To maximize the reporting quality of this scoping review, we followed the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) recommendations [30] (checklist provided in Checklist 1). The protocol for this scoping review was registered with the International Platform of Registered Systematic Review and Meta-Analysis Protocols (INPLASY) [31]. All data generated in this review are provided in Multimedia Appendix 1 and are accessible in the institutional repository [32].

In this scoping review, we searched for primary studies in the 3 principal health databases: PubMed (National Center for Biotechnology Information), PsycINFO (ProQuest), and Web of Science (Clarivate). The search was restricted to medical and psychological databases to capture studies directly relevant to health outcomes. Consequently, studies primarily published in engineering or computer science, which may focus on algorithm development or sensor-based data processing, were not included. The search strategy followed the PRESS (Peer Review of Electronic Search Strategies) [33] and PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses–Search extension) guidelines [34] and consisted of 2 groups of search terms referring to (1) HL and (2) ML. We also added a third group of terms preceded by the Boolean operator “NOT” to improve the specificity of the search strategy.

This scoping review adopts a health-focused perspective, in which HL is treated as a multidimensional construct rather than the sum of isolated behaviors [6,35,36]. Therefore, the umbrella term HL was combined using the operator “OR” with an interaction block including (1) physical activity, (2) diet, (3) sleep, and (4) stress. This block is aligned with the multiple health behavior change and lifestyle medicine frameworks, in which the interaction between behaviors is a central construct [37,38].

The search strategy was adapted to the specific syntax of each database (Table S1 in Multimedia Appendix 1). The search was conducted on October 10, 2025, with language restrictions (English and Spanish) but without limitations on publication years.

Mendeley was used as reference management software; the results of the search strategy were entered, and duplicates were merged or removed. An ad hoc checklist was used to extract information from the included papers. The checklist was divided into 5 sections:

The review was presented as a narrative synthesis, and the information was summarized in tables and figures. The information extracted from the studies was divided into 3 blocks: type of data, ML process, and software. For data extraction, we focused on lifestyle components, health outcomes, data sources, acquisition methods, and data typology. Regarding the ML process, we focused on the whole process, consisting of preprocessing, modeling, validation, evaluation, and XAI methods. For model evaluation, only the procedures used for final performance assessment were extracted. When studies reported cross-validation, we classified it as the final performance estimation method unless authors explicitly stated its use for hyperparameter optimization. To identify the top-ranked lifestyle components in the XAI analysis, we systematically examined the figures and tables reported in each study. In this review, the term ranked refers to features that were highlighted by the XAI algorithm. In contrast, unranked indicates that appeared in the model but were not reported in the XAI visualization, while not available denotes features that were not included in the ML model. Software used in each study was also recorded.

A total of 2249 papers were retrieved from the databases, and 65 studies met the eligibility criteria and were included in this scoping review (refer to Figure 1). There was very good agreement between reviewers during the selection process: 96% (n=52; κ=0.84, 95% CI 0.61-1.0) in the title and abstract screening and 94% (n=35; κ=0.88, 95% CI 0.71-1.0) in the full-text screening.

From this point forward, the “Results” section is structured following the ML workflow depicted in Figure 2, which illustrates the 5 key steps in the ML pipeline. The process begins with data acquisition, followed by preprocessing to prepare the data. Then, SL algorithms are applied and evaluated to determine their effectiveness. Finally, explainability techniques are used to understand the models. The dashed lines indicate that modeling, evaluation, and explainability can improve earlier stages, making the process iterative. Each stage of the process corresponds to a subsection. Finally, we examined the software used throughout the entire process in the included studies.

The 65 papers included in this review were published between 2004 and 2025, with 57 (87.7%) published since 2019. Figure 3 shows the annual productivity output stratified by lifestyle components. The studies were carried out in several geographical regions across 4 continents (Table S2 and Figure S1 in Multimedia Appendix 1). The mean sample size was 29,905.14 participants, with the smallest study including 8 participants and the largest including 470,778 participants.

