Kolmogorov–Arnold Networks for Cross-Domain Time-Series Modeling in Health and Activity Monitoring

Hamza Haruna Mohammed, Gabriel Kiss, Frank Lindseth

Keywords: Kolmogorov–Arnold Networks (KAN) Interpretable Deep Learning Time Series Modeling Cross-Domain Transfer Learning Wearable Sensor Data Circadian Rhythms Digital Health Attention Mechanisms

Abstract: Wearable and clinical time-series provide complementary views of human health but differ in sampling, noise, and labels, hindering cross-domain modeling. We present KAN-Health, a Kolmogorov–Arnold Network–based framework that harmonizes heterogeneous sources into a small set of daily metrics and applies spline-based univariate transforms with additive mixing for intrinsic interpretability. We pretrain on a large wearable dataset (PMData) and freeze spline layers while fine-tuning only the mixing/attention components on a clinical ADHD dataset (Hyperaktiv), preserving transparent feature mappings during transfer. Across leave-one-subject-out evaluation, KAN-Health improves F1 and MCC over Random Forest, Logistic Regression, Gradient Boosting, and a Transformer baseline on Hyperaktiv, and yields higher MCC in both transfer directions. Visualizations of the learned splines align with clinical expectations (e.g., circadian regularity and sleep efficiency). KAN-Health demonstrates that interpretable KANs can match or exceed black-box baselines while enabling cross-domain adaptation with fewer trainable parameters.

Serve As Reviewer: ~Hamza_Gbada1, ~Fabien_Danieau1

Submission Number: 66