Prediction of Pathologic Complete Response for Rectal Cancer Based on Pretreatment Factors Using Machine Learning
Kevin A Chen1, Paolo Goffredo2, Logan R Butler1
1Division of Gastrointestinal Surgery, Department of Surgery, University of North Carolina at Chapel Hill, Chapel Hill, North Carolina.
Diseases of the Colon and Rectum
|November 23, 2023
Summary
Machine learning accurately predicts pathologic complete response in locally advanced rectal cancer before treatment. This can help identify patients suitable for nonoperative management, improving treatment strategies.
Area of Science:
- Oncology
- Machine Learning
- Predictive Modeling
Background:
- Pathologic complete response (pCR) is a key prognostic indicator for locally advanced rectal cancer (LARC).
- Accurate prediction of pCR in the pretreatment setting is crucial for treatment de-escalation and nonoperative management strategies.
- Existing predictive models for pCR in LARC are limited by small datasets and low accuracy.
Discussion:
- Machine learning models, particularly gradient boosting, demonstrate superior performance in predicting pCR compared to traditional logistic regression.
- Key predictors of pCR include absence of lymphovascular invasion, absence of perineural invasion, lower CEA levels, smaller tumor size, and microsatellite stability.
- A concise model incorporating the top 5 predictors maintains high predictive performance.
Key Insights:
- Gradient boosting achieved an AUC of 0.777, outperforming logistic regression (AUC 0.684).
- The study utilized a large, national dataset (53,684 patients) from 2010-2019.
- Identified significant pretreatment predictors for pCR in LARC.
Outlook:
- Developed machine learning models can accurately predict pCR for LARC patients.
- Future refinement with nonoperatively treated patients could enable precise identification of candidates for watch-and-wait strategies.
- These models hold potential for personalized treatment planning in LARC.
