Development and Internal Multicenter Validation of a Deep Learning Model for Predicting Post-Hepatectomy Liver...

Area of Science:

• Hepatobiliary surgery
• Artificial intelligence in medicine
• Oncology

Background:

• Post-hepatectomy liver failure is a critical complication following liver resection for hepatocellular carcinoma (HCC).
• Existing scoring systems and statistical models offer limited accuracy in predicting this complication.
• Multicenter clinical data is crucial for developing robust predictive models.

Purpose of the Study:

• To develop and internally validate a deep learning (DL) model for predicting post-hepatectomy liver failure.
• To compare the predictive performance of the DL model against traditional logistic regression.
• To leverage multicenter data for enhanced model generalizability.

Main Methods:

• A retrospective analysis of 498 patients undergoing HCC liver resection across six centers.
• Integration of preoperative biochemical, intraoperative surgical, and tumor characteristics into a deep neural network.
• Validation using area under the receiver operating characteristic curve (AUC), precision-recall curves, calibration plots, and decision curve analysis (DCA).

Main Results:

• The DL model achieved superior predictive performance with AUCs of 0.914 (training), 0.892 (validation), and 0.906 (test), significantly outperforming logistic regression.
• Key predictors identified include ALBI and MELD scores, prothrombin time, intraoperative blood loss, and resection extent.
• Robust calibration and clinical utility demonstrated through DCA.

Conclusions:

• The developed deep learning model offers significantly improved prediction of post-hepatectomy liver failure compared to logistic regression.
• This model demonstrates potential for enhancing perioperative risk stratification and surgical planning in HCC patients.
• Internal validation across a multicenter cohort supports the model's reliability and clinical applicability.