All procedures were approved by the Commission Cantonale d’Ethique de la Recherche sur l’être humain with the number 2023‐00975. Written informed consent was obtained from all participants before starting data collection. Participants received no monetary compensation.

Data used in this work came from the Providemus alz project [28], an ongoing observational cohort of community-dwelling adults (45.61‐77.62 years old), being conducted at the University of Geneva, Switzerland. Participants (cognitively healthy individuals) volunteered to wear a consumer-grade smartwatch (Withings Steel HR) continuously and actively complete wave-based assessments approximately quarterly (every 90 days) over 2 years using a custom-made mobile app (mQoL, based on the mQoL Living Lab infrastructure [29]) available for iOS and Android devices. Data collection started in March 2024 and lasted until April 2026.

The provided smartwatch was used to collect technology-reported outcomes (like PA, heart rate, and sleep) passively (referred to as passive data in the rest of this paper). Patient-reported outcomes (like depressive symptoms) and performance-reported outcomes (like processing speed) were actively obtained (referred to as active data in the rest of this paper) via an interaction with the participant through the mQoL app available for iOS and Android devices.

For this paper, we used Providemus alz data collected until August 20, 2025, corresponding to a maximum of 6 repeated assessments using validated questionnaires and tests (waves) and a maximum of 528 days of wearable data per participant. Figure 1 illustrates the data collection timeline.

Following the data quality criteria applied in previous research [30,31], data from 67 participants were included in the analysis (Table 1 provides the baseline characteristics). Validated assessments covered 20 cognitive functioning and mental health outcomes, but in this paper, we focused only on 2 clinically relevant measures: processing speed (measured using the Trail Making Test [32] part A) and depressive symptoms (a suboutcome of the Hospital Anxiety and Depression Scale [HADS] [33]).

In total, 29 variables served as predictors and were sampled at 2 frequencies (daily and once per wave), as presented in Textbox 1.

For the predictors sampled daily, wear time was defined in a previous publication [28]. The percentage of missing windows per day was used so the model could account for the incompleteness of some linear or quadratic fits (introduced later in this section). The sleep regularity index quantified night-to-night consistency in sleep-wake timing; it captures temporal regularity of sleep behavior, computed by comparing sleep-wake states across consecutive days in 15-minute time bins and adjusting for time zone shifts.

For the predictors sampled once per wave, Cognitive Reserve Index questionnaire (CRIq) and Mediterranean diet score (MDS) values were only obtained at baseline and, therefore, had a constant value across all waves. Chronological age was computed based on the date of the active assessment. The difference between subjective and chronological age was obtained by first asking the participants at every wave to select the decade with which they identified the most then subtracting its midpoint value from the chronological age (negative values indicating feeling older than their actual age). The “number of assessment tries” was used to account for the cases in which, for technical or personal reasons, participants started to complete the assessments but did not finish them, having another try some time later in the same wave; it served as an indirect measurement of potential learning effect within a wave. At the end, 29 variables were used as predictors.

The 2 outcomes studied, processing speed and depressive symptoms, were collected once per wave and matched with daily and once-per-wave data while preserving temporal order. Baseline characteristics of both outcomes across all the data used in this research are provided in Table 2. For each outcome measured on day D, the analysis included only predictor data (daily and once-sampled) recorded before D.

No data imputation was performed on either the predictors or the outcomes at any stage of the analysis; all models were trained and evaluated using only the available observed data.

To address the study objectives concerning the relationships among 24-hour movement behaviors, the temporal dynamics, and cognitive functioning and mental health outcomes, we applied the PEPHA framework. It is an adaptable analytic approach designed for observational and ecological settings, enabling the integration of high-frequency wearable data with lower-frequency, validated outcome measures. By leveraging naturally occurring behavioral variation, the framework characterizes, at both population and individual levels, which behavioral dimensions are potentially associated with a given outcome and how these associations vary across time within an observation period based on historical individual data.

The framework was implemented using a regressive support vector machine (SVM) with a radial basis function (RBF) kernel. This model was selected for its robustness in moderate sample sizes, ability to capture nonlinear associations, and stability in high-dimensional feature spaces (characteristics typical of longitudinal aging datasets derived from wearable sensors). Within PEPHA, the SVM is not used for outcome optimization or causal inference but as a stable function approximator that enables systematic comparison of behavioral representations across time windows, lag structures, and individuals. All framework parameters are generic and configurable (eg, wave duration), allowing replication and adaptation to other datasets and outcomes through the accompanying open-source implementation [37].

PEPHA begins by aligning high-frequency passive data with lower-frequency outcome assessments. Each outcome collected at wave W+1 is paired with the passive data recorded during the 90 days preceding it (wave W), so that predictors always refer to the 24-hour movement behaviors before the outcome. As the assessment day of each outcome can differ, that 90-day window is dynamically computed depending on the outcome being studied at a time. An illustration of this alignment is in Figure 2.

The framework uses two key parameters at this stage: (1) TARGET_DAYS, which defines the retrospective window length (eg, 90 d), and (2) MIN_VALID_WAVES, which defines the minimum number of valid waves required for a participant to be included in the next steps.

A wave for a specific participant is considered valid if (1) at least 50% of its daily data are available (defined as having at least 10 h of wear time, as in [31]) and (2) the outcome value at W+1 is present.

