New spectral embedding method improves rare disease EHR analysis

A team led by Feiqing Huang (Harvard) introduces a robust spectral embedding framework for rare disease cohorts in electronic health records (EHRs). Rare disease data suffer from high dimensionality and small sample sizes, making traditional unsupervised learning difficult. The new method leverages a knowledge matrix extracted from a broader population that shares partial subspace overlap with the target cohort. Unlike prior approaches, it does not require strict one-to-one signal alignment, allowing for more realistic structured sharing.

The procedure works in two steps: first identifying and removing irrelevant components from the knowledge matrix, then applying a projection-based method to recover shared and heterogeneous components separately. Simulations and real-world analysis on a multiple sclerosis cohort show significant improvements over competing methods, especially when shared signals are weak and only partially aligned. The work, published on arXiv (2606.11570), has implications for clinical decision support and patient stratification in rare diseases.