Four lifestyle domains were identified in the review: physical activity, diet, sleep, and stress. Most studies (48/65, 74%) integrated data from more than 1 lifestyle domain (refer to Table 1). The most studied component was physical activity, featured in 61 of 65 studies. Diet and sleep appeared in 33 of 65 (51%) and 34 of 65 (52%) studies, respectively, while stress appeared in 15 of 65 (23%) studies (refer to Figure 3). Only 17 of 65 (26%) studies focused exclusively on a single domain (15 on physical activity [40-54] and 2 on sleep [55,56]).

Regarding data sources, 33 of 65 (51%) studies relied on self-acquired data, while 32 of 65 (49%) studies used either private or public health datasets, such as UK Biobank [57]. Among the studies using self-acquired data, the mean sample size was 10,503.41 participants. Six studies focused exclusively on females [46,47,58-61], and 27 studies reported a female proportion ranging from 41% to 89.4%. Studies using health repositories exhibited a greater mean sample size of 80,406.38 participants. Nineteen studies reported a female proportion ranging from 41.33% to 70%, 12 did not report sex distribution, and 1 study focused exclusively on females.

Figure 4 summarizes the different data collection modes used for each of the 4 lifestyle domains identified (Table S3 in Multimedia Appendix 1 provides a detailed description of each measure). Single-item measures were used in 40 of 65 (61.5%) studies assessing physical activity, 18 of 65 (27.7%) studies assessing diet, and 22 of 65 (33.8%) studies assessing sleep. Despite the heterogeneity of these items, distinct categories emerged within each lifestyle domain.

For physical activity, the emerging categories included engagement in physical activities [47,50-52,54,58,62-66], intensity [10,42,49,57,67-71], frequency [43,44,60,72-86], and environmental factors [87,88]. Within the diet domain, categories included frequency of consumption [58,60,66,72,84], types of products [75,76,80,82], environmental factors [87], and consumption habits [59,62,65,68,70,71,86,89]. Regarding sleep, categories included sleep duration [57,60,62,66-68,70,72,77-79,81,82,84,85,87], perceived sleep quality [63,89], and sleep-related problems [69,80]. Finally, within the stress domain, the emerging category focused primarily on stress level [10,57,79,82,86,89,90].

Questionnaires were used to collect physical activity data in 13 studies. The standardized questionnaires included the Global Physical Activity Questionnaire (GPAQ) [53,72,91], the International Physical Activity Questionnaire (IPAQ) [92], the Lifetime Total Physical Activity Questionnaire (exercise and sport subscale) [93], the Nutritional and Social Healthy Habits (NutSo-HH) scale [94], the Indian Migration Study Physical Activity Questionnaire (IMS-PAQ) [95], physical fitness test [41], the Short Questionnaire to Assess Health-Enhancing Physical Activity (SQUASH) [73,74], the physical activity scale from the Active Living Index [88], the Pregnancy Physical Activity Questionnaire [61], and the Physical Activity Scale for the Elderly [96]. For diet assessment, 12 studies used the Food Frequency Questionnaire (FFQ) [57,73,74,83,88,91,92,95], the PrimeScreen questionnaire [61], the Mini Nutritional Assessment [97], the NutSo-HH scale [94], and a nonstandardized questionnaire consisting of items from different questionnaires [93]. The standardized questionnaires used to measure sleep were the Pittsburgh Sleep Quality Index [10,56,61,86,91,92,96,97], the NutSo-HH scale [94], the Munich Chronotype questionnaire and Sleep Disturbance Scale for Children [64], and the Epworth Sleepiness Scale [96]. To measure stress, the stress subscale of the Depression Anxiety Stress Scale (DASS) [92,93], the INTERHEART stress questionnaire [59], the Psychosocial Well-being Index-Short Form [91], the Perceived Stress Scale [61,64,88], and the Profile of Mood States [81] were used. Regarding data collection through sensors, most studies used wearable devices. One study used a smartphone to obtain points of interest related to physical activity and diet [98], and 1 sleep study used polysomnography [55]. Finally, 2 studies used words related to physical activity and diet, 1 derived from Google Trends [99] and the other from Twitter (Twitter, Inc) [100].