In the Providemus alz application, these rules ensured that only consistent and interpretable wave-outcome pairs were retained, yielding a robust set of participant-wave samples for modeling. Given the outcome assessment under noncontrolled mobile settings, processing speed results were only considered valid if within the mean ±2.5 SD of the data initially collected. Depressive symptoms scores were not filtered, as they were assessed regardless of completion times and using a standardized and validated questionnaire.

Following our previous analyses of the Providemus alz cohort [31], we tested the modeling of processing speed and depressive symptoms using levels (absolute scores) and absolute drifts (deviations from prior medians), given the differences found earlier. This distinction reflects the 2 temporal perspectives that PEPHA can capture. In the levels variant, the model uses the temporal slopes of daily predictors during the 90 days preceding each assessment to predict the state of the outcome at that assessment (eg, how recent [90 d] fluctuations in PA or sedentary lifestyle relate to concurrent cognitive or affective levels). In the drift variant, the model instead links the change in outcome values between consecutive waves to the patterns of change in 24-hour movement behavioral dynamics within each 90-day window, thus focusing on how habits changes accompany longer-term shifts in brain-health trajectories.

For the methods of this research and following the data inclusion criteria described earlier, the levels models were trained on data from 66 participants (average per participant: 4.67, SD 0.69) for processing speed and from 67 participants (average per participant: 4.52, SD 0.82) for depressive symptoms. The drift models, in contrast, included 55 participants (average per participant: 4.91, SD 0.29) for processing speed and 54 participants (average per participant: 4.76, SD 0.58) for depressive symptoms. This difference reflects the fact that drift models require 2 consecutive waves to estimate within-person change and therefore begin at wave 3, whereas levels models begin at wave 2, which meant participants who stopped adhering to the study’s protocol before wave 3 could not be modeled in the drift case.

Once valid daily data are selected, PEPHA transforms them into compact temporal summaries to reduce the influence of short-term extremes and noise. The 90-day period is divided into consecutive, nonoverlapping windows of WINDOW_DAYS (typically between 2 and 7), balancing the need for sufficient observations to compute reliable summary statistics with the ability to capture short-term behavioral variation. Each window is summarized as a 24-hour movement behavioral “snapshot” from which 8 descriptive statistics are then computed: mean, median, SD, IQR, minimum, maximum, skewness, and kurtosis. This produces a temporal sequence of windows for every predictor.

The sequence of descriptive statistics is then split into 2 halves—early (Half1) and late (Half2)—corresponding to the first and second halves of the observation period between the waves (45 d each). For each half, PEPHA fits both linear and quadratic regressions to estimate the trend slopes that summarize change over time. Optionally, the SE of each slope can also be computed to account for trend fit, controlled by a parameter in the framework referred to as USE_SE. The result is a structured set of slope-based features describing not only how much each 24-hour movement behavioral variable changes (linear) but also how within the wave these changes occur (quadratic).

In PEPHA and across both levels and absolute drift variants, the model learns from temporal trends (relying on predictors being the linear and quadratic fits) rather than static levels (eg, having access to the actual sequence of PA data), emphasizing how 24-hour movement behaviors evolve rather than how high or low they are. This design allows PEPHA to identify which behavioral dimensions and which temporal patterns best explain individual differences in outcomes, independent of baseline offsets.

For the example application of the framework here presented, no data imputation was applied; missing values were retained throughout the pipeline, preserving the natural irregularities of real-world behavior. Regressions were fitted directly on the available data, respecting the temporal spacing of missing windows rather than interpolating them. The resulting features provide a compact and interpretable representation of temporal behavior for each wave.

Because PEPHA integrates multiple analytical components (temporal windowing, regression feature extraction, and support vector modeling), the performance of its modeling phase depends on several interacting complexity parameters. Selecting these parameters manually or through grid search would be inefficient and prone to suboptimal results. To ensure reproducibility and avoid arbitrary choices, PEPHA uses Bayesian optimization via Optuna [38], an adaptive search algorithm that iteratively explores the parameter space to identify hyperparameter combinations that balance predictive accuracy, correlation strength, and variance fidelity under a consistent evaluation procedure.

Parameters jointly optimized include the following. MIN_VALID_WAVES is the minimum number of valid waves per participant to be included in the modeling and analysis. WINDOW_DAYS is the length of each summarization window in the predictors data. USE_SE indicates whether SEs of the fitted linear and quadratic trends are used as predictors aside from the trends themselves. CORR_THRESHOLD is the maximum absolute correlation allowed between redundant predictors before one is dropped; it is applied to avoid information repetition, as detailed later in this section. SVR_C, SVR_epsilon, and SVR_gamma are core regressive SVM model hyperparameters controlling regularization, insensitivity margin, and kernel width.

Each Optuna trial performs 10-fold grouped cross-validation (grouped by participant to make sure one participant can never have data in both train and test sets at the same time, which would represent a data leakage case) and returns a single objective score that integrates 3 complementary criteria:

Objective=0.5\ ×\ MAEz-1.2\ ×\ r+0.5\ ×\ |σpredσtrue-1|(1)

where MAEz is the mean absolute error (MAE) normalized by the SD of the true outcomes (a scale-invariant measure of prediction error), r is the Pearson correlation between predicted and observed outcomes, and σpred/σtrue quantifies how well the predicted variability matches that of the real data.

