Tri-SfSVD: New AI framework for clustering messy longitudinal health data
Uncovers hidden patient subgroups from sparse, irregular temporal data without imputation.
A new paper on arXiv introduces Tri-SfSVD (Sparse functional Singular Value Decomposition for triclustering), a unified framework that discovers biclusters and triclusters in longitudinal data. Unlike prior methods that require ad hoc imputation or assume shape-homogeneous clusters, Tri-SfSVD integrates continuous trajectory estimation with simultaneous selection of subjects, features, and temporal subregions. This is achieved by imposing sparse penalties across all three dimensions directly on observed data, making it robust to high-dimensional, sparsely sampled, and irregularly timed measurements.
In simulations, Tri-SfSVD outperformed existing functional biclustering approaches in high-dimensional settings. Applied to Inflammatory Bowel Disease (IBD) multi-omics data, it identified three biclusters linking patient subgroups with distinct clinical characteristics to specific microbial pathway groups. On multi-channel EEG data from an alcohol study, it found three triclusters connecting sample clusters to localized brain activity patterns and temporal subregions. The method promises interpretable subgroup discovery in complex diseases and neuroscience.
- Tri-SfSVD simultaneously clusters subjects, features, and time regions without imputation or shape constraints.
- Outperformed existing methods in high-dimensional simulations with sparse, irregular longitudinal data.
- Identified 3 biclusters in IBD data linking patient subgroups to microbial pathways, and 3 triclusters in EEG data linking alcohol phenotypes to brain activity patterns.
Why It Matters
Enables interpretable subgroup discovery from messy longitudinal health data, improving disease subtyping and neuroscience research.