Concerning the modeled inputs, 56 of 65 (86.1%) studies used multimodal data. The input modalities were lifestyle (60/65, 92.3%), sociodemographic (49/65, 75.3%), clinical (29/65, 44.6%), anthropometric (14/65, 21.5%), psychological (20/65, 30.7%), physical (3/65, 4.62%), environmental (10/65, 15.38%), physiological (3/65, 4.62%), and behavioral (2/65, 3.07%). The model outcomes included lifestyle domains in 14 (22% studies; 5 physical activity [40,44-46,90], 5 sleep [10,48,86,97,101], 3 diet [62,66,94], and 1 stress [61]) and other health outcomes in 51 (75% studies; with mental health, cancer, cardiovascular diseases, and diabetes being the most frequent categories; refer to Table 1).

Cross-sectional data were acquired in 36 (55%) studies [42-44,51-53,56-58,60-62,65,66,71-74,77-80,84,85,89,92,94,95,97], longitudinal data in 18 (28%) [10,41,50,54,55,59,64,67-69,81,82,96] studies, time-series data in 7 (11%) [45,46,48,75,76,101,102] studies, combined longitudinal and time-series data in 1 study [90], textual data in 2 studies [99,100], and combined cross-sectional and geographical data in 1 study [103].

The preprocessing phase was divided into variable transformation, missing imputation, resampling, and dimensionality reduction. At least one of these preprocessing phases was reported by 59 (90.76%) studies (refer to Table 1 and Table S4 in Multimedia Appendix 1 for more details).

There were 26 (40%) studies that reported normalization or other arithmetic or statistical transformations of variables before the modeling phase [40,41,45,48,50,51,53,58,59,62,64,67,69,73-75,79,81,83,84,89,91-94,102]. Six (9.23%) studies recoded categorical variables into quantitative variables, 5 used one-hot encoding [10,53,64,67,102], and 1 used principal component analysis with quantile transformer scaler [42].

Missing data imputation was reported in 37 (56.92%) studies. Twelve papers simply removed cases with missing data [10,42-44,61,64,77,78,80,93,98,101], while others applied cutoff percentages for missing values (eg, 10% [87], 30 % [73], or >50% [58,72]), and 1 study removed observations with missing values in the output [102]. Techniques included single imputation (mean, median, or mode) [51,52,58,59,70,83,88,92], multiple imputation by chained equations [57,63,72,73], k-nearest neighbor [89,102], regression-based algorithm [57,81], random forest (RF)–based multiple imputation [55,69,84], the MissForest algorithm [67,68], imputation based on peers with similar health profile group [82], imputation using training data [104], and replacement of missing values with the last available data [75].

Resampling techniques were reported in 21 (32.30%) papers. Eighteen studies balanced the datasets using methods such as using the minority class as a reference, undersampling the majority class [42,61], or the synthetic minority oversampling technique (SMOTE) [43,44,52,60,63,67,71,72,84,86,89,91,94,95,97]. One paper compared the results of SMOTE against the adaptive synthetic algorithm [65]. Finally, 1 study stabilized variations in underrepresented outcome classes using bootstrap resampling [82]. Regarding cancer studies where cases were fewer than controls, 2 different strategies were applied to the same dataset: sample-size equalization by randomly grouping cancer-free participants based on the number of cancer survivors [74], while another study matched cases and controls by sex, age, and education level, then selected a random sample resulting in 50% cases and 50% controls [73].

Dimensionality reduction was used in 38 (58.46%) studies using 3 approaches. The first approach involved assessing the relationship between features and outcomes by removing redundant information [55,61,64,70,83,85,87,93,98,99]. Other methods included factor analysis [10] and principal component analysis [42,59,69]. The second approach optimized models to achieve lower prediction error [58,66,104]. The third approach involved automatic selection of predictors during model training [41,49,57,62,72-74,77,78,81,90,95,102]. For more information, refer to Table S4 in Multimedia Appendix 1.

Final model evaluation procedures were explicitly reported in 57 (87.69%) reviewed studies, with the hold-out test set being the most applied strategy. These included a hold-out test set [10,42,48,49,52,53,57,60,61,63,65-67,69,71-73,75,79,82-85,94,95,100,101], k-fold cross-validation (final evaluation) [40,41,43,44,47,50,51,58,68,70,74,78,81,87,88,91,92,97,98,104], nested cross-validation [64,93], leave-one-out cross-validation (for small datasets with n<150) [45,59,76]. Two studies used external datasets for model performance assessment [55,86]. For time-series data, 2 studies divided the dataset based on the time of acquisition, keeping an out-of-sample dataset for model evaluation [46,99], while 1 study [102] used 5-fold cross-validation in an offline setting followed by an online weighted resampling methodology to address drift.