This formulation penalizes models that minimize MAE by collapsing variability (a common issue in regression with small samples) or that produce predictions with poor correlation with the target values despite low error. The 3 terms respectively ensure accuracy, directional fidelity, and variance realism, yielding models that generalize better across heterogeneous behavioral patterns. This multiterm objective extends beyond conventional MAE minimization by discouraging variance collapse, a frequent pitfall when small samples lead models to reproduce the mean rather than data variability.

By minimizing this composite objective, Optuna identifies the parameter configuration that provides the best trade-off between fit and generalization without over-regularizing or inflating variance. Once the optimal parameters are found, they are used consistently in the subsequent PEPHA stages for feature importance analysis, temporal order testing, and change point detection.

For the proof-of-concept application of PEPHA presented in this paper, Table 3 details the Optuna parameters explored across 500 tries. For the gamma range of values, “scale” is an adaptive γ based on feature variance; “auto” is a fixed γ=1/number of features, where γ is the kernel coefficient controlling how tightly the regressor fits around each training point (lower=smoother, higher=more localized).

The resulting optimal configuration was then used for the models described in the following sections, ensuring that feature importance and temporal analyses operated under the best validated parameter set.

After generating slope-based predictors and deciding on the optimal complexity parameters to use, the goal of PEPHA is to estimate how strongly each 24-hour movement behavioral predictor potentially relates to the cognitive and mental health outcomes across the population and for an individual. This is done by training a regressive SVM model that maps all predictor features (linear and quadratic slopes, optional SE, and control variables that are reported once per wave, such as age or education) onto each of the outcomes of interest.

To avoid duplicated information conveyed by the predictor’s slopes and SEs, highly correlated predictor pairs (mean and median, IQR and SD, minimum and maximum, kurtosis and skewness) above a certain |ρ|=CORR_THRESHOLD in both halves have the second of them (median, SD, maximum, skewness) removed from the predictor set. Spearman rank correlation was chosen for this redundancy filtering step because it captures monotonic but potentially nonlinear relationships between predictors. In contrast, Pearson correlation, which is used in the Optuna parameter optimization, quantifies linear agreement on the continuous scale of the outcomes. This distinction ensures that redundancy removal remains sensitive to nonlinear codependencies, while model evaluation accurately reflects quantitative predictive power.

Key parameters controlling this stage include CORR_THRESHOLD, SVR_C, SVR_epsilon, and SVR_gamma.

Training then follows a leave-one-subject-out (LOSO) cross-validation (CV) strategy with z-normalization applied using training-fold statistics only (ie, normalizing per fold to avoid data leakage). Model accuracy is assessed using MAE, expressed in both natural units and as a percentage of the outcome’s scale (scaled MAE [SMAE]). For processing speed, SMAE values were obtained by dividing MAE by 14 (the empirical range between the minimum and maximum scores after filtering data within mean±2.5 SD), whereas for depressive symptoms values were divided by 22 (the theoretical range of the HADS).

Given the use of SVM as the PEPHA model, feature relevance was quantified via permutation importance, assessing the increase in prediction error when a given feature was randomly permuted. At this stage, averaging across all folds (all participants) to obtain the population ranking of features most predictive of the outcome, PEPHA provided a first-level answer to RQ1—identifying which aspects of 24-hour movement behavior most inform the resulting outcomes of processing speed and depressive symptoms.

RQ2 asked whether the timing of behavioral change within each 90-day wave period was potentially associated with the processing speed and depressive symptoms at its end and whether individualized “potential temporal association windows” can be identified within each observation period.

In the PEPHA framework, this is operationalized through a half-swap analysis, in which predictors from the first and second halves of each wave (Half1 and Half2) are swapped to estimate how temporal order is linked to outcome predictions at the individual level. Note that the internal orders of both halves internal are maintained (Half1 from day 1 to 45 is kept in the same order 1 to 45). Thus, this test quantifies each participant’s sensitivity to when behavioral change occurs rather than how much change occurs. This sensitivity is interpreted as a marker of temporal association or responsiveness within the observation window, rather than as evidence that a behavior should occur at an exact time point in real-life intervention settings.

For each individual and wave, the model predicted the outcome twice: (1) using the original half-order (Half1 followed by Half2) and (2) using a swapped version in which early and late trends are interchanged (Half2 followed by Half1).

The difference between these 2 predictions,

Δ=yobserved−yswapped(2)

represents the individual’s timing sensitivity—how much the moment of behavioral change affects the expected outcome.

Statistical strength and direction of this effect were then summarized across waves for the same individual, using metrics such as the mean Δ, SD, Wilcoxon signed-rank test, Cohen d with bootstrap CIs (95% CI), and the probability of superiority (proportion of waves with Δ>0). These metrics quantified individual sensitivity of the outcome to the order of 24-hour movement behavior trends and changes, partially answering RQ2.

To illustrate how PEPHA operates at the individual level, we present 2 real participants in this proof-of-concept application, each analyzed for both outcomes: processing speed and depressive symptoms. These examples act as concise “use cases” showing how the framework uncovers personalized temporal-order dependencies between behavioral dynamics and outcomes.

