Perioperative Risk Assessment to Predict Hemodynamic Instability in Cardiac Surgery Patients
- Program: Biomedical Engineering
- Course: EN.580.480 Precision Care Medicine
Project Description:
This project aims to enhance the prediction of hemodynamic instability in cardiac surgery patients undergoing cardiopulmonary bypass, addressing the critical challenge of perioperative instability that can lead to severe complications and increased healthcare utilization. The project integrates comprehensive preoperative, intraoperative, and postoperative data, focusing on dynamic changes during surgery often overlooked by current models. A novel method was developed to extract stiffness indices from time-series pulmonary arterial pressure data, transforming raw measurements into predictive inputs. Machine learning models were implemented and optimized through extensive hyperparameter tuning and cross-validation for classification tasks. Correlation matrices focused on the target variable to determine feature importance, ensuring that the most predictive elements were utilized to forecast patient outcomes effectively. This comprehensive approach improves predictions of postoperative outcomes and offers real-time insights that enhance patient management.
Project Photo:
A platypus wearing headphones, typing code on a computer. Behind the platypus scientist are floating icons representing cardiac surgery: an EKG readout, a detailed heart diagram, and a hospital ICU scene.
Project Poster
Open full size poster in new tab (PDF)
Project Post Summary:
Cardiac surgeries often involve risks of hemodynamic instability, which can lead to extended ICU stays and increased healthcare costs. Current models primarily use pre- and post-operative data, neglecting intraoperative dynamics that are crucial for real-time patient management.
This project employs a structured approach to predict postoperative outcomes in cardiac surgery patients, integrating time-series feature extraction from intraoperative data and predictive modeling using machine learning. The methodology includes data preprocessing, exploratory analysis, and advanced feature engineering. A risk classification model categorizes patients by ICU stay risk, and feature validation using Shapley analysis assesses the impact of novel engineered indicators, creating a comprehensive framework to improve perioperative care.
Analysis demonstrated that intraoperative medication dosages and the novel indicator, stiffness, are key predictors for the classification target of ICU length of stay. The successful use of Random Forest and KNN classifier models demonstrates that intraoperative data enhances risk assessment for cardiac surgery patients.
Course Faculty
- Joseph L Greenstein
- Casey Overby Taylor