Multimorbidity measurement strategies for predicting hospital visits

Multimorbidity is a major challenge for healthcare systems. While multiple methods exist to measure it from electronic health records (EHRs), their relative performance for predicting hospital utilization is unclear. We conducted a retrospective cohort study using 15 years of EHR data (2007–2022) from a Portuguese hospital, including 925,508 patients and 9·7 M visits. Three phenotyping strategies were compared: a multi-source rule-based dictionary, Clinical Classification Software Refined (CCSR) mapping, and drug-based mapping. Five multimorbidity indices were evaluated: Charlson Comorbidity Index (ChCI), Elixhauser, Multimorbidity Weighted Index (MWI), RxRisk, and disease counts. Outcomes included emergency department (ED) visits, hospital admissions, unplanned admissions, and readmissions over 30–365 days, modeled with logistic regression (LR) and XGBoost. Best performance was achieved with XGBoost (AUROC 0.671–0.681 for ED visits, 0.663–0.668 for admissions, 0.773–0.781 for unplanned admissions, 0.798–0.848 for readmissions), with AUPRC patterns consistent with AUROC rankings. In LR models comparing indices, multimorbidity improved AUROC by 0.024–0.150 over demographics. MWI performed best for ED visits and unplanned admissions, while ChCI led for 30-day admissions (AUROC 0.613) and readmissions (AUROC 0.734). Combining multimorbidity with healthcare utilization features yielded the highest discrimination (AUROC 0.596–0.777), with readmissions most predictable. Multimorbidity ranked among the top predictors across outcomes. Multi-source phenotyping enhances chronic condition detection and prediction. Weighted indices offer modest benefits over disease counts, with optimal choice depending on the prediction task. Multimorbidity adds consistent value alongside utilization measures, supporting its routine use in risk stratification to improve population health management.