Participants were selected based on the literature on cognitive aging and affective association, which consistently identify sex and age as the main drivers of change (women tend to have overall worse brain health at younger ages, but men’s decline is steeper with age [39-41]). Accordingly, we identified:

To estimate when, within a wave, the largest behavioral shifts occur, PEPHA finally performs change point detection on the original daily sequences across waves of the same individual. Using the ruptures/pruned exact linear time (PELT) algorithm, change points are detected for each daily variable. It was selected because it provides an exact solution for multiple change point detection while maintaining near-linear computational complexity, allowing efficient analysis of daily behavioral time series where several structural shifts may occur within a wave. Such points are obtained via 2 cost models: (1) ℓ₂ (mean-shift), selecting the boundary with the largest Cohen d between segments, and (2) RBF (distributional-shift), detecting broader shape changes.

Each detected change point is characterized by its position (in days before the outcome assessment), magnitude of change, and data completeness. Aggregating these across waves yields individualized maps of potential “potential temporal association windows”—periods when outcomes may be most sensitive to changes in behavioral patterns. They signal candidate potential temporal association windows that may help prioritize when favorable or unfavorable configurations of behavior are likely to matter most (eg, reductions in sedentary time ≈ 20 d before assessment of processing speed), completing RQ2’s answer by revealing when behavioral timing is potentially most relevant for a given outcome and a given individual.

To operationalize PEPHA’s individualized component and for a proof-of-concept implementation, we applied such change point analysis to the top features with the highest permutation importance for a given participant and outcome. This approach enabled the identification of behavioral inflection points within each 90-day pre-assessment window.

Predictors were then organized according to groups. Within each, the mean number of days before assessment and its 95% CI were computed to identify clusters of potential temporal sensitivity (windows in which behavioral changes are most predictive of outcome variability).

Finally, to translate the detected change points into actionable temporal guidance, we visually overlapped all waves of the same individual for the 3 predictors with the highest importance for a given outcome. By aligning the daily trajectories relative to the outcome assessment, we examined how the behavioral pattern differs before versus after each detected change point. This step allowed interpretation of how the temporal ordering of behavioral patterns unfolds across waves, thereby informing how a potential halves-swap of the wave should be structured to reproduce the behavioral patterns associated with improved predicted outcomes per individual.

PEPHA produces 2 outputs at its end. At the population level, LOSO cross-validated SMAE and feature-importance tables identify the behavioral domains most predictive of the outcome studied. At the individual level, reports combine half-swap metrics (Δ, Cohen d, p, probability of superiority) with change point computations and confidence intervals. An illustration of PEPHA’s step-by-step methods is provided in Figure 3.

Together, these layers translate complex longitudinal data into personalized, temporally specific insights applicable to aging populations. PEPHA may offer a reproducible, interpretable basis for data-driven, timing-aware intervention planning, guiding future research toward individualized strategies that promote healthy trajectories of aging.

For the proof-of-concept application of the framework, all analyses were performed in Python 3.11 using standard scientific libraries, including pandas (2.2.2), NumPy (1.26), SciPy (1.13), scikit-learn (1.5), ruptures (1.1) for change point detection, and matplotlib (3.9) and seaborn (0.13) for visualization. Computations were executed on an Apple M1 Pro processor with 16 GB RAM running macOS 26, although the full pipeline is hardware-agnostic and compatible with both CPU- and GPU-based environments.

The code implementing the full PEPHA framework (from data preprocessing to feature extraction, modeling, temporal order testing, and change point analysis) was written to ensure full reproducibility. All configuration parameters (eg, window size, penalty for change point detection) were defined in a central location of the script, allowing flexible adaptation to other datasets and research contexts.

Given the extensive Optuna hyperparameter search performed for each outcome domain (see the Methods section), in this section, we report only the results obtained with the parameter configuration yielding the lowest composite Optuna objective function (see equation 1). Table 4 summarizes the most optimal complexity parameter values for the 4 model types: processing speed levels, processing speed drift, depression levels, and depression drift.

The optimized parameters revealed systematic differences between the levels and drift variants of PEPHA. Levels models converged toward a minimum of 2 valid waves per participant, whereas drift models required at least 3 valid waves, consistent with the greater temporal stability needed to estimate changes in outcome trajectories. The temporal window size was predominantly 6 days, except for processing speed levels, which favored shorter 2-day windows, likely reflecting the higher short-term sensitivity of absolute processing speed performance to extreme fluctuations in 24-hour movement behavioral predictors. The inclusion of SE features did not generally improve model performance, except in the case of depression drift, where additional uncertainty information appeared to aid generalization.

Levels models tended to perform best with lower correlation thresholds for redundant feature removal (≈0.60), allowing moderately correlated predictors to remain, while drift models benefited from stricter pruning (≈0.75‐0.80). Similarly, levels models favored smaller support vector regression (SVR) regularization constants (C), promoting smoother fits, whereas drift models required larger C values to capture stronger nonlinear relations. No clear trends emerged for ε or γ.

As a result of the optimization by Optuna, the final objective formula value, its components, and the final N of participant data included in each finally built model are presented in Table 5.

Drift models consistently achieved lower objective scores (see equation 1) than levels models. However, and due to the lower value of minimum waves required for inclusion, levels models preserved slightly broader population coverage (more included participants). Across outcomes, MAEz values were close to 1, consistent with the expected difficulty of forecasting cognitive functioning and mental health subtle signals in ecologically collected data.

Given that the primary objective of this paper was to introduce the PEPHA framework rather than to optimize predictive performance per se, we do not emphasize the relatively high MAEz values, serving as a proof-of-concept step only. In the next section, we report the final model performance for each outcome using the best-found hyperparameters and a LOSO CV protocol.

