Spectral structure learning for clinical time series

• DOI: 10.48550/arxiv.2502.11680
• Publication Date: 2025-02-17
• Authors: Ivan Lerner; Anita Burgun; Francis Bach
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Machine Learning (cs.LG)

We develop and evaluate a structure learning algorithm for clinical time series. Clinical series are multivariate observed in multiple patients irregularly sampled, challenging existing algorithms. assume that our times realizations of StructGP, k-dimensional multi-output or multi-task stationary Gaussian process (GP), with independent sharing the same covariance function. StructGP encodes ordered conditional relations between series, represented directed acyclic graph. implement an adapted NOTEARS algorithm, which based on differentiable definition acyclicity, recovers graph by solving continuous optimization problems. Simulation results show up to mean degree 3 20 tasks, we reach median recall 0.93% [IQR, 0.86, 0.97] while keeping precision 0.71% [0.57-0.84], recovering edges. further regularization path is key identifying With proposed model dependencies, flexibly adapt different regularity, enabling us learn these dependencies from observations.