When do learned representations improve longitudinal health prediction? A persistence-anchored, information-matched benchmark of wearable data and personalized models

Description This repository provides the data, source code, experimental results, and reproducibility materials accompanying the research article “When Do Learned Representations Improve Longitudinal Health Prediction? A Persistence-Anchored, Information-Matched Benchmark of Wearable Data and Personalized Models” by Anand Gavai and Jaap Heringa. The study investigates whether self-supervised temporal representation learning, particularly Joint-Embedding Predictive Architectures (JEPA), improves longitudinal wellbeing prediction compared with simpler statistical, persistence-based, and personalized models. Using the publicly available LifeSnaps dataset, comprising 71 participants with Fitbit-derived physiological and behavioural measurements and ecological momentary assessments, we evaluate the prediction of mood, stress, and energy at 1-, 3-, and 7-day horizons. The benchmark emphasizes four methodological principles: Persistence-anchored evaluation: Comparing predictive models against simple carry-forward and recent-history baselines. Information-matched comparisons: Ensuring that differences in prediction performance are not attributable to unequal access to historical observations. Participant-level validation: Using participant-disjoint cross-validation and subject-level bootstrap confidence intervals. Representation and personalization analysis: Evaluating JEPA-inspired representations, dimensionality-reduction controls, model capacity, label efficiency, and individualized prediction. The results show that, under the evaluated conditions, learned temporal representations provide no statistically discernible predictive advantage over strong, simpler models using comparable information. Recent self-reported wellbeing is a particularly informative predictor, while the additional value of wearable-derived representations and person-specific calibration is limited. Repository contents The reproducibility package includes: A derived daily LifeSnaps dataset with documented provenance. Python implementations of baseline models, JEPA-inspired representation learning, and evaluation procedures. Experimental configurations and numerical results supporting the manuscript and supplementary material. Statistical analysis, confidence intervals, and model-comparison procedures. Scripts for regenerating research figures and results. Automated tests addressing reproducibility, data leakage, and temporal consistency. Documentation covering research methods, limitations, and protocol deviations. Reproducibility and reuse The repository is intended to support independent verification, methodological comparison, and reuse in longitudinal machine learning research. The evaluation framework can be adapted to other datasets involving temporally correlated outcomes, sparse labels, and personalized prediction. The findings are specific to the LifeSnaps cohort, the evaluated wellbeing outcomes, and the tested model configurations. They do not establish clinical effectiveness, causal intervention effects, or the general superiority or inferiority of JEPA architectures. Original data source LifeSnaps: https://doi.org/10.5281/zenodo.7229547 The original LifeSnaps dataset is credited to its creators. This deposit provides derived research data and reproducibility materials rather than claiming authorship of the original cohort data.