Introduction:
The National Institutes of Health Stroke Scale (NIHSS) scores are often not readily available in structured claims databases. We have previously demonstrated that a machine learning model can be used to determine proxies for NIHSS scores. Our current work focuses on creating a model applicable across different databases to validate our approach and enable further outcome studies.
Methods:
We identified 1,415 eligible hospital-admitted patients in the Optum® de-identified Integrated Claims-EMR database who were diagnosed with ischemic or hemorrhagic stroke, or a transient ischemic attack and had NIHSS scores in medical notes. These patients were split into a training (N=1,192) set for model development and a hold-out test (N=223) set to evaluate model performance. Furthermore, model performance was externally validated using the 286 eligible stroke patients in IBM’s Claims-EMR database (CED). Potential predictors for stroke severity included relevant procedures, diagnoses, patient demographics, and information about the patient hospital stay.
Results:
The optimal model, a random forest model, achieved a coefficient of determination (R
2
) between the actual and predicted NIHSS scores in the hold-out Optum dataset of 0.48 and of 0.42 in the secondary CED dataset. The final model incorporated a total of 47 predictors. The strongest predictors included transient ischemic attack diagnosis, length of hospital stay, critical care procedures, patient age, and hemiplegia diagnosis.
Conclusion:
This study shows that machine learning can be used to determine proxies for NIHSS scores across different real-world databases. Ultimately, this will enable large claims-based outcome studies involving stroke severity to improve our understanding of how stroke severity affects healthcare utilization, total cost of care, and the financial impact on the larger community.
Length of hospital stay is shorter in South Asian patients with acute pulmonary embolism