Among the 11 studies that addressed regression problems, the reported evaluation metrics included mean absolute error [44,75,87,102], mean squared error [44,68,75], root-mean-square error [44,45,68,78,98,99,102], mean absolute percentage error [102], and coefficient of determination (R²) [44,68,79,81,98,99].

For classification problems, reported evaluation metrics included specificity and sensitivity (recall) [41-44,47,48,58,63,66,72,76,82,85,86,91,93,95,104], precision and recall [40-43,48,61,70,80,86,100,101], the confusion matrix [42,48,80], error rate as the proportion of misclassified observations (1 – accuracy) [48,58,72,90], Cohen κ [58,72], F₁-score [41,43,49,59,70,80,86,91,100,101,104], and model training time [58]. The most frequently used metrics were accuracy [10,41-45,47-49,58,59,61,63,66,70,72,80,82,83,85,86,88,91,93,101] and area under the receiver operating characteristic curve [41-43,46-49,58,59,63,70,72-74,76,83,85,86,91,93,95,100,101,104] (Table S5 in Multimedia Appendix 1). Additionally, 1 epidemiologic study used the Brier score to assess cardiovascular mortality [53].

To enhance explainability, 22 (33.84%) studies implemented specific XAI methods to clarify the contribution of each predictor and the direction of the relationship with the outcome. Of these, 21 studies used Shapley Additive Explanations (SHAP) values, while 1 used local interpretable model-agnostic explanations (LIME). As shown in Figure 6, XAI methods, particularly SHAP values, have been applied across different health domains, showing a remarkable increase in use in 2025. The most common visualization methods were beeswarm and bar plots, with the contribution of each feature ranked along the y-axis, placing the most important feature at the top of the plot. The number of features plotted ranges from 6 to 50, with 20 being the most common (refer to Table 2).

Because the nonlifestyle features included in each model vary substantially across studies, the ranking reported in Table 2 should be interpreted in relative terms. When a lifestyle behavior appears among the top-ranked features, this indicates that it contributed more strongly to the model than nonlifestyle variables included in the analysis. Conversely, lifestyle components appearing in lower positions acted as secondary predictors.

In a study on cardiometabolic disease [76], the ML solution was first explained at the individual participant level to provide specific behavioral feedback, and then at the group level to reveal the ranking of features for succeeding in behavioral changes. In both models, physical activity and diet variables were among the top contributors. However, in another study [65], neither regular physical activity nor diet habits were among the top 6 variables for predicting diabetes in adults.

In 4 mental health studies [64,67,80,85], SHAP values were used to rank the contribution of each feature. In 3 of the 4 studies, lifestyle variables were among the top contributors. In [80], physical activity and fruit and vegetable consumption; in [64], sleep variables were the top predictors of stress in young adolescents during the COVID-19 pandemic; and in [85], sleep duration was identified as important for predicting depression. However, in [67], the intensity of physical activity was unranked because it was excluded from the XAI analysis during a feature selection step prior to modeling. In the same study, sleep hours were the most important variable for predicting depression among adults. Regarding mental health studies, psychological resilience was assessed among medical students [56], with sleep disturbance being a key factor affecting their resilience.

Three studies focusing on older adults in China had cognitive function as their outcome [50,52,63]. In these studies, physical activity variables were ranked as top predictors. Additionally, in 1 of these studies [63], sleep quality was ranked as the eighth top feature out of 20. Furthermore, when loneliness was assessed among older adults from China [51], exercise was ranked as the seventh variable out of 15.

In a study predicting quality of life [91], stress, sleep quality, and physical activity emerged as the strongest predictors, with the eating index appearing in the top 15 variables.

Two studies focused on sleep as the specific outcome. In [86], the SHAP values ranked stress score, vegetable consumption, and sedentary time among the top 5 variables for predicting sleep disturbance. In [97], nutritional status was the third most important variable for predicting the risk of sleep disorders in older adults.

