Early Prediction of Adverse Stroke Outcomes Using Nonclinical Factors and Missing Data: A Machine Learning Study.

Area of Science:

• Neurology
• Medical Informatics
• Machine Learning

Background:

• Early prediction of stroke outcomes is crucial for clinical decision-making and resource allocation.
• Traditional prognostic tools often lack complete clinical information.
• Machine learning (ML) offers a potential solution by utilizing available data, including non-clinical factors and data missingness.

Purpose of the Study:

• To evaluate ML models for predicting adverse stroke outcomes at 90 days post-admission.
• To assess the utility of non-clinical data and missingness patterns alongside traditional predictors.
• To compare the performance of Gradient Boosted Machine (GBM) models.

Main Methods:

• Utilized routine hospital data from UK clinical sites (NHS SafeHaven) for model training.
• Developed three GBM models incorporating clinical, non-clinical, and missingness features.
• Validated models using 10% of the data, evaluating accuracy, Area Under the ROC Curve (AUC), and Brier score.
• Employed SHapley Additive exPlanations (SHAP) for feature importance analysis.

Main Results:

• The study included 3530 stroke patients; clinical data exhibited significant missingness (>63% for five features).
• Models incorporating non-clinical and missingness features achieved 71% accuracy and an AUC of 0.76.
• Non-clinical factors (e.g., time to assessment/admission) and missing data (e.g., pulse, LDL) were significant predictors of adverse outcomes.

Conclusions:

• Non-clinical factors and data missingness are valuable for early prediction of 90-day adverse stroke outcomes.
• These factors, often well-documented in electronic health systems, can complement traditional clinical predictors.
• ML models leveraging these data sources can enhance stroke outcome prognostication.