Online interpretable dynamic prediction models for clinically significant posthepatectomy liver failure based on
Yuzhan Jin1,2, Wanxia Li1,2, Yachen Wu3
1School of Basic Medicine and Clinical Pharmacy, China Pharmaceutical University, Nanjing.
International Journal of Surgery (London, England)
|June 18, 2024
Summary
Machine learning models accurately predict posthepatectomy liver failure (PHLF) risk. These models outperformed traditional clinical scores, offering a valuable tool for patient management and improving outcomes after liver surgery.
Area of Science:
- Hepatobiliary surgery
- Medical artificial intelligence
- Clinical decision support
Background:
- Posthepatectomy liver failure (PHLF) is a major cause of mortality following liver resection.
- Accurate prediction models for PHLF are currently lacking, hindering effective patient management.
- This study aimed to develop and validate precise machine learning (ML) models for predicting clinically significant PHLF.
Purpose of the Study:
- To develop and compare the predictive performance of ML models against established clinical scores for PHLF.
- To identify key preoperative and postoperative variables associated with PHLF risk.
- To create interpretable ML models for clinical decision support in liver surgery patients.
Main Methods:
- Development of five preoperative and postoperative ML models using data from 226 hepatectomy patients.
- Comparison of ML models with MELD, FIB-4, ALBI, and APRI clinical scores.
- Internal validation using cross-validation and external validation on an independent temporal dataset, assessing AUC and AUPRC.
- SHapley Additive exPlanations (SHAP) analysis for model interpretability.
Main Results:
- Clinically significant PHLF occurred in 10.2% of patients.
- Preoperative ML models identified creatinine, total bilirubin, and Child-Pugh grade as key predictors.
- Postoperative models additionally included the extent of resection.
- Artificial neural network (ANN) models demonstrated superior performance (AUCs up to 0.851 preoperatively and 0.766 postoperatively) compared to clinical scores (AUCs up to 0.714).
- External validation confirmed the robustness of the ANN models (AUCs 0.720 and 0.731).
Conclusions:
- Interpretable, dynamic ML models significantly outperform traditional clinical scores in predicting PHLF risk.
- These ML models can serve as effective clinical decision support tools for identifying high-risk patients.
- The developed models offer improved preoperative and postoperative risk stratification for patients undergoing hepatectomy.
