New Trans-GLMC method improves transfer learning with hospital suicide risk data

A new statistical method called Trans-GLMC tackles a key challenge in transfer learning: source heterogeneity. When multiple auxiliary datasets are available to help a target population with limited data, not all sources are equally useful—and their usefulness often falls into latent clusters. Existing methods treat sources as either informative or not, missing this structure. Researchers from UConn and other institutions tested Trans-GLMC on a real-world suicide risk study using the Connecticut Hospital Information Management Exchange (CHIME) dataset, which includes 636,758 patients across 27 hospitals. Because suicide attempts are rare per facility, individual hospital models are unstable, yet naive pooling ignores hospital-level differences in patient mix.

Trans-GLMC first computes coefficient-based distances between target and candidate sources to recover hidden clusters. It then combines global fusion, within-cluster refinement, and target debiasing to produce an estimator that adapts to the detected structure. The method establishes a non-asymptotic error bound that improves on unclustered transfer learning when a meaningful target cluster exists, and matches it otherwise. In simulations and the CHIME study, Trans-GLMC improved facility-specific prediction, identified interpretable hospital communities with mutual transferability, and recovered clinically coherent suicide-risk factors. The work was published on arXiv in June 2026.