Abstract 88P
Background
Chemotherapy with metastasectomy plays a crucial role for colorectal liver metastasis (CRLM). Chemotherapy-induced hepatotoxicity (CIH) can interrupt maintaining chemotherapy, especially in high-risk patients with CRLM who underwent hepatic metastasectomy. This study aimed to develop the predictive models for CIH in patients with CRLM.
Methods
This study included 2890 cases with CRLM who underwent hepatic metastasectomy between 2015 and 2021 at Korea University Anam Hospital. CIH was defined as follows: 1) if baseline AST/ALT ≤ 45mg/dL, follow-up AST/ALT during chemotherapy ≥ 100 mg/dL or 2) if baseline AST/ALT > 45mg/dL, follow-up AST/ALT during chemotherapy ≥ 100 mg/dL and at least 2-fold the baseline level. The data included several potential risk factors for hepatotoxicity. Three machine learning methods, including XGBoost, Light GBM, and Random Forest (RF), were used to develop the predictive model for CIH.
Results
CIH occurred in 56 events (1.9%). The AUCs for prediction of CIH was 0.80, 0.77, and 0.83 in XGBoost, Light GBM, and RF, respectively. Both XGBoost and Light GBM provided an accuracy of 0.97, a sensitivity of 0.60, and a specificity of 0.95 and 0.93, respectively. RF provided an accuracy of 0.97, a sensitivity of 0.50, and a specificity of 0.97. The F1 Score is the highest for LightGBM and XGBoost (0.75), suggesting the best performance in balancing sensitivity and precision. Despite its highest AUC, Random Forest has a lower F1 score (0.66), hinting at a trade-off between sensitivity and specificity.
Conclusions
These machine learning models can identify patients with a greater likelihood of risk from CIH. It can be helpful for patients with metastatic colorectal cancer requiring long-term exposure to chemotherapy. Further studies are necessary to validate and update these models in independent datasets in future.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
The authors.
Funding
Has not received any funding.
Disclosure
All authors have declared no conflicts of interest.