As presented in Table 6, across all models, prediction errors remained within 8% and 20% of the range of values of the assessment tool (SMAE). The depression-related models achieved the lowest error rates, indicating that affective outcomes may exhibit more consistent relationships with wearable-derived 24-hour movement behavioral dynamics than cognitive performance does.

Performance patterns differed across outcomes and between the levels and drift variants. For processing speed, the drift model outperformed the levels model, yielding lower errors and higher correlation with observed values. Its σpred/σtrue ratio (0.98) indicated realistic variance reproduction, suggesting that gradual changes in behavioral habits across consecutive 90-day periods more effectively captured the dynamics underlying cognitive change than short-term fluctuations did.

For depressive symptoms, the opposite pattern emerged: The levels model achieved the lowest errors and the strongest correlation, indicating that short-term behavioral dynamics are more predictive of concurrent depressive levels. However, this model underestimated interindividual variability (σpred/σtrue=0.47), whereas the drift variant reproduced outcome dispersion almost perfectly (≈1.0) despite slightly higher errors.

In line with the study’s aims and RQs, the following sections focus exclusively on the levels models, as they directly address how changes in 24-hour movement behavior during the 90 days preceding an assessment relate to the observed outcome levels. The drift models, which capture habit changes across waves rather than immediate behavioral dynamics, will be explored in future work.

Across both the processing speed and depressive symptoms outcomes, the 10 features whose permutation led to the largest SMAE increase are reported in Tables 7 and 8 and represent the most potentially influential predictors within their respective models.

When analyzing only the top 10 features for processing speed, most of them reflected the predictors in the intensity-specific PA duration group (5/10, 50%), were represented by kurtosis as the feature (5/10, 50%), and had linear trends as the predominant trend (60%). Interestingly, sleep regularity index was the top predictor.

Regarding depressive symptoms’ top 10 features, 24-hour heart rate dynamics was the group for the majority (6/10, 60%), with skewness the most represented feature (5/10, 50%) and linear and quadratic trends equally predominant. Control-related variables such as wearing time kurtosis also contributed modestly.

However, although the results presented in this section highlight only the 10 strongest predictors for each outcome, PEPHA’s design enables a broader examination of which behavioral and control dimensions systematically contributed to model accuracy across all features. The following section therefore expands this analysis to the full set of predictors, quantifying their overall importance and group-level representation within the population models.

For processing speed, 144 features (29.63% of 486 total, based on the initial 8 features per 22 predictors sampled daily on each of the 2 halves and using linear and quadratic fits, added to the 7 predictors sampled once, and after removing the correlation-derived duplicates) yielded positive permutation values, indicating that their exclusion increased model error. Control variables such as education, total CRIq score, MDS, sex, subjective age difference, and chronological age did not appear among these positive contributors. Conversely, sleep regularity index, the proportion of missing windows, the number of tries on the Trail Making Test, and the wearing time were beneficial for modeling. For the other groups of predictors, all their variables contributed positively to prediction.

When grouped by the 24-hour movement behavioral domain, the proportional representation of informative predictors indicates locomotor activity, 24-hour heart rate dynamics, and intensity-specific PA duration as the most impactful groups for the modeling of processing speed across the population studied, as shown in Table 9.

Predictor-wise and considering the entire set of cases for which permutation led to a model performance drop (positive mean MAE increase), sleep regularity index (with a sum SMAE impact of 0.446% and 7% of the informative list features), heart rate skewness (0.327% and 9.2%, respectively), seconds spent in heart rate zone 3 (0.301% and 6.3%, respectively), and step frequency (0.297% and 7%, respectively) emerged as the most informative predictors. Analysis of the temporal features revealed a nearly balanced representation across halves (Half1: 50%; Half2: 48.6%; note that controls are not considered in this metric), with the second half achieving a higher sum of SMAE drop when permuted (1.886%) compared with the first half (1.763%). Quadratic trends contributed slightly more to the modeling (58.5% vs 40.1% for linear), with quadratic also obtaining a bigger model contribution (1.843% SMAE) compared with linear (1.763% SMAE). All parameters originated from slope-based representations, as SEs were not included in the best-performing processing speed configuration.

For depressive symptoms, 207 features (39.35% of 526) had positive permutation values. Like for processing speed, control variables such education, total CRIq, MDS, sex, and subjective age difference did not have predictive value. However, chronological age was important for the modeling of depressive symptoms. Additionally, the number of HADS tries did not appear among positive contributors, in contrast to Trail Making Tries tries for processing speed.

Table 10 contains the proportional representation and sum of SMAEs for each predictor group when modeling depressive symptoms.

Across the list of informative predictors, the predictors with the highest impact on the model’s performance were heart rate minimum (0.332% of the difference in SMAE and 6.3% of the informative features list), heart rate skewness (0.227% and 3.9%, respectively), and seconds spent in heart rate zone 2 (0.205% and 6.3%, respectively). A similar proportion of informative features originated from each half of the wave (Half1: 51.2%; Half2: 47.8%, with sums of SMAEs being 1.107% and 1.219%, respectively), suggesting that late-period dynamics of activity may carry stronger affective relevance. Like processing speed, the depression model did not benefit from the inclusion of SE parameters. Linear trends were mostly as represented (48.8%, sum MAE difference 1.087%) as the quadratic ones (50.2% and 1.239%, respectively).

