A Claims-based Machine Learning Model to Classify Modified Rankin Scale at Stroke Discharge
Mamoon Habib1, Rafaella Cazé de Medeiros2, Syed Muhammad Ahsan2, Aidan McDonald Wojciechowski1, Joseph Newhouse3, M. Brandon Westover4, Lee Schwamm5, Lidia Maria Moura1
1Massachusetts General Hospital, 2Neurology, Massachusetts General Hospital, 3Harvard University, 4Neurology, Beth Israel Lahey Health Medical System, 5Yale New Haven Health System
Objective:
We validated a Medicare claims-based classifier to assess stroke severity, as measured by the Modified Rankin Scale (mRS) at discharge, among Medicare beneficiaries hospitalized for acute ischemic stroke (AIS).
Background:
The modified Rankin Scale (mRS) has been widely used to assess the severity of AIS in electronic health records (EHRs) and registries. However, national large-scale claims-based studies have been limited by the lack of valid measures of stroke severity.
Design/Methods:
We linked Paul Coverdell National Acute Stroke Program (PCNASP) to Medicare Claims data and sampled AIS patients aged 65+ admitted from 01/2018 to 12/2020 across nine states. PCNASP was primarily used to extract the outcome, discharge mRS, while the variables for the classifier were derived from Medicare claims data. Hospitals with an mRS completion rate below 90% were assigned to the training set, while those with 90%+ were assigned to the test set. A random 20% sample from hospitals with 90%+ completion was added to the training set for representativeness. Binary logistic regression with Lasso penalty classified discharge mRS, dichotomized as ≤2 for favorable and >2 for unfavorable outcomes. The classifier was evaluated on the test set using metrics like ROC AUC, Precision-Recall AUC and sensitivity.
Results:
The sample included 68,636 patients (mean [sd] age: 79.5 [8.7], female: 37,439 (54.5%)), with 27,986 (mean [sd] age: 79.8 [8.71], female: 15,394 (55.0%)) in the testing set. The predictors included 63 features, such as demographics, medical history, treatments, and discharge outcomes. The classifier achieved an ROC AUC of 0.85 (95%CI: 0.85 – 0.86), Sensitivity of 0.90 (95%CI: 0.90 – 0.91), Precision-Recall AUC of 0.90 (95%CI: 0.90 – 0.91).
Conclusions:
Among Medicare beneficiaries hospitalized for AIS, the claims-based classifier demonstrated excellent performance in ROC AUC, Precision-Recall AUC, and sensitivity for mRS classification. These results configure it as a tool for national surveillance of stroke outcomes.
Disclaimer: Abstracts were not reviewed by Neurology® and do not reflect the views of Neurology® editors or staff.