TNAM: Temporal neural additive models for multivariate clinical time-series

Multivariate clinical time series prediction provides critical support for life-saving medical decisions and healthcare resource optimization. However, existing models often produce unreliable predictions due to two fundamental challenges: (1) Variable heterogeneity, where distinct measurement frequencies cause dominant variables to mask signals from others; and (2) Cross-variable dependencies, where the failure to model complex physiological interactions results in incomplete representations of clinical trajectories. To address these challenges, we present a Temporal Neural Additive Model (TNAM). Inspired by the additive paradigm, TNAM first employs a set of variable-specific LSTM networks to isolate the unique dynamic patterns of each variable as independent additive components. To subsequently capture complex physiological interdependencies, we introduce a frequency-domain interaction module that models second-order spectral interactions. Comprehensive evaluations on five diverse real-world medical datasets demonstrate that TNAM consistently outperforms state-of-the-art methods across both mortality and early sepsis prediction tasks. In challenging cross-dataset evaluations, TNAM demonstrates superior generalization capability, highlighting its robustness to distribution shifts across different clinical environments. These results suggest that the proposed architecture can provide useful support for clinical decision-making, early warning, and personalized treatment planning. The code is publicly available at https://github.com/Hgnnhd/TNAM .