This section focused on answering RQ2—whether the timing of behavioral change within each 90-day wave period could be linked to the processing speed and depressive symptoms at its end. It does so by searching for the potential temporal association windows per participant and outcome case, thus providing an individualized result and insight.

In the individualized analysis of processing speed for participant MO, the behavioral dimensions highlighted differed from the population-level pattern. Among the top 10 features (Table 13), 24-hour heart rate dynamics accounted for one-half of the features, instead of intensity-specific PA duration, which was the most relevant group at the population level. Although population results did not reveal a predominant half in the top 10 features, for the participant MO’s case Half 2 was represented in 70 % of the top features, displaying a higher relevance of the second 45-day period in the outcome.

Recalling the positive Cohen d observed earlier for this individual, swapping the 2 halves of the behavioral window would lead to a decrease in the processing speed score of this individual, corresponding to improved performance (since lower values reflect faster completion times).

To examine when such adjustments might be most effective, we overlapped the 95% CIs of all the predictors’ change points detected according to our methods. The per-predictor plots presented in Figures 4 and 5 show the distribution of individual change points, with the bars indicating the 95% CI for each predictor for which a change point was detected and red markers representing the mean. Cohen d–based change points identify the time within the observation window at which the magnitude of behavioral change between earlier and later segments is largest, independent of outcome relevance. In contrast, RBF-derived change points reflect moments when behavioral shifts most strongly influence the model’s predicted outcome. Thus, the former captures maximal behavioral differences, whereas the latter captures maximal outcome-relevant differences.

The aggregated distribution (Figures 6 and 7) revealed a mean change point at 40.5 days before the assessment (median 38.0, 95% CI 26.4 to 62.1) based on Cohen d and 45.9 days before the assessment (median 48.0, 95% CI 30. to 54.5) based on the RBF-derived importance scores.

When analyzing aggregate overlaps of detected change points for only the most informative group of predictors (24-hour heart rate dynamics), the results (Figure 8) obtained with Cohen d revealed a mean change point at 45.1 days before assessment (median 42.5, 95% CI 29.8 to 62.6) and with RBF at 48.3 days before assessment (median 50.0, 95% CI 35.8 to 59.5).

However, after knowing that MO processing speed would benefit from a halves-swap of their behavior data and knowing the change point across waves of the top 3 predictors of their processing speed, one final question remains: how should that change happen? For example, should the IQR of the minimum 24-hour heart rate increase or decrease after the change point of 63.3 days before the outcome assessment?

To answer that final step, Figure 9 presents the overlap of the 5 waves of MO’s data for the top 3 features as well as their evolution over time.

For the strongest predictor, the IQR of daily minimum heart rate, the wave-overlap plots (Figure 9A) indicate that, before this change point (earlier in the wave, further from the assessment), the values were generally lower and more stable across waves. After the change point (closer to the assessment), the IQR increased and became more variable. This pattern suggests a transition from relatively stable nocturnal/resting cardiovascular dynamics earlier in the wave toward greater dispersion closer to the assessment. Importantly, the prior order-swap result indicated that reversing the temporal order of the 2 halves improved MO’s predicted processing speed. This suggests that performance would benefit if the more stable, low-IQR pattern occurred closer to the assessment (26 d to 0 d prior) while the more variable pattern occurred earlier in the wave (90 d to 26 d prior).

For the second strongest predictor, the minimum of daily seconds spent in intense activity (Figure 9B), the results indicated that, earlier in the wave, the minimum values remained low overall but with fewer and more sustained increases visible as extended plateaus rather than brief excursions. Later, they were generally low but punctuated by frequent, short spikes, indicating intermittent engagement in higher-intensity activity. This pattern reflects a shift from a more sustained (but temporally constrained) intense activity presence earlier to an irregular, burst-like exposure later in the wave toward. The order-swap results suggested that the relevant adjustment was, again, not a simple increase or decrease in intense activity but rather a reorganization of its timing. Specifically, positioning the earlier-wave pattern of more sustained minimum intensity closer to the assessment (60 d to 0 d before) and shifting the later-wave pattern of brief, irregular spikes to an earlier period (90 d to 60 d) would be associated with improved processing speed performance. For the third strongest predictor, the kurtosis of the daily median heart rate (Figure 9C), 2 change points were detected: one at 41.7 days before the assessment using Cohen d and another at 60.0 days before the assessment using the RBF-derived importance scores.

The wave-overlap plots show that, relative to the Cohen d change point, earlier in the wave, the kurtosis values were generally lower and less variable, indicating heart rate distributions closer to a more regular, light-tailed shape. Later in the wave (closer to the assessment), kurtosis increased and exhibited greater variability, reflecting more peaked and heavy-tailed distributions with occasional extreme values. In contrast, the RBF-derived change point at 60 days primarily separated periods by the maximum attained kurtosis, with higher extreme values occurring predominantly closer to the assessment. Together, these results indicate a transition from more regular, stable heart rate distribution shapes earlier in the wave toward increasingly irregular and extreme dynamics later. The order-swap results suggest that shifting the lower kurtosis, less variable regime to the later portion of the wave (49 d to 0 d before the assessment) and relocating the higher, more variable kurtosis patterns to the earlier portion (30 d to 60 d before) would be associated with improved processing-speed performance.