In [88], “favorable” food stores and global stress were identified as the top variables for predicting incidence of cardiovascular disease, with the availability of outdoor activities ranking in the top 5. In a longitudinal study investigating the association of diet with long-term reduction in waist circumference, SHAP values highlighted the importance of high-quality components in reducing visceral fat [83]. This study also measured exercise with a single item of frequency, which was not included in the top ranking. Regarding a diet study [94], the adoption of different diets in the Spanish population was assessed, with fish consumption positioned as the most important variable for all diets studied (Mediterranean, intermittent fasting, vegan, and vegetarian). In the same study, practicing sport was ranked among the top variables across the different diets. However, sleep quality was only included in the ranking for intermittent fasting and the vegan diet.

Physical inactivity emerged as the most important feature in explaining county-level obesity using LIME [87]. In this study, the food environment and insufficient sleep, both measured as single items, were not included as top predictors of obesity prevalence.

In a cross-national study assessing frailty [68], sleep variables were among the 20 most important variables across all the cohorts studied. Physical activity was included in the US and UK cohorts and was also ranked in the top 20 variables. In the UK cohort, fruit consumption was the 19th out of 20 key contributors to frailty.

One study on osteoporosis [84] used the Life’s Essential 8 scores for physical activity, diet, and sleep. Sleep and physical activity scores were among the 20 most important variables, but the diet health score was unranked. Similarly, the Life’s Essential 8 scores were used to determine infertility risk in women [60], with the diet score being the only one ranked among the top 10 variables.

Finally, 1 study combined lifestyle factors with metabolites associated with psoriasis [71], with physical activity intensity ranked among the 20 key factors out of 150 variables, whereas dietary habits were not ranked.

Of the included papers in the review, 23 (35.38%) used R software (R Core Team) [105] for data analysis (Table 3). The R packages used were data.table for data manipulation, tidyverse [106] as a general package for data science, Multivariate Imputation by Chained Equations [107] for missing data imputation, missMDA for performing multiple imputation with principal component analysis, FactoMineR for exploratory data analysis and principal component analysis, Boruta [108] for feature selection through a wrapper algorithm, caret (Classification and Regression Training) [109] for creating models, randomForest for RF analysis, randomForestSCR for RF for survival, regression, and classification analysis, rpart for recursive partitioning and regression trees, xgboost for extreme gradient boosting, bartMachine for Bayesian additive regression trees, kernlab and e1071 for support vector machines, survival for survival analysis, lime for local interpretable model-agnostic explanations, and finally, SuperLearner [110] to choose the optimal learner for a given prediction problem with a k-fold cross-validation algorithm.

In contrast, 24 (36.92%) studies developed the models in Python (Python Software Foundation), using the following libraries: Scikit-learn, used in all studies for predictive data analysis; pandas, for manipulating tabular data; NumPy, for mathematical functions; Keras and TensorFlow, for implementing deep learning; lightGBM, for performing light gradient boosting machine; SHAP, to explain ML solutions; creme, for online ML; Bayesian optimization, as a global optimization package to find the maximum value of an unknown function in as few iterations as possible; imbalanced-learn, to combine either undersample or oversample methods; and TextBlob, emoji, nltk, and profanity, for processing and analyzing textual data.

Finally, 4 studies used both Python and R, and 6 studies used other software programs such as SPSS, MATLAB, and KNIME. Eight papers did not report the software used [10,40,59,60,75,82,85,94].

This scoping review of 65 studies provides the current state of the application of supervised ML algorithms for the analysis of lifestyle data. The increase in studies in this field since 2019 indicates that it is a noteworthy area of study. The diversity in the sample origin, alongside the accessibility to new AI tools and novel methods for monitoring health outcomes (eg, wearables), denotes global attention to lifestyle. This section addresses the methodological shortcomings found in the reviewed studies.

In relation to lifestyle data, we found that most studies adopted a multidomain approach, integrating more than just a single component. This strategy enhances and facilitates a more comprehensive understanding of health problems related to the 4 lifestyle domains considered in this review: physical activity, diet, sleep, and stress. The distribution of lifestyle domains identified here was similar to that reported in a previous scoping meta-review [113], although we observe that sleep has gained prominence in recent years, now reaching a level comparable to diet. Although these results highlight the growing recognition of the interrelated nature of lifestyle behaviors, the imbalance in the distribution of these factors limits the capacity of current studies to fully model and understand the interaction among the 4 lifestyle components and their combined effects on health.

