# Development and internal-external validation of a risk prediction model for acute pain after HAIC for patients with liver cancer using logistic regression and XGBoost algorithm

## TL;DR

This study develops and validates a model using XGBoost to predict acute pain after chemotherapy in liver cancer patients, showing better performance than logistic regression.

## Contribution

The novel contribution is the development and validation of an XGBoost-based predictive model for post-HAIC pain in liver cancer patients.

## Key findings

- XGBoost model outperformed logistic regression in predicting acute pain after HAIC with higher AUC values.
- XGBoost model showed better calibration and decision curve analysis results in external validation.
- The model identified oxaliplatin dosage, initial HAIC treatment, and age as top risk factors for acute pain.

## Abstract

To develop a clinical model for the early prediction of moderate-to-severe pain after Hepatic Artery Infusion Chemotherapy (HAIC) in liver cancer patients using XGBoost algorithm and then compare its prediction capacity with the logistic model.

A multicenter cohort study presented according to TRIPOD + AI statement, which was conducted in 3 tertiary hospitals in Shanghai from May 2022 to June 2024. Lasso regression was used to screen for risk factors. Logistic regression and XGBoost algorithm were tested and compared by Brier, area under the curve (AUC), calibration curve, Hosmer–Lemeshow test, intercept and slope, and decision curve analysis (DCA).

The study included 1303 patients, with 725 for model development, 578 for external validation. In the XGBoost model, the top 3 most important variables were oxaliplatin dosage, initial HAIC treatment and age. XGBoost model and logistic regression model showed discriminative ability with AUC values of 0.729, 0.714, 0.707 and 0.722, 0.715, 0.684 in the modeling, internal validation, and external validation sets, respectively. The calibration and decision curve analyses of both models showed favorable results in both modeling and validation sets, except for the calibration of logistic regression model in external validation. XGBoost model performed better across all evaluated dimensions in external validation. Based on the risk score generated by the XGBoost model, the population was categorized into low, intermediate, and high-risk subgroups for stratification.

XGBoost model has higher accuracy and stronger robustness in predicting acute moderate-to-severe pain after HAIC in patients with liver cancer, which will facilitate risk assessment and implement precise and early interventions.