Deep Learning to Predict Mortality After Cardiothoracic Surgery Using Preoperative Chest Radiographs

Area of Science:

• Artificial Intelligence in Medicine
• Cardiovascular Surgery
• Radiology

Background:

• The Society of Thoracic Surgeons Predicted Risk of Mortality (STS-PROM) is limited to specific cardiac surgeries and requires extensive data input.
• Estimating mortality risk for all cardiac surgery types, including non-STS index procedures, is crucial for comprehensive patient care.

Purpose of the Study:

• To develop and validate a deep learning model (CXR-CTSurgery) for predicting postoperative mortality risk using preoperative chest radiographs.
• To compare the performance of CXR-CTSurgery against the STS-PROM for both STS index and non-STS index cardiac surgical procedures.

Main Methods:

• A deep learning model, CXR-CTSurgery, was developed using data from 9283 cardiac surgery patients at Massachusetts General Hospital (MGH).
• The model was validated internally on 3615 MGH patients and externally on 2840 Brigham and Women's Hospital (BWH) patients.
• Model performance was assessed using the C-statistic for discrimination and the observed-to-expected (O/E) ratio for calibration, compared to STS-PROM.

Main Results:

• CXR-CTSurgery demonstrated comparable discrimination to STS-PROM for STS index procedures at both MGH and BWH.
• The deep learning model showed similar C-statistics for non-STS index procedures compared to STS index procedures.
• CXR-CTSurgery exhibited superior calibration (better O/E ratios) compared to STS-PROM for STS index procedures in both testing cohorts.

Conclusions:

• Deep learning models utilizing chest X-rays can effectively predict postoperative mortality in cardiac surgery patients.
• CXR-CTSurgery offers a valuable alternative or supplement to STS-PROM, particularly when STS-PROM is not applicable or calculable.
• This AI-driven approach enhances risk stratification for a broader range of cardiac surgical procedures.