Bursting the bubble: data leakage and inflated deep learning accuracy in multivariate time-series frailty classification

Background Accurate prediction of frailty in older adults is crucial for preventing adverse outcomes, yet distinguishing frail, pre-frail, and non-frail states remains challenging. A recent study applied InceptionTime to the GSTRIDE dataset and reported near-perfect multi-class frailty prediction (>98% accuracy), exceeding values typically observed in comparable studies. Methods We conducted a methodological re-evaluation and replication of this pipeline to assess the robustness of reported performance. Corrections included subject-wise data partitioning, feature scaling within training folds, and non-overlapping sliding time windows applied separately to each subset to prevent potential leakage. Results Reimplementation of the original pipeline reproduced the previously reported high accuracy. After applying the corrected framework, overall recall and precision decreased (50.9 and 52.3%, respectively), providing a more conservative, data-specific estimate of model generalizability. Per-class analysis indicated reductions across all categories, with Frail-class recall dropping to 21.4%, highlighting the particular challenge of identifying high-risk individuals. Conclusion The findings suggest that methodological factors, such as data leakage, likely contributed to the previously reported high performance. Under rigorous controls, frailty prediction is challenging, particularly for the frail class, underscoring the need for careful evaluation of model generalizability. Significance This study illustrates the importance of transparent, methodologically sound pipelines in clinical AI research. By providing a reproducible framework for frailty prediction, we aim to support future studies in obtaining realistic performance estimates and developing clinically meaningful models.