Abstract 5023: Machine learning prediction of perineural invasion in hepatobiliary cancers

Background: Hepatobiliary cancers, which affect the liver, gallbladder, and bile ducts, are aggressive malignancies with a poor prognosis. Factors such as perineural invasion (PNI) can further complicate these cancers, as it can facilitate the spread of cancer cells to distant organs, reducing treatment effectiveness and overall survival. In this study, we aimed to develop machine learning (ML) models to predict PNI in patients with hepatobiliary cancers. Methods: After data processing and cleaning, our dataset included profiles of 433 samples of hepatobiliary cancers from the Memorial Sloan Kettering Cancer Center 2024 HCC dataset. Binary logistic regression (BLR) was used to investigate predictors of PNI. Various ML models were developed by python libraries, such as “scikit-learn” and “TensorFlow”, then trained and tested on 80-20 split. Ten classification models were built with several features including demographic, clinicopathological, and genomic variables such as fraction genome altered, microsatellite instability, and tumor mutational burden. GridSearch cross-validation was used for hyperparameter tuning. Model performance was evaluated on test sets and SHapley Additive exPlanations (SHAP) analysis was applied to interpret feature contributions. Results: Of the 433 samples, 25.9% (n = 112) were extrahepatic cholangiocarcinoma, 24.7% (n = 107) were gallbladder cancer, 13.6% (n = 59) were hepatocellular carcinoma, 32.6% (n = 141) were intrahepatic cholangiocarcinoma, and 3.2% (n = 14) were biphenotypic tumors. Cancer subgroup was a significant predictor in BLR analysis, with extrahepatic cholangiocarcinoma having the highest proportion of PNI. Among the ML models, Decision Tree demonstrated the highest precision at 0.902. On the other hand, according to area under the curve (AUC), Logistic regression (LR), Gradient Boosting (GB), and Random Forest (RF) were the top three performing models (AUC = 0.892, 95% CI [0.845, 0.939]; AUC = 0.889, 95% CI [0.841, 0.936]; and AUC = 0.888, 95% CI [0.841, 0.936], respectively). Cancer subgroup was the most influential feature among these top three models, and primary tumor site ranked second in the RF model, according to SHAP analysis. Multilayer perceptrons were the least-performing, achieving AUC values of 0.797 (95% CI: 0.736-0.858) for the single-hidden-layer model and 0.758 (95% CI: 0.693-0.823) for the four-hidden-layer model. Conclusion: This study demonstrates the use of ML in predicting PNI in hepatobiliary cancers, with LR, GB, and RF models achieving acceptable performance. These models can improve disease management and evaluation, allowing for personalized treatment strategies by predicting tumor aggressiveness and invasive potential. Citation Format: Jehad A. Yasin, Abd-alrahman Obeid, Lujia Chen, Mohammad-Amer A. Tamimi, Fares A. Qtaishat, Osama Younis, Ahmad E. Saeed, Yousef A. Ateiwi, Ramez M. Odat, Mohammad M. Alghaniem, Nada Al-Awamleh, Azhar Saeed, Anwaar Saeed. Machine learning prediction of perineural invasion in hepatobiliary cancers [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 5023.