Answers are grounded in this paper and cite their sources.
Background : Longitudinal predictive models can anticipate subsequent diagnoses, but connecting their predictions to clinical interpretation and management remains challenging. We developed PRISM, an agentic framework that combines diagnosis prediction with biomarker analysis, comorbidity assessment, intervention analysis, and knowledge extraction, using inflammatory bowel disease (IBD) as a use case. Methods : Using a cohort of 113,729 patients, we developed a decoder-only transformer that predicted diagnoses from structured longitudinal electronic health records. Four specialized agents guided analyses using feature masking, Shapley attribution, medication insertion, and statistical association screening. Evaluation included 318,145 prediction positions from 11,382 held-out patients, with analyses conducted at cohort and individual levels. Results : Exact-code Top-1, Top-3, Top-5, and Top-10 accuracy was 15.28%, 28.69%, 36.67%, and 49.04%, respectively. Top-10 accuracy reached 86.52% when any of the next ten diagnoses was accepted. Within outcome-specific baseline Top-1 subgroups, masking conventional CRP, high-sensitivity CRP, and fecal calprotectin reduced mean predicted IBD probability by 3.02, 3.48, and 1.37 percentage points. Masking IBD history reduced predictions of intestinal obstruction, nutritional deficiencies, and inflammatory arthropathies by 4.43, 3.67, and 4.98 points. Prednisone, metronidazole, and vedolizumab insertion reduced mean next-diagnosis nutritional-deficiency probability by 1.89, 1.89, and 1.22 points. Knowledge extraction identified a direct association between IBD and opioid-related disorders and an indirect relationship linking IBD, intestinal obstruction, and subsequent peritonitis or intra-abdominal abscess. Conclusions : PRISM integrated longitudinal prediction with clinically guided interpretation, intervention exploration, and cohort-scale association screening. Combining cohort summaries with individual analyses revealed differences obscured by averages and generated hypotheses for clinical investigation.
No takes yet. Share an insight, caveat, or question.
Shalmani et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: