A prognostic model built from inpatient electronic medical records across all acute-care facilities in Calgary, Alberta analyzed 15,160 heart failure admissions (2011–2019) to predict readmission at 30, 90, and 365 days. Variables selected via a modified Delphi expert-consensus process were extracted from clinical notes using natural language processing (NLP) alongside structured EMR elements. Competing risk survival models treated death as a rival outcome, and performance was assessed via C-statistics, sensitivity, specificity, and predictive values.
Heart failure carries a 30-day readmission rate hovering near 20–25% in most health systems, generating enormous cost and patient burden. What distinguishes this approach is the deliberate mining of free-text clinical notes — historically an underutilized data layer — rather than relying solely on billing codes or structured fields that often lag clinical reality. Coupling NLP with expert-curated variable selection addresses a persistent weakness in prior readmission tools like LACE or GWTG-HF, which plateau around C-statistics of 0.60–0.65. Whether this model meaningfully surpasses that benchmark remains unclear from the abstract alone.
Key limitations include single-region generalizability (one Canadian health system), a retrospective design that cannot confirm causal pathways, and the absence of social determinants of health. As a preprint posted on medRxiv and not yet peer-reviewed, the reported performance metrics and clinical conclusions should be treated as preliminary until independent validation and formal peer review are complete. Potentially incremental rather than paradigm-shifting, but methodologically noteworthy.