Concerning the data acquisition process, over half of the studies acquired their own data. This acquisition process implies control over variables and reduces the time required for cleaning [114]. Interestingly, both self-acquired datasets and those sourced from private or public health repositories demonstrated gender parity in the analyzed datasets. However, we detected a major limitation in this part of the process regarding the data acquisition methodology. In most studies, data were collected through single items, such as regular physical activity (response “yes” or “no”) [58] or usual time of waking up and going to bed [62], resulting in low representativeness of the construct being measured. The result of this acquisition method is high heterogeneity in measures, which hinders their generalizability. Therefore, the quality of data must be one of the challenges to be addressed, and specifically, the consistency in measures [115,116].

Nevertheless, the current accessibility and precision of health sensors such as wearables [117] and the Internet of Things [118] may contribute to transferability and actionability in the population [119]. The growth in technology allows the integration of different data forms as well as more objective measures of lifestyle, substantially reducing the impact of retrospective bias by tracking real-time data in an ecological situation [120]. Therefore, merging questionnaires and sensor data may be the key to identifying relationships between lifestyle measurements and personalizing interventions or changes in specific behaviors. This integration would include physiological, psychological, and behavioral factors, which are the most common analysis types in the ML community to extract clinical insights [121].

Regarding the analysis of ML, 2 different approaches emerged in the reviewed studies: 1 focused on prediction through classification and regression problems, and the other focused on interpretability through feature selection. The first is already an acknowledged approach, while the second typically constitutes an important component of the ML process, specifically during the preprocessing stage. However, feature selection studies do not use model evaluation metrics, which can limit their statistical validity and the generalizability of results. Remarkably, the family of ML algorithms most closely related to feature selection is tree-based because it provides indices of the importance of each variable. Although most papers in this scoping review combined different families of algorithms and compared their results, the most common model family was tree-based, which was applied for each data typology identified. Specifically, RF is the most used algorithm, which may be due to its robustness in handling missing values, the consideration of complex interaction in the data [122], and its lower sensitivity to variable scales [123]. Despite the benefits of RF, the underusage of DL algorithms represents a critical missed opportunity for robustly analyzing complex and multimodal data. In this review, DL algorithms were underused for lifestyle data. This result may reflect a gap in expertise or access to computational resources among lifestyle researchers, potentially limiting the application of more complex models. With ongoing advances in computational power and algorithmic efficiency, it is expected that the use of DL algorithms will become more widespread in the near future [124].

Regarding the preprocessing stage, most studies detailed some phases of the process, but there is no consensus on the description of this stage of ML. Variable transformation is a crucial step for certain algorithms, particularly for SVM and specific architectures of DL that exhibit sensitivity to the raw form of the variables. In this review, 12 out of 19 studies that performed the SVM algorithm, and 13 out of 26 studies focused on DL, reported variable transformation. Additionally, it is worth noting that these algorithms cannot handle missing values, requiring imputation before the modeling phase. Among the SVM studies, 11 out of 19 reported techniques for missing imputation, and 17 out of 26 DL studies explicitly addressed this. In contrast, tree-based algorithms are less sensitive to variable scales and missing values, yet incorporating these feature engineering steps could enhance model performance [123]. It is noteworthy that the preprocessing steps, specifically how missing values are addressed, have been identified as a potential concern for transparency. This procedural aspect could introduce sampling biases, thereby influencing the generalizability and comprehension of the dataset context [125].

Resampling techniques, aiming to balance the dataset, are commonly implemented in classification problems. SMOTE has been the most widely used technique in this review, especially because it achieves better results than a simple undersampling of the majority class. In the health domain, imbalanced datasets are common, and SMOTE oversamples the minority class with synthetic examples and randomly undersamples the majority class to balance the dataset [126].

Finally, dimensionality reduction enables capturing the most relevant information for the outcome while eliminating noise and redundant information. In this review, dimensionality reduction was the most frequent preprocessing step, appearing in 38 studies. Notably, not only could SVM and DL models benefit from removing irrelevant predictors, but also in tree-based algorithms, dimensionality reduction minimizes model complexity, resource consumption, and data acquisition costs [123].

The division of the original dataset is an essential step for assessing the performance of the ML solution. In this review, 8 studies did not report how they split their data to assess the model, which denotes a lack of generalization of their results. This omission represents a common issue in ML research that should be carefully addressed to minimize bias [127]. Train and test division, also known as “hold out,” is a method with considerable variability due to the use of a unique random data distribution [127]. Therefore, other methods might be more suitable. For example, leave-one-out cross-validation, which trains the model on “n – 1” observations and makes predictions on the remaining one. Although effective for small datasets, it is computationally intensive with large datasets [128]. K-fold cross-validation involves randomly dividing the original dataset into k groups. K-fold cross-validation not only offers computational advantage over leave-one-out, but also gives more accurate estimations due to the bias-variance trade-off [128]. In time-series data, only 1 paper [102] introduced a different form of data split, considering the dependencies of the entire series. Ideally, this type of data should be treated with a method called rolling forecast origin resampling, which estimates the model with historical data and evaluates it with the most recent data [129]. In other words, the training set should ideally comprise observations that occurred before those in the test set; however, this method was not found in this review.

Regarding the evaluation of ML models, the choice of metrics depends on the nature of the problem, whether regression or classification. Specific evaluation metrics tailored to each problem are crucial for correct evaluation, aligning with the priorities and needs of each field. For instance, in medical studies where the cost of treatment in terms of health is high, it becomes crucial to identify true patients over false positives. In contrast, if the treatment has minimal side effects and has demonstrated benefits, sensitivity might not be as important as specificity. It is worth highlighting that in the field of data science, precision and recall are more commonly used, whereas in medical fields, specificity and sensitivity are more prevalent [18]. These differences may cause misunderstandings between the 2 domains.

Regarding model explainability, 22 studies incorporated a dedicated step in the ML process for explainability. SHAP values and LIME were the only XAI methods applied to lifestyle data, and these 2 methods were the most common in a recent systematic review on XAI methods [130]. XAI-related studies in this review were published since 2019, with a notable increase in the number of publications in 2025, comprising half of the papers. This exponential growth was also found in [27,28], where the trend in published papers occurred between 2016 and 2022. Thus, our review demonstrates this exponential distribution in the health and behavioral sciences, where XAI methods are gaining prominence.

Although tree-based algorithms, especially decision trees, are known for facilitating interpretation, SHAP values can be applied to any type of model [131]. The adoption of XAI in lifestyle studies remains low (33.84%). One possible explanation for this is that explainability algorithms are often not integrated into a standard ML pipeline, thereby increasing the technical complexity of the workflow. However, some efforts are being made by R and Python developers to incorporate XAI algorithms into pipelines using libraries such as H2O [132].

In this review, lifestyle components (physical activity, diet, sleep, and stress) consistently appeared among the top-ranked features in models using XAI techniques, highlighting their substantive contribution relative to nonlifestyle variables. These findings align with prior research emphasizing the integration of diverse lifestyle components [133,134]. However, the level of interpretability achieved also depends on the quality of data used in the models, which in some studies did not meet expected standards. For instance, in a study where the focus was on obesity, the diet component was not among the top-ranked features [87]. In this county-level study, a food environment index was measured as a single item, potentially inadequately representing the diet component of lifestyle.

Therefore, integrating XAI methods into the ML process could enable tailored interventions based on model results, provided that measures are collected accurately. Furthermore, the adoption of XAI algorithms contributes to increased trust and verification of the fairness of the models. This approach can also facilitate the translation of findings to stakeholders and health systems, thereby enhancing transparency, promoting the adoption of models in society, and supporting informed decision-making [27].

Competition between Python and R for ML software dominance in data science is currently intense. Both Python and R are freely distributed, object-oriented software with large and active communities. Python, as a programming language, offers specific implementations through libraries tailored for statistical analysis, including ML and DL. In contrast, R is a statistical software that integrates fundamental statistics into its base functionalities. While Python requires libraries for each stage of analysis, its well-established libraries streamline the process. On the other hand, R faces challenges due to its heterogeneous libraries, which hinder replicability and require expertise in varying syntax across packages. To address this concern, the meta-package tidymodels (Max Kuhn and Hadley Wickham) resolves these issues by integrating all necessary packages for each ML step, using a unified syntax. Additionally, tidymodels integrates user-friendly interfaces and promotes good methodological practice, thereby preventing user errors [135]. Conversely, Python presents a preferable environment for DL with the TensorFlow and PyTorch frameworks. In this regard, the possibility of developing ML projects on powerful computational cloud-based platforms, such as Google Colaboratory (also referred to as Google Colab) [136], offers Python a remarkable advantage over R local environments.

