SIR 2024
Interventional Oncology
Rajesh Shah, MD
Director of Interventional Radiology
California Pacific Medical Center
Financial relationships: Full list of relationships is listed on the CME information page.
Nishita Kothary, MD
Professor of Radiology
Stanford University School of Medicine
Disclosure information not submitted.
To determine if the 30-day mortality could be accurately predicted using machine learning in patients undergoing any treatment, including chemotherapy, loco-regional therapy or surgery for hepatocellular carcinoma.
Materials and Methods:
63,793 unique patients undergoing 119,927 procedures with a diagnosis of hepatocellular carcinoma (HCC) with a treatment date between 01/2000 and 12/2022 were selected from the Veterans Affairs Informatics and Computing Infrastructure. Features selected for inclusion were laboratory studies including albumin, hemoglobin, alpha fetoprotein, white blood cell count, platelets, ALT, AST, bilirubin, Cr, INR, demographics including age, race, sex, location of treatment, and measures of comorbities including the Charlson-Deyo index, MELD, and ALBI. Thirty-day mortality was calculated using the procedure date and the death date. Procedures evaluated included surgical resection, laparascopic ablation, percutaneous ablation, chemoembolization, radioembolization, transplant, monoclonal antibody therapy (e.g. Sorafenib), and immunotherapy. We undersampled the data to balance the classes. A Random Forest (RF) machine learning (ML) classifier was used to train and test the data in an 80:20 stratified split to predict 30-day mortality after treatment. Area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy, and F1 score were obtained. Feature importance was determined.
Results: Data was available for 15,717 unique patients undergoing 53,161 procedures. The RF classifier showed an AUROC of 0.859, Accuracy of 0.786, Sensitivity of 0.786, Specificity of 0.787, PPV of 0.761, NPV of 0.809, and F1 score of 0.773. The top four features based on RF feature importance were albumin, MELD score, hemoglobin, and ALBI score.
Conclusion:
An ML RF model shows high accuracy and positive predictive value in predicting 30-day mortality after treatment for HCC. Further refinements to the classifier are needed to improve performance including exploring other ML techniques and potentially delving into deep learning approaches. An accurate ML classifier could help in decision making for therapies for HCC.