Together, these complementary analyses indicate that MO’s processing speed performance was most sensitive not to absolute levels of PA or heart rate metrics but to their temporal ordering within each wave. Across the 3 strongest predictors (variability in minimum heart rate, minimum engagement in intense activity, and the distributional shape of median heart rate), a consistent pattern emerged: Configurations characterized by greater stability, regularity, or sustained structure were beneficial when they occurred closer to the assessment, whereas more variable, burst-like, or extreme patterns were better positioned earlier in the wave. The change point analyses identified when these configurations naturally shifted, while the order-swap result clarified the direction of benefit, indicating that reversing the observed before/after sequence would potentially improve processing speed. Taken together, these findings suggest that, for MO, the pattern associated with improved predicted processing speed was characterized by a temporal reorganization of existing 24-hour movement behavioral patterns, in which stable cardiovascular dynamics and more sustained activity engagement were positioned closer to the assessment, rather than by uniform increases or decreases in activity or heart rate levels.

This paper introduced and empirically validated PEPHA, an interpretable AI framework that provides a way to identify individual potential temporal association windows and the behavioral dimensions that contribute to health-relevant outcomes. PEPHA is purposefully hypothesis-generating rather than causal: It surfaces who, which behavioral dimensions, and during which periods outcomes appear more sensitive to behavioral variation. Importantly, these windows should not be interpreted as indicating that a person should begin a behavior at an exact number of days before an assessment but rather as identifying periods of increased association or responsiveness within which favorable or unfavorable behavioral patterns may matter more.

At the population level and using data from the Providemus alz study, PEPHA highlighted distinct temporal and behavioral conclusions for processing speed and depressive symptoms. Processing speed was characterized by a broad contribution of locomotor activity, 24-hour heart rate dynamics, and intensity-specific PA duration, with nonlinear temporal patterns playing a prominent role. This suggests that cognitive functioning in aging is potentially associated not only with overall levels of 24-hour movement behavior but also with how behavioral and physiological signals fluctuate and evolve within an observation period prior to assessment. In contrast, depressive symptoms were primarily associated with sedentary behavior alongside intensity-specific PA duration and heart rate dynamics, with greater relevance of 24-hour movement behavioral patterns occurring later within the observation window (second half of the 90 d prior to the assessment). Together, these findings underscore that cognitive functioning and mental health in aging are potentially linked to similar configurations of behavioral dynamics but in different relevance order, emphasizing the importance of temporal distribution and patterning of 24-hour movement behaviors rather than aggregate activity alone. However, the empirical findings presented here primarily serve to illustrate the framework’s analytical capabilities rather than to establish novel relationships between PA and cognition or mental health or to provide clear intervention plans to be implemented.

Beyond identifying which behavioral dimensions matter on a population level, PEPHA also served to uncover when behavioral change appears most consequential individually. Although not finding a clear trend in the population-aggregated analysis, the framework isolated participant-specific “potential temporal association windows” (ie, periods where altering the sequence or stability of 24-hour movement behaviors most influenced predicted outcomes). To demonstrate this individualized component, 2 real participants (female<55 years old; male, >60 years old) were analyzed as illustrative cases. Despite comparable prediction accuracy, only the male participant exhibited a consistent, meaningful effect of temporal order, indicating that behavioral sequencing played a role in the cognitive outcome (estimated 8% decrease in predicted processing speed scores when the halves are swapped, suggesting that a change in behavior order was potentially associated with an increase of cognitive performance). For this individual, PEPHA localized the main window of sensitivity to change approximately 60 days before the cognitive assessment for 2 physiological signals (heart rate minimum and median) and 30 days for intense activity duration, suggesting that this individual’s processing speed may have been particularly sensitive to favorable or unfavorable behavioral configurations during those periods. Rather than indicating that a specific behavior should be initiated exactly 30 days or 60 days before assessment, these findings point to windows of increased potential temporal association in which the sequencing and stability of 24-hour movement behavior may be more relevant to the outcome. Future population-wide analyses will extend this approach across individuals, enabling systematic characterization of how such individualized “potential temporal association windows” vary as a function of demographic, cognitive reserve, and clinical profiles.

From a practical perspective, PEPHA is intended to support personalized interpretation rather than automated clinical decision-making. For example, a clinician could use the framework to identify which behavioral dimensions appear most relevant for an individual and whether these associations are concentrated within specific periods preceding an assessment. These outputs could support personalized monitoring, support personalized behavior change trials, guide conversations regarding lifestyle patterns, and generate individualized hypotheses for future evaluation. Importantly, they should not be interpreted as behavioral prescriptions or recommendations regarding intervention timing but rather as interpretable summaries of personalized temporal associations requiring prospective validation.

Although the absolute prediction errors can still be considered large (an expected property of ecological data collected in free-living settings), the proposed framework achieved its primary goal on this first proof-of-concept application: to derive individualized, interpretable analyses of 24-hour movement behaviors capable of generating personalized hypotheses regarding the temporal relevance of behavioral patterns for healthy aging. This implementation was not compared against nontemporal behavioral representations, as its objective was not to demonstrate superiority over existing predictive models but to establish the feasibility and interpretability of a temporal modeling framework. Future work should directly quantify the incremental value of incorporating temporal information through comparisons with nontemporal baselines and more complex temporal architectures.

