, and disability, good mental health, self-perceived good health, and the absence of mobility problems and pain.³⁷ Another study with middle-aged and older adults found that after 10 years, 15.7% of participants were experiencing healthy ageing, defined as the absence of elevated depressive symptoms, functional impairment, cognitive deficit, and disabling diseases. These studies employed different definitions and had longer follow-ups. Consequently, older participants did not achieve or maintain healthy ageing. This helps explain, to some extent, the higher incidence of healthy ageing in the present study.¹⁰

The model trained with RF was notable for its higher ROCAUC and PRAUC values. RF can combine many decision trees using robust statistics and has been employed in predictive models applied to complex datasets in the health field.³⁸ Although it is easy to justify the decision to use higher-performance algorithms such as RF, the DT algorithm produces explanatory models that have gained widespread acceptance in the health field because they are easier to interpret.^39,2 justifying our choice to compare different ML algorithms.

All the models achieved PPV values greater than 50; that is, for every two observations classified as healthy ageing by the algorithms, at least one corresponded to healthy ageing. The model trained with RF reached the highest PPV value (62 [CI 95% 58–64]). This finding suggests that the RF algorithm is effective, as the model used here was trained on an imbalanced dataset, which means fewer observations fell into the “yes” category for the outcome variable. With imbalanced datasets, ML algorithms find it challenging to predict the category with fewer observations.¹⁶

Some predictors significantly contributed to predicting healthy ageing, as shown by the SHAP algorithm. The highest levels of household wealth positively influenced the outcome. A higher socioeconomic level is linked to greater access to healthcare services.⁴⁰ While a longitudinal study with middle-aged and older adults found that higher individual and household income was associated with healthy ageing,⁴¹ another study indicated that lower socioeconomic levels correlated with longer waiting times for appointments with specialists and surgery in countries with centralised health systems.⁴²

As expected, being in a younger age group contributed positively to the prediction of healthy ageing; however, lower self-perceived age, while contributing less to the prediction, also had a positive effect. Feeling younger was associated with good physical health, cognitive function, psychological well-being, life satisfaction, and reduced use of healthcare services among older adults.⁴³ Nevertheless, a prior study discovered that the connection between self-perceived age and physical health and longevity in middle-aged and older adults was minimal, thereby supporting our findings.⁴⁴ Moreover, although self-perceived age is a recognised indicator of the ageing process, its effect on healthy ageing may be influenced by age-related bias. Younger older adults are more inclined to report feeling younger, irrespective of their actual health status, which may exaggerate the effect of self-perceived age on healthy ageing within this subgroup. This reinforces the importance of assessing the interaction between self-perceived and chronological age, as undertaken in the present study. Our analyses revealed that this interaction did not significantly enhance the prediction of healthy ageing, suggesting that self-perceived age contributed meaningfully to healthy ageing outcomes independently of chronological age.

Normal physical performance contributed positively to healthy ageing. A decline in physical performance, specifically loss of muscle strength, is linked to sarcopenia, disability, and falls. Conversely, maintaining physical performance is crucial for maintaining physical health, mobility, and independence in older adults.⁴⁵ Furthermore, the chair-rise test effectively identifies vulnerabilities such as sarcopenia and frailty, and it is cost-effective in clinical practice.^46,17

Lifestyle factors significantly influence ageing phenotypes, with physical activity being one of the most crucial components for healthy ageing. Moderate-intensity physical activity, such as walking, reduces the risk of morbidity and mortality in older adults.⁴⁷ Conversely, a lower level of physical activity is a determining factor for frailty among older adults.⁴⁸ Moreover, a longitudinal study found that inactive older adults experienced poorer healthy-ageing trajectories after five years compared to those engaging in any level of physical activity.⁴⁹

It should be emphasised that physical activity and physical performance measured through clinical tests are distinct domains that evaluate different aspects of health and physical functioning. Including these predictors in models for predicting healthy ageing is crucial for disease prevention and treatment focused on rehabilitation.^{47, 48}

Interestingly, not using the internet or email contributed negatively to the prediction of healthy ageing. For older adults, the internet serves as an important means of maintaining and expanding their contacts within social networks. This technology also provides access to healthcare information linked to improved physical health. Specifically, for frail individuals, internet use, such as online shopping or banking, can help compensate for functional limitations and restricted mobility. Additionally, it allows them to engage with their social environment based on their needs.⁵⁰ Furthermore, evidence indicates that initiating internet use improved cognitive function in older adults who were not previously users. Conversely, ceasing internet use was associated with accelerated cognitive decline over the 16-year study period.⁵¹ Our findings suggest that not using the internet at baseline negatively affected cognition and functionality, both of which are components of intrinsic capacity used in this study's definition of healthy ageing.

