Finding hidden patterns in messy medical time-series data

Medical time-series data—like repeated microbiome or brain imaging measurements—are high-dimensional, irregularly sampled, and sparse. Tri-SfSVD combines functional data analysis with sparse penalties to find biclusters (subject-feature groups) and triclusters (subject-feature-time groups) without imputation or rigid assumptions. Applied to IBD microbiome data, it linked patient subtypes to distinct bacterial pathways; on EEG data, it uncovered alcohol-related phenotypes tied to localized brain activity patterns across time windows.