Tree-based survival analysis improves mortality prediction in cardiac surgery.

Area of Science:

• Cardiovascular Surgery
• Medical Informatics
• Machine Learning

Background:

• Machine learning classification tools are established for predicting cardiac surgical outcomes.
• ML-based survival analysis is an underexplored method for predicting post-cardiac surgery mortality.

Purpose of the Study:

• To benchmark the performance of tree-based survival models against Cox proportional hazards (CPH) modeling for predicting cardiac surgery mortality.
• To identify key risk factors for mortality using the best-performing ML model.

Main Methods:

• Utilized a national database of 144,536 patients (147,301 surgery events).
• Compared three ML models: decision tree (DT), random forest (RF), and gradient boosting machine (GBM) using 2-fold cross-validation.
• Assessed performance using the concordance index (C-index).

Main Results:

• Gradient Boosting Machine (GBM) achieved the highest C-index (0.803), outperforming Random Forest (0.791), Decision Tree (0.729), and Cox Proportional Hazards (0.596).
• The top predictors of mortality included age, procedure type, length of hospital stay, early postoperative drain output, and duration of inotrope use.

Conclusions:

• Tree-based survival analysis presents a non-parametric and high-performing alternative to traditional CPH modeling.
• GBMs provide interpretable insights into non-linear relationships, highlighting critical risk factors and guiding future cardiac surgery research.