# Individualized Triplet Chemotherapy Decision-Making in Metastatic Colorectal Cancer: A Machine-Learning-Driven Study

## TL;DR

This study uses machine learning to help doctors decide which patients with advanced colorectal cancer are most likely to benefit from intensive chemotherapy.

## Contribution

A machine learning model was developed to personalize triplet chemotherapy decisions in metastatic colorectal cancer.

## Key findings

- A model with 10 key variables achieved a high predictive performance (AUC of 0.919) for progression-free survival.
- The reduced model using top 10 features still performed well (AUC of 0.869).
- Variables like tumor location and blood biomarkers were identified as most influential.

## Abstract

Managing metastatic colorectal cancer is challenging, and although intensive chemotherapy regimens such as FOLFOXIRI or FOLFIRINOX can provide substantial benefit, they are not appropriate for every patient. This study aimed to develop a tool that helps clinicians identify which individuals are most likely to benefit from these intensive treatments. Using data from 136 patients, the researchers built a model based on 10 key variables—including tumor sidedness and several routine blood biomarkers such as ferritin, CA19-9, and CRP—to predict which patients are more likely to experience longer periods without disease progression. This predictive approach may support more personalized treatment decisions by helping ensure that patients receive the most suitable therapy while avoiding unnecessary toxicity. Future prospective studies will be needed to validate the model’s clinical utility.

Objective: The optimal patient subgroup that derives substantial benefit from triplet chemotherapy (FOLFOXIRI/FOLFIRINOX) as first-line treatment for metastatic colorectal cancer (mCRC), and the clinical scenarios in which its increased toxicity is justified, remain uncertain. This study employed a machine learning–based approach to develop a predictive biomarker capable of identifying patients most likely to benefit from triplet therapy. Methods: Clinical data from 136 patients in the Ankara University de novo mCRC cohort were retrospectively reviewed. 66 clinical and biochemical variables were analyzed. Consistent with the existing literature, progression-free survival (PFS) ≥ 270 days was selected as the primary outcome. Individual treatment effect (ITE) estimation was performed using the T-Learner method with separate regression models for each treatment arm (μ1 − μ0). Model performance was evaluated using leave-one-out cross-validation (LOOCV). Feature importance was assessed using SHAP analysis, after which a reduced model was constructed using only the most influential variables. Results: The model incorporating all features demonstrated the highest predictive performance, with a ROC AUC of 0.919. SHAP analysis identified the top 10 predictive variables: primary tumor localization, ferritin, CA19-9, CRP, uric acid, TSH, triglycerides, total protein, LDL, and platelet count. The reduced model built using only these 10 features achieved an AUC of 0.869 for predicting PFS ≥270 days. Conclusion: This machine learning–based model presents a promising framework for improving patient selection for triplet chemotherapy in mCRC. Prospective validation in larger cohorts will be essential to support its integration into clinical decision making.