Roux-MMC: Advancing Cardiac Surgery Risk Prediction

Roux-MMC advances cardiac surgery risk prediction by outperforming the STS benchmark across eight post-operative outcomes, enabling clinicians to better identify and manage high-risk patients using interpretable ML models trained on real-world data from Maine Medical Center.

The Stakes

Post-operative complications in cardiac surgery demand precise risk stratification for optimal preparation and intervention. The Society of Thoracic Surgeons (STS) model has long set the standard, but machine learning offered potential for greater nuance using pre- and intraoperative variables.

My Role

Contributed to developing and validating Roux-MMC, a suite of ML models predicting eight adverse events: mortality, stroke, renal failure, reoperation, prolonged ventilation, major morbidity/mortality, prolonged length of stay (PLOS), and short length of stay (SLOS).

Data Processing

Analyzed a retrospective cohort of 9,841 STS Adult Cardiac Surgery Database (ACSD) records (2012–2021) and prospective cohort of 2,305 (2022–2024) from Maine Medical Center. Handled real clinical challenges like missing values, inconsistent coding, variable definitions, and PHI safeguards using Python and SQL for cleaning and structuring.

Feature Engineering

Collaborated with clinicians to engineer features capturing surgeon-recognized patterns beyond standard risk scores, including intraoperative variables that boosted predictive power. Emphasized quantifiable insights from domain expertise over automated selection.

Model Development

Developed models prioritizing reliability, interpretability, and calibration alongside accuracy, using SHAP for explanations and validation across retrospective/prospective splits to ensure clinical robustness.

Results

Roux-MMC achieved 0.020–0.167 greater AUROC than STS across all eight outcomes in the prospective cohort:

• Prolonged ventilation: 0.911 AUROC
• Mortality: 0.882 AUROC

These gains highlight improved high-risk patient identification for targeted care.

Translation to Practice

Created clinician-friendly visualizations: SHAP-based risk factor breakdowns, patient comparisons, and uncertainty indicators to bridge ML outputs with surgical decision-making. Ensured explanations focused on intuitive factors like age, comorbidities, and interactions.

Key Lessons

• Interpretability ensures trust in high-stakes settings.
• Domain collaboration yields superior features.
• Data quality and cleaning dominate effort.
• Models must generalize across cohorts without spurious correlations.

Why It Matters

This work demonstrates deploying AI in life-critical domains—translating data into actionable insights that integrate with clinical workflows and support patient outcomes. Publication on medRxiv (DOI: 10.1101/2025.02.24.25322811v1).