Bias in Prediction Models to Identify Patients With Colorectal Cancer at High Risk for Readmission After Resection.

Area of Science:

• Medical Informatics
• Health Services Research
• Machine Learning in Healthcare

Background:

• Machine learning (ML) models are increasingly used for medical predictions.
• Evaluation and reporting of potential biases in these models are often lacking.
• Racial bias in ML models can lead to health disparities.

Purpose of the Study:

• To develop clinical prediction models for 30-day readmission after colorectal cancer (CRC) surgery.
• To examine the potential for racial bias in these developed models.

Main Methods:

• Utilized the American College of Surgeons' National Surgical Quality Improvement Program (ACS-NSQIP) data (2012-2020).
• Compared four ML methods: logistic regression (LR), multilayer perceptron (MLP), random forest (RF), and XGBoost (XGB).
• Assessed model bias using false negative rate (FNR) difference, false positive rate (FPR) difference, and disparate impact.

Main Results:

• Included 112,077 CRC patients across four race groups (White, Black, Other, Unknown/Not Reported).
• Significant differences in AUROC, FPR, and FNR were observed between race groups for all models.
• The 'Other' race category showed higher FNR and consistently lower FPR, meeting disparate impact thresholds for unfairness.

Conclusions:

• Predictive models for post-colorectal surgery readmission may exhibit unequal performance across different race groups.
• Such biases can perpetuate inequalities in healthcare delivery and patient outcomes.
• Urgent need to address and mitigate racial bias in clinical prediction models.