Machine learning prediction model for post- hepatectomy liver failure in hepatocellular carcinoma: A multicenter study

Jitao Wang, Tianlei Zheng, Yong Liao, Shi Geng, Jinlong Li, Zhanguo Zhang, Dong Shang, Chengyu Liu, Peng Yu, Yifei Huang, Chuan Liu, Yanna Liu, Shanghao Liu, Mingguang Wang, Dengxiang Liu, Hongrui Miao, Shuang Li, Biao Zhang, Anliang Huang, Yewei Zhang, Xiaolong Qi, Shubo ChenFrontiers in Oncology 2022 · DOI

TL;DR

This paper introduces a novel machine learning model, built using the LightGBM algorithm, to predict post-hepatectomy liver failure (PHLF) in patients with hepatocellular carcinoma (HCC). The model demonstrates superior predictive performance compared to existing clinical scoring systems, offering a more accurate and potentially life-saving tool for patient management.

What Problem Does This Paper Address?

Hepatocellular carcinoma (HCC) is a significant global health concern, and radical liver resection is a primary treatment. However, a major challenge following this surgery is post-hepatectomy liver failure (PHLF), a serious complication that can lead to death. While various scoring systems and nomograms exist to predict PHLF, none are universally recognized or consistently accurate. This leaves a critical gap in accurately identifying patients at high risk, hindering personalized treatment and potentially leading to preventable adverse outcomes. The authors aim to bridge this gap by developing a more robust and accurate predictive model using machine learning.

Methods

The core of this study lies in the development and validation of a machine learning model. The researchers employed the LightGBM (Light Gradient Boosting Machine) algorithm, known for its speed and efficiency in handling large datasets. They utilized a multicenter dataset of 875 HCC patients who underwent hepatectomy.

The process involved:

  • Data Splitting: The dataset was divided into training (70%), validation (10%), and testing (20%) cohorts to ensure robust model evaluation and prevent overfitting.
  • Feature Selection: A total of 25 clinically relevant variables were considered. The Shapley Additive Explanations (SHAP) method was used to quantitatively assess the importance of each variable and understand their contribution to the model's predictions. This is a key technique for interpreting complex ML models.
  • Model Optimization: The LightGBM algorithm was fine-tuned by adjusting parameters like the number of iterations, leaves, and maximum depth to achieve optimal performance.
  • Performance Evaluation: The model's accuracy was assessed using standard metrics: Area Under the Receiver Operating Characteristic Curve (AUC), sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV).
  • Clinical Utility Assessment: Decision Curve Analysis (DCA) was employed to evaluate the clinical usefulness of the developed model and compare it against traditional scoring systems.

Datasets

The study utilized a multicenter dataset comprising 875 patients with HCC who underwent radical hepatectomy. This dataset was collected from four hospitals: The Xingtai People's Hospital, The Second Affiliated Hospital of Nanjing Medical University, Fifth Medical Center of People's Liberation Army (PLA) General Hospital, The First Affiliated Hospital of Dalian Medical University, and Tongji Hospital Affiliated to Huazhong University of Science and Technology. The data was split into training (n=612), validation (n=88), and testing (n=175) cohorts. No new benchmark datasets were introduced; rather, existing patient data was leveraged.

Key Insights and Ideas

  • Platelet Count is King (for PHLF Prediction): The SHAP analysis revealed that platelet count (PLT) was the most significant predictor of PHLF. This aligns with existing research but highlights its paramount importance in this ML model.
  • Multivariate Power of ML: The study underscores the advantage of ML in integrating multiple clinical parameters (25 in this case) to achieve superior predictive power compared to simpler, single-indicator scoring systems.
  • LightGBM's Efficacy: The choice of LightGBM proved effective, enabling the development of a model that is both accurate and computationally efficient.
  • Clinical Utility Beyond Accuracy: DCA demonstrated that the ML model not only achieves higher accuracy but also offers greater clinical utility, meaning it can help clinicians make better decisions across a wider range of risk thresholds.

Key Results

The developed ML PHLF model demonstrated strong predictive performance across all cohorts:

  • AUCs: 0.944 (training), 0.870 (validation), and 0.822 (testing).
  • Sensitivity and Specificity: The model achieved high sensitivity (e.g., 87.6% in training, 100% in validation) and good specificity (e.g., 85.9% in training, 64.4% in validation), indicating its ability to correctly identify both patients who will and will not develop PHLF.
  • Superiority over Traditional Models: The ML model consistently outperformed established clinical scoring systems like ALBI, FIB-4, APRI, MELD, and CTP in terms of AUC across all cohorts.
  • Web Calculator: A web calculator based on the ML model has been developed to facilitate its clinical application.

Limitations and Future Work

The authors acknowledge several limitations:

  • Selection Bias: While a multicenter design aimed to mitigate this, some selection bias may still exist.
  • Interpretability: ML models, including LightGBM, can be considered "black boxes." While SHAP provides some interpretability, further research into more interpretable ML algorithms is suggested.
  • Overfitting Potential: This is a general concern with ML models, and ongoing validation is crucial.
  • Subgroup Analysis: The current model predicts overall PHLF risk. Future work should investigate its predictive value in specific patient subgroups (e.g., CTP class B and C) and for other PHLF diagnostic criteria.

The authors propose future research to address these limitations, including exploring more interpretable ML approaches and validating the model in diverse patient populations and diagnostic criteria.

Summary updated

Abstract

A novel ML model for the prediction of PHLF using common clinical parameters was constructed and validated. The novel ML model performed better than did existing noninvasive models for the prediction of PHLF.
BibTeX
@inproceedings{wang2022machine,
  author = {Jitao Wang and Tianlei Zheng and Yong Liao and Shi Geng and Jinlong Li and Zhanguo Zhang and Dong Shang and Chengyu Liu and Peng Yu and Yifei Huang and Chuan Liu and Yanna Liu and Shanghao Liu and Mingguang Wang and Dengxiang Liu and Hongrui Miao and Shuang Li and Biao Zhang and Anliang Huang and Yewei Zhang and Xiaolong Qi and Shubo Chen},
  title = {{Machine learning prediction model for post- hepatectomy liver failure in hepatocellular carcinoma: A multicenter study}},
  booktitle = {Frontiers in Oncology},
  year = {2022},
  doi = {10.3389/fonc.2022.986867},
}
Figure