Background: Thirty-day readmission following hospital discharge for stroke is an important quality measure for US hospitals. Current US prediction models for post stroke readmission based on electronic medical records from single healthcare systems or hospitals have modest discrimination (AUC range 0.64 - 0.74). Aim: To develop 30-day all-cause readmission prediction model using a machine learning (ML) based method trained on linked stroke registry and administrative claims data. Methods: Using probabilistic linking, we matched acute stroke (ICD-10 I61-I63) discharges from 31 hospitals participating in the Michigan Acute Stroke registry between 2016-2020 to multipayer administrative claims data provided by the Michigan Value Collaborative for Medicare and Blue Cross Blue Shield of Michigan commercial beneficiaries. Stroke registry data included patient demographics, clinical characteristics, past medical history, and treatments. Claims data was used to identify readmissions within 30 days of discharge. We used multivariable LASSO logistic regression- a simple ML technique to predict 30-day all-cause-readmission and evaluated the prediction accuracy using a hospital-split internal validation scheme to generate hospital-specific and pooled AUC estimates with 95% confidence intervals (Figure 1). Results: Of 19,382 linked stroke discharges, 2,724 (14.1%) were readmitted within 30-days. Readmitted patients were older, more likely to be male, black, and have higher stroke severity (NIHSS >5). Registry hospitals were either primary (64%) or comprehensive (26%) stroke centers. Hospital-specific 30-day readmission ranged between 9.9%-23.1% ( P <.001) with an average of 14.1% (95% CI:13.6%-14.5%) (Figure 2). Hospital specific AUC estimates ranged between 0.60-0.80 with a pooled AUC of 0.68 (95% CI:0.65-0.70) (Figure 2). Conclusions: ML prediction model fitted to linked registry data can be used to predict hospital-specific and statewide readmissions post stroke.
Read full abstract