Nevertheless, this first empirical application of PEPHA should be interpreted considering several methodological and data-related limitations. First, the proof of concept relied on the Providemus alz dataset, which, despite its unique longitudinal high-frequency nature, comprised a relatively small and demographically homogeneous sample. Consequently, population-level estimates may underrepresent interindividual diversity in PA patterns; cognitive baselines; and contextual factors such as work status, comorbidities, or seasonal variability. Extending PEPHA to larger and more heterogeneous cohorts will be essential to establish the generalizability of its individualized findings.

Second, the number of repeated outcome assessments per participant was limited (3 to 5), which constrains the precision of temporal order and change point estimates. The observed wide CIs for individual-level metrics such as Cohen d or the probability of superiority reflect this inherent uncertainty. Consequently, the individualized behavioral patterns and temporal association windows presented in this work should be interpreted as illustrative applications of the PEPHA framework rather than definitive evidence of stable person-specific temporal effects. Their robustness and generalizability should be evaluated in future studies with substantially longer longitudinal follow-up and a greater number of repeated assessments per participant. Such datasets will enable more reliable estimation of individualized temporal patterns as well as formal evaluation of their reproducibility under different modeling assumptions.

Third, this first implementation of PEPHA deliberately prioritized interpretability over maximal temporal representational power. By summarizing behavioral trajectories using linear and quadratic temporal trends, the framework produced compact and clinically interpretable descriptors of behavioral evolution that can be directly related to individual outcomes. To minimize the inevitable trade-off between interpretability and information preservation, the framework summarized each temporal window using a set of descriptive statistical metrics (eg, central tendency, variability, distributional shape) that have been widely adopted in previous wearable-sensing and digital phenotyping research to characterize behavioral and physiological time series. Nevertheless, this design inevitably sacrifices part of the temporal richness present in continuous wearable data. In particular, higher-order temporal characteristics such as circadian organization, burst-like behavioral episodes, recurrent patterns, and complex nonlinear transitions may not be fully preserved by the current feature representation. We intentionally accepted this trade-off in this first implementation because the primary objective was to demonstrate an interpretable methodological framework capable of generating individualized temporal hypotheses rather than maximizing predictive performance. Future work should investigate richer temporal representations, including sequence-based architectures such as long short-term models (LSTMs) or transformer models combined with explainability techniques (eg, temporal SHAP values or attention-based interpretation), to determine whether additional temporal information can be exploited while maintaining clinically meaningful interpretability. Such approaches, however, require substantially larger longitudinal datasets than those available for this initial proof-of-concept implementation and would, therefore, fall out of scope of this initial research.

Finally, this first proof of concept tested the framework on cognitive functioning and mental health outcomes derived from ecological self-assessments and mobile digital games. These measures, though validated, remain subject to within-person noise and daily fluctuations. Integrating additional objective or clinical endpoints (eg, biomarkers or mobility indicators) could strengthen external validity and clarify the mechanistic pathways linking behavior and brain health.

Overall, these limitations do not detract from the framework’s conceptual contribution. Instead, they outline a clear research agenda: scaling PEPHA to multisite datasets, integrating multimodal signals, extending it to other health outcomes, extending it to randomized control trial implementation for causality analysis of its results, and closing the loop between individualized behavioral prediction and adaptive, real-world intervention design.

In the context of healthy aging, the PEPHA approach supports a paradigm in which AI no longer prescribes uniform recommendations but instead derives personalized “potential temporal association windows.” These can inform hypotheses about intervention timing, periods of increased sensitivity, and the tailoring of PA guidance to each person’s physiological and lifestyle rhythms. Such personalization directly responds to the need of using AI for tailoring PA interventions based on multidimensional, real-world data.

The framework’s modular structure also ensures its transferability beyond the current use case. Its parameterization, such as window length, feature selection thresholds, and outcome types, can be adapted to diverse datasets, from clinical PA trials to community-based monitoring studies. The framework’s interpretability further supports clinical translation, allowing practitioners to visualize how behavioral dynamics influence health trajectories and to communicate AI-derived recommendations transparently to older adults.

In summary, PEPHA lays the groundwork for a new generation of explainable and actionable AI tools that bridge observational sensing and behavioral intervention. By enabling individualized insight into both the mechanisms and timing of behavior-health relationships, it aligns with the broader goal of transforming digital health data into personalized, adaptive strategies that promote functional capacity, autonomy, and quality of life in aging populations.

In this article, we presented and applied a novel framework, with which we demonstrated that naturally occurring 24-hour movement behavior can reveal both which behavioral dimensions and when changes most potentially influence cognitive functioning and mental health in adults.

In addition to population-derived conclusions, PEPHA allowed us to study the individual level, detecting unique and personalized “potential temporal association windows” and showing that behavioral order and timing can modulate cognitive and mental health outcomes, particularly in specific aging profiles. The empirical findings should be interpreted as illustrative applications of the framework rather than definitive evidence of novel behavioral associations.

Although preliminary, these findings illustrate how interpretable AI can move beyond prediction to prescription, transforming high-frequency observational data into actionable, personalized hypotheses for optimizing PA in everyday life. Future work must expand PEPHA to larger, more heterogeneous samples and evaluate its integration into adaptive intervention platforms, ultimately contributing to the development of explainable, data-driven, and timing-sensitive PA personalization for healthy aging.