Prospective and external validation of stroke discharge planning machine learning models.

Area of Science:

• Neurology
• Medical Informatics
• Data Science

Background:

• Stroke patient discharge planning is complex.
• Accurate prediction of patient outcomes is crucial for effective discharge.
• Machine learning models offer potential for improving discharge predictions.

Purpose of the Study:

• To validate previously developed machine learning models for stroke patient discharge planning.
• To assess model performance on external and prospective datasets.
• To predict modified Rankin Scale (mRS) at discharge, discharge destination, survival to discharge, and length of stay.

Main Methods:

• Utilized datasets from Royal Adelaide Hospital and Lyell McEwin Hospital.
• Applied pre-derived machine learning models to prospective and external patient data.
• Evaluated model performance using area under the receiver operator curve (AUC), sensitivity, and specificity.

Main Results:

• Models demonstrated strong predictive performance for discharge mRS ≤ 2 (AUC 0.85-0.87), home discharge (AUC 0.76-0.78), and survival to discharge (AUC 0.91-0.92).
• Prediction of length of stay using admission data remained challenging (AUC 0.62-0.66).
• Successful prospective and external validation of six-variable machine learning models was achieved.

Conclusions:

• Machine learning models show promise for predicting key discharge planning factors in stroke patients.
• Validated models can provide valuable insights for discharge decisions.
• Further research is needed to confirm patient and system-level benefits of implementing these predictive models.