Harnessing hybrid stacking ensemble learning for accurate pulmonary embolism diagnosis using tabular clinical data

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Clinical Decision Support Systems

Background:

• Pulmonary Embolism (PE) is a critical condition requiring prompt diagnosis for effective treatment.
• Accurate prediction of PE from clinical data is essential for timely medical intervention.
• Existing diagnostic methods may benefit from enhanced predictive modeling.

Purpose of the Study:

• To develop and evaluate a Hybrid Stacking Ensemble (HSE) framework for Pulmonary Embolism (PE) prediction.
• To leverage diverse machine learning models and an optimization algorithm for improved predictive performance.
• To assess the framework's accuracy and robustness using the RSNA-STR-PE dataset.

Main Methods:

• A hybrid stacking ensemble (HSE) framework was designed, integrating SAINT transformer, XGBoost, LightGBM, and MLP base learners.
• The Marine Predators Algorithm (MPA) was employed to optimize the weighting of base model outputs.
• A logistic regression meta-learner with L2 regularization was used for final prediction fusion.

Main Results:

• The MPA-optimized-HSE achieved a prediction accuracy of 92.3% and an AUROC of 0.91.
• The proposed ensemble model significantly outperformed individual base learners in PE prediction.
• The framework demonstrated robust performance and interpretability on tabular clinical data.

Conclusions:

• The MPA-optimized-HSE offers a powerful and reliable method for enhancing PE prediction from clinical data.
• This approach shows significant potential for improving diagnostic accuracy in real-world clinical settings.
• The study highlights the efficacy of ensemble methods combined with optimization algorithms in medical diagnostics.