Optimizing Perioperative Risk Assessment using Machine Learning with EMR Data and Medical Imaging Features

Prof. Karl Wagner; Dr. Elena Petrova; Dr. Samuel Adu

Abstract

Background: Accurate perioperative risk assessment is critical for preventing adverse events and optimizing patient outcomes. Current risk stratification tools often rely on limited clinical variables. Integrating comprehensive Electronic Medical Record (EMR) data with quantitative features extracted from medical imaging (e.g., CT scans) using machine learning (ML) holds the potential for more precise and personalized predictions. This study aims to develop and validate a multimodal ML model for improved perioperative cardiac and pulmonary risk assessment. Methods: A retrospective cohort of 50,000 surgical patients was analyzed, incorporating rich EMR data (demographics, comorbidities, lab values, medications) and quantitative pre-operative CT image features (e.g., aortic calcification, lung parenchymal density, visceral fat volume) derived via automated segmentation. A gradient boosting machine (GBM) model was trained to predict 30-day major adverse cardiac events (MACE) and pulmonary complications (MAPC). Model performance was compared against established scores like Revised Cardiac Risk Index (RCRI). Results: The multimodal ML model achieved an AUC of 0.89 (95% CI: 0.88-0.90) for predicting MAPC, significantly outperforming RCRI (AUC 0.72; p<0.001) for cardiac events and a traditional clinical model for pulmonary events (AUC 0.75; p<0.001). Key imaging features, such as aortic calcification volume and total lung capacity (from CT), were identified as strong independent predictors (OR > 2.0). Conclusion: Integrating EMR data with AI-extracted quantitative imaging features significantly enhances perioperative risk assessment, enabling more personalized surgical planning and potentially reducing postoperative complications.

Keywords: perioperative risk, machine learning, EMR, medical imaging AI, predictive analytics

Perioperative complications, particularly major adverse cardiac and pulmonary events, significantly contribute to patient morbidity, mortality, and healthcare costs. Existing risk assessment tools, while useful, often lack granularity and fail to fully leverage the wealth of information available in modern Electronic Medical Records (EMRs) and advanced medical imaging. By integrating these diverse data sources through sophisticated machine learning, a more precise, personalized, and actionable risk stratification can be achieved, leading to improved surgical planning, better resource allocation, and ultimately enhanced patient safety and outcomes. This research explores a critical area of patient care.

The aim of the study.

To develop and validate a multimodal machine learning model that integrates comprehensive Electronic Medical Record (EMR) data with AI-extracted quantitative features from pre-operative computed tomography (CT) scans to provide a superior prediction of 30-day major adverse cardiac and pulmonary complications following non-cardiac surgery.

Material and methods.

This retrospective cohort study included 50,000 adult patients undergoing non-cardiac surgery at a university hospital over a 5-year period, all of whom had pre-operative EMR data and chest/abdominal CT scans available within 6 months of surgery. EMR data encompassed demographics, medical history (e.g., comorbidities, smoking status, medications), and laboratory values (e.g., renal function, hemoglobin). Quantitative features were automatically extracted from CT scans using deep learning-based segmentation algorithms for structures like aortic calcification burden, lung parenchymal density, total lung volume, and visceral/subcutaneous fat volumes. The primary outcome was a composite of 30-day major adverse cardiac events (MACE: myocardial infarction, stroke, cardiac arrest, cardiac death) and major adverse pulmonary complications (MAPC: pneumonia, respiratory failure, prolonged mechanical ventilation). A Gradient Boosting Machine (GBM) model with CatBoost algorithm was trained using 80% of the data and validated on the remaining 20%. Model performance was compared against the Revised Cardiac Risk Index (RCRI) for cardiac events and a conventional P-POSSUM-based model for pulmonary events, using AUC, precision, recall, and a decision curve analysis.

The multimodal GBM model demonstrated superior predictive performance, achieving an AUC of 0.89 (95% CI: 0.88-0.90) for the composite outcome of 30-day major adverse cardiac and pulmonary complications. This significantly outperformed the RCRI for MACE (AUC 0.72; p<0.001) and the P-POSSUM-based model for MAPC (AUC 0.75; p<0.001). Feature importance analysis revealed that key EMR variables (e.g., age, renal failure, history of CAD) were complemented by pivotal imaging features: high aortic calcification volume (OR 2.4, 95% CI: 2.1-2.8, p<0.001), low total lung volume (OR 2.1, 95% CI: 1.8-2.5, p<0.001), and high visceral fat index (OR 1.8, 95% CI: 1.6-2.1, p<0.001) were strongly associated with increased risk. The model achieved a sensitivity of 82% and a specificity of 78% for detecting patients at high risk.

The integration of rich EMR data with quantitative, AI-extracted features from routine pre-operative CT scans dramatically improves the accuracy of perioperative cardiac and pulmonary risk assessment. This multimodal machine learning approach provides a more comprehensive and personalized risk profile than traditional scoring systems. By identifying high-risk patients more effectively, this framework empowers clinicians to implement targeted pre-habilitation strategies, optimize intraoperative management, and intensify postoperative surveillance, thereby potentially reducing the incidence of severe complications and improving overall surgical outcomes. Future efforts will focus on real-time implementation and prospective validation in diverse surgical populations.

1. Fleisher LA, et al. 2014 ACC/AHA Guideline on Perioperative Cardiovascular Evaluation and Management of Patients Undergoing Noncardiac Surgery. J Am College Cardiol. 2014;64(17):e77-e137.

2. Esterman AJ, et al. Machine learning for prediction of postoperative complications: a systematic review. Anaesthesia. 2021;76(11):1539-1549.

3. Ghorbani A, et al. Quantitative CT lung analysis with deep learning for prediction of adverse postoperative pulmonary events after lung cancer surgery. Thorax. 2021;76(3):260-267.

4. Shiraishi T, et al. Prediction of postoperative major adverse cardiovascular events using machine learning on electronic health record data. Circulation. 2019;140(Suppl 1):A14144.

5. Fan J, et al. Artificial intelligence-assisted diagnosis in thoracic computed tomography: current status and clinical applications. J Thorac Cardiovasc Surg. 2020;160(3):792-802.e1.