Stroke mortality prediction based on ensemble learning and the combination of structured and textual data

Area of Science:

• Medical informatics
• Machine learning in healthcare
• Cerebrovascular disease research

Background:

• Predicting short-term mortality in severe cerebrovascular diseases like stroke is medically significant.
• Existing models may not fully leverage diverse patient data for accurate mortality prediction.

Purpose of the Study:

• To develop and evaluate novel multi-level prediction models for six-month mortality in haemorrhagic and ischaemic stroke patients.
• To enhance prediction performance by integrating bioassay data with structured and textual radiology reports.

Main Methods:

• Combined multiple machine learning classifiers: Random Forest, AdaBoost, Extremely Randomised Trees, XGBoost, TabNet, and DistilBERT.
• Utilized a large dataset comprising 19,616 haemorrhagic and 50,178 ischaemic stroke patients.
• Incorporated bioassay data and radiology text reports into a multi-level prediction framework.

Main Results:

• Achieved high performance metrics for haemorrhagic stroke patients: AUROC = 0.89, AUPRC = 0.70, Precision = 0.52, Recall = 0.78, F1-score = 0.63.
• Achieved strong performance for ischaemic stroke patients: AUROC = 0.88, AUPRC = 0.54, Precision = 0.34, Recall = 0.80, F1-score = 0.48.
• Demonstrated enhanced prediction by combining laboratory, structured, and textual radiology data.

Conclusions:

• Developed novel, high-performing six-month mortality prediction models for stroke patients.
• These models can significantly aid in mortality risk assessment and early identification of high-risk individuals.
• Potential for more efficient healthcare resource utilization for stroke survivors through improved risk stratification.