Who benefits from adjuvant chemotherapy? Identification of early recurrence in intrahepatic cholangiocarcinoma patients after curative-intent resection using machine learning algorithms

TL;DR

This study uses machine learning to identify intrahepatic cholangiocarcinoma patients likely to experience early recurrence after surgery, helping determine who benefits from adjuvant chemotherapy.

Contribution

The novel use of machine learning models, particularly LightGBM, to predict early recurrence and guide adjuvant chemotherapy decisions in ICC patients.

Findings

Early recurrence is an independent prognostic risk factor for overall survival in ICC patients.

LightGBM outperformed other models in predicting early recurrence with high sensitivity and specificity.

Adjuvant chemotherapy significantly improved survival for patients predicted to experience early recurrence.

Abstract

It is vital to enhance the identification of early recurrence in intrahepatic cholangiocarcinoma (ICC) patients after curative-intent resection and to determine which patients could benefit from adjuvant chemotherapy (ACT). This study aimed to evaluate the effectiveness of machine learning algorithms in detecting early recurrence in ICC patients and select those who would benefit from ACT to improve prognosis.

The study analyzed 254 intrahepatic cholangiocarcinoma (ICC) patients who underwent curative-intent resection to identify early recurrence predictors. Through logistic regression and feature importance analysis, we determined key risk factors and subsequently developed machine learning models utilizing the top five predictors for early recurrence prediction. The predictive performance was validated across area under the ROC curve (AUC).

Early recurrence was an independent prognostic risk factor for overall survival (OS) in ICC patients after curative resection (P<0.001). The feature importance ranking based on machine learning algorithms showed that AJCC 8th edition N stage, number of tumors, T stage, perineural invasion, and CA125 as the top five variables associated with early recurrence, which was consistent with the independent risk factors of multivariate logistic regression model. Using the aforementioned five variables, we developed four machine learning prediction models, including logistic regression, support vector machine, LightGBM, and random forest. In the training set, the AUC values were 0.849, 0.860, 0.852, and 0.850, respectively. In the testing set, the AUC values were 0.804, 0.807, 0.841, and 0.835, respectively. Among the various prediction models, LightGBM demonstrated superior performance compared to other models in the testing set, exhibiting higher sensitivity, specificity, and accuracy. The effectiveness of ACT on prognosis for different recurrence times, as predicted by the LightGBM model, indicated that ACT could significantly prolong median OS and RFS for ICC patients predicted to experience early recurrence in both the training and testing sets (P<0.05). Conversely, for ICC patients predicted to have late recurrence, ACT did not improve OS and RFS (P>0.05).

The prediction models established in this study demonstrate good predictive capability and can be used to identify patients who may benefit from ACT.

Figures