Based on the review’s results and to enhance transparency and replicability in multidisciplinary sciences [137], we provide comprehensive methodological and reporting guidelines and a checklist for ML projects. Although various studies have proposed guidelines and checklists [138,139], the rapid expansion of ML algorithms in health domains necessitates iterative evaluation to incorporate new steps into the ML research workflow. The guidelines and checklist (Checklist 2) are based on the 5 stages of the ML workflow, as depicted in Figure 2, with added software tools.

Our guidelines focused on SL algorithms for classification and regression problems. The choice of algorithms depends on the measurement of the outcome. We recommend using multiple algorithms to compare results and select the best fit, given that there is significant variability across problems [145]. Additionally, comparing different families of algorithms is also advisable, as some improvements exist within the same family. When it comes to replicability and transparency, reporting algorithm hyperparameters is crucial in ML problems, as different configurations can yield varying results [146].

Avoiding overfitting requires dividing the original dataset appropriately. This step is fundamental in ML implementation and should be considered in every study to ensure that the extracted insights are reliable and generalizable. This division depends on data typology; we identified 3 different typologies relevant at this stage, defined by whether time is an implicit factor in the data acquisition process.

Python and R were the most widely used software in the review. However, the variety of libraries can complicate the process. To address this, the metapackage “tidymodels” provides unified syntax, enhancing replicability [135]. For DL, Python offers a more powerful environment thanks to PyTorch and TensorFlow frameworks and the possibility of developing analyses on cloud services such as Google Colab. Furthermore, we encourage researchers to make their code publicly available in open repositories such as GitHub.

At the level of the review itself, some limitations emerge. Although time-series data were frequently used in the analysis of lifestyle behaviors, the primary focus of this review was on the methodological framework of supervised ML, rather than on specific time-series modeling approaches. Future research should specifically address this gap by examining the current use of time-series models applied to wearable data to better capture human behavior. Moreover, the exclusion of UL algorithms limits the scope of ML algorithms covered in this review. For example, cluster analysis has been used to classify children according to their eating behaviors and identify features related to obesity [155], with findings suggesting that interventions such as reducing eating speed may help prevent childhood obesity. This scoping review focuses on supervised ML approaches, but future research could examine the current application of UL in lifestyle data. Finally, a further limitation of this review is that the search strategy prioritized studies using the umbrella term “healthy lifestyle.” While this approach captured a representative sample of research explicitly conceptualized as HL, it may have excluded multidimensional studies examining combinations of physical activity, diet, sleep, or stress that did not use this terminology. Although an interaction block was incorporated to address the interconnectedness among these domains, we acknowledge that some multidomain research may not have been retrieved due to the specificity of the search string.

This review has identified several limitations within the studies reviewed that need to be addressed. First, ensuring data quality remains a significant challenge that must be addressed by carefully selecting data acquisition methods to build reliable and robust models. Second, the evaluation process is crucial for preventing overfitting, and using hold-out cross-validation can lead to high variance partitioning. Therefore, it is recommended to implement k-fold cross-validation at various stages, such as during validation; for time-series data, rolling forecast origin resampling is recommended [129].

In conclusion, this scoping review provides a comprehensive analysis of lifestyle using ML models and serves as a guideline for future research. While the relationship between lifestyle and health is well-established, ongoing efforts are needed to refine how we measure lifestyle to create robust models. It is essential to focus not only on model performance but also on data representativeness, which is closely related to the granularity established during data collection. Although RF algorithms are prominent in lifestyle data analysis, it is recommended to compare their performance with other algorithms within and across families. Future research should also incorporate SHAP values to enhance interpretability within the ML workflow. Additionally, the tidymodels metapackage (R software) with H2O for XAI can assist researchers in evaluating process quality with unified syntax, thereby contributing to replicability.