Optimized Machine Learning Models to Predict In-Hospital Mortality for Patients with ST-Segment Elevation Myocardial...

Area of Science:

• Cardiology
• Medical Informatics
• Machine Learning

Background:

• ST-segment elevation acute myocardial infarction (STEMI) poses a significant risk of in-hospital mortality.
• Accurate prediction of mortality is crucial for timely intervention and improved patient outcomes.
• Machine learning (ML) offers potential for developing robust predictive models.

Purpose of the Study:

• To optimize ML models for predicting in-hospital mortality in STEMI patients.
• To evaluate the impact of random under-sampling (RUS) on model performance.
• To compare prediction accuracy using full hospitalization data versus simplified pre-reperfusion data.

Main Methods:

• 5708 STEMI patients were analyzed, with data split into training (75%) and testing (25%) sets.
• Four ML models were trained using RUS, with performance assessed by accuracy, sensitivity, specificity, G-mean, and AUC.
• Models were trained and compared using a full dataset and a simplified dataset of variables available before reperfusion therapy.

Main Results:

• RUS significantly improved model performance (G-mean) compared to models without RUS, across both full and simplified datasets.
• Support Vector Machine (SVM) models demonstrated superior performance, achieving high accuracy, sensitivity, specificity, G-mean, and AUC.
• SVM models trained on the simplified pre-reperfusion dataset achieved performance comparable to, and in some metrics (sensitivity), superior to models trained on the full dataset.

Conclusions:

• Random under-sampling (RUS) is an effective technique for enhancing the performance of predictive models in STEMI mortality prediction.
• ML models utilizing a simplified set of variables, collected prior to reperfusion therapy, can accurately identify high-risk STEMI patients.
• This approach facilitates early risk stratification and potentially guides immediate clinical management decisions.