Ensemble Machine Learning Model Incorporating Radiomics and Body Composition for Predicting Intraoperative HDI in...

Area of Science:

• Endocrinology
• Oncology
• Medical Imaging
• Machine Learning in Medicine

Background:

• Intraoperative hemodynamic instability (HDI) poses significant risks during pheochromocytoma/paraganglioma (PPGL) surgery, potentially leading to severe complications.
• Effective prediction of HDI is essential for improving surgical outcomes in PPGL patients.

Purpose of the Study:

• To assess the risk of intraoperative HDI in patients undergoing PPGL surgery.
• To develop and evaluate predictive models for intraoperative HDI.

Main Methods:

• Retrospective analysis of 199 consecutive PPGL patients, divided into hemodynamic instability (HDI) and hemodynamic stability (HDS) groups.
• Development of prediction models using ensemble machine learning (EL) and multivariate logistic regression.
• Models incorporated computed tomography-based body composition, tumor radiomics, and clinical data.

Main Results:

• The EL model demonstrated superior predictive performance, with an Area Under the Curve (AUC) of 96.2% in the training cohort and 93.7% in the validation cohort.
• EL model's AUC significantly outperformed the logistic regression model (74.4% training, 74.2% validation).
• The EL model exhibited favorable calibration and clinical applicability.

Conclusions:

• An ensemble machine learning model integrating preoperative CT-derived body composition, tumor radiomics, and clinical data can effectively predict intraoperative HDI in PPGL patients.
• This predictive capability can aid in optimizing surgical management and improving patient outcomes.