The results obtained from the model tested here indicated that abdominal obesity contributed negatively to the prediction of healthy ageing. Abdominal obesity is considered a more accurate health risk factor than obesity measured by BMI, particularly in older adults, and it is associated with disability, falls, and mortality.⁵²

The present study has several limitations. First, because many of the variables used are self-reported, measurement errors may introduce bias in the analysis. Although this is a significant limitation, self-reported variables are common in population studies. Chronic and degenerative diseases cited in the literature to define healthy ageing, such as renal failure, multiple sclerosis, and amyotrophic lateral sclerosis,^11,12 were not captured in the assessments. Therefore, this study's definition of healthy ageing was adapted according to the healthy ageing criteria available in the dataset. Another limitation is that the follow-up lasted only four years, which may not have been sufficient to identify changes that could have had a greater effect on the incidence of healthy ageing, given the participants' age and the modifiable factors analysed. Nevertheless, older adults may experience short-term fluctuations in health status that contribute to increased vulnerability and frailty, as shown by a recent study.⁵³ Therefore, investigating healthy ageing over shorter time frames remains important, as it can reveal health behaviours, clinical factors, and psychosocial determinants that may help optimise population health in older age. In addition, we chose this short follow-up period because it allowed us to gather more observations, which is important as ML models require a substantial number of observations in both categories of the outcome variable. Attrition during follow-up can also introduce bias. However, this limitation is inherent in longitudinal studies like ELSA, as ongoing participation depends on various factors beyond the sampling strategies and protocols used, including the incentives offered to motivate respondents to stay in the study. While this remains a notable limitation, ML models showed strong performance in distinguishing individuals who experienced healthy ageing from those who did not during the follow-up period. Finally, the relatively imbalanced nature of the dataset may have affected some ML metrics. However, since there was no major imbalance, we decided not to minimise the differences in the observations between the two outcome categories. We maintained a true characteristic of population studies: there are certainly not similar numbers of observations for individuals experiencing healthy ageing and those not experiencing this process.

A positive point we should highlight is that we utilised robust algorithms, such as RF and NNET, along with an explanatory one, DT. Comparing the results of each of these algorithms is essential to verify their consistency. Furthermore, the study employed the SHAP algorithm as an additional method to better understand the prediction results. Additionally, we explored interactions among predictors, and the findings indicated that neither chronological age nor self-perceived age functioned as significant moderators in the relationship between sociodemographic and age-related factors and healthy ageing. Lastly, we used a structured dataset with predictors and an outcome variable that can be selected in other datasets for external validation.

Conclusions and Implications

This study has developed ML models for predicting healthy ageing. These models can assist researchers, healthcare professionals, and managers in formulating effective public policies and guiding healthcare professionals in developing and implementing strategies to prevent and treat disabling diseases. The explanatory model created here can also enhance communication between researchers and healthcare professionals, helping to promote health by involving the population in decisions to change lifestyle habits. Future longitudinal studies are recommended to build on the variables and modelling methods outlined in this research to perform external validation. In particular, using ML algorithms that include explainability features could offer further understanding of the changing factors influencing healthy ageing over time.

Data statement

The data are available upon reasonable request. Data from the ELSA Study can be accessed through the UK Data Service for researchers who meet the criteria for accessing confidential data under the conditions of the End User Licence: https://discover.ukdataservice.ac.uk. The data can be retrieved at https://www.elsa-project.ac.uk/accessing-elsa-data and requested via the site https://ukdataservice.ac.uk/help/. Although the ELSA data cannot be directly accessed in this study, for the sake of reproducibility, we provide an example of the code used for the estimation of the explanatory algorithm via SHAP values (Supplementary material 1C).