Interpretable-MTLNet: A Kolmogorov–Arnold Network for Multitask Mental Health Prediction
Keywords: Machine Learning, Artificial Intelligence, Mental Health Prediction, Wearables, Multitask Learning, KAN
TL;DR: multitask neural architecture that jointly detects depression and anxiety from daily wearable time series while exposing mathematically transparent relations.
Abstract: Depression and anxiety are among the most prevalent mental health disorders worldwide, yet often underdiagnosed due to reliance on self-reporting and infrequent clinical assessments. Wearable devices offer a scalable path toward continuous mental health monitoring, but existing models often sacrifice interpretability for accuracy. We present Interpretable-MTLNet, a multitask neural architecture that jointly detects depression and anxiety from daily wearable time series while preserving mathematical transparency. The model couples multi-scale temporal convolutions with Kolmogorov–Arnold Network (KAN) layers to learn scale-aware embeddings and task-specific spline-based heads. This design provides consistent global and local explanations via activation-weighted univariate curves and symbolic surrogates. Evaluated on 40,000 participants with Fitbit data from the All of Us Research Program, Interpretable-MTLNet achieves a macro-AUROC of 0.731 across tasks, outperforming strong baselines under subject-level splits and remaining robust in imbalanced settings. These findings suggest that KAN-based architectures can deliver both accuracy and interpretability for detecting mental health problems from wearable data, advancing trustworthy digital phenotyping.
Submission Number: 41
Loading