Development of a transformer model for predicting the prognosis of patients with hepatocellular carcinoma after radiofrequency ablation

Abstract

Introduction Radiofrequency ablation (RFA) is a widely accepted, minimally invasive treatment modality for patients with hepatocellular carcinoma (HCC). Accurate prognosis prediction is important to identify patients at high risk for cancer progression/recurrence after RFA. Recently, state-of-the-art transformer models showing improved performance over existing deep learning-based models have been developed in several fields. This study was aimed at developing and validating a transformer model to predict the overall survival in HCC patients with treated by RFA. Methods We enrolled a total of 1778 treatment-naïve HCC patients treated by RFA as the first-line treatment. We developed a transformer-based machine learning model to predict the overall survival in the HCC patients treated by RFA and compared its predictive performance with that of a deep learning-based model. Model performance was evaluated by determining the Harrel’s c-index and validated externally by the split-sample method. Results The Harrel’s c-index of the transformer-based model was 0.69, indicating its better discrimination performance than that of the deep learning model (Harrel’s c-index, 0.60) in the external validation cohort. The transformer model showed a high discriminative ability for stratifying the external validation cohort into two or three different risk groups (p < 0.001 for both risk groupings). The model also enabled output of a personalized cumulative recurrence prediction curve for each patient. Conclusions We developed a novel transformer model for personalized prediction of the overall survival in HCC patients after RFA treatment. The current model may offer a personalized survival prediction schema for patients with HCC undergoing RFA treatment.

Can Machine Learning Predict Favorable Outcome After Radiofrequency Ablation of Hepatocellular Carcinoma?

“…Our model demonstrated an accuracy and an AUC of 90.6% and 0.95, respectively, for the training set, while they were 78.9% and 0.80, respectively, for the validation set. These numbers exceeded the accuracy and AUC of the two models reported by Tong et al 16 and Sato et al 15 Probably, the gap in AUC between our model and theirs was caused by an overfitting bias in our model. Our study opens the door for health care systems with limited resources such as ours to develop their own tools in the era of personalized medicine.…”

“…In our study, the incidence of alcohol consumption was 0% compared with 20% alcohol consumption in the Japanese population. 15 Moreover, the percentage of patients with worse hepatic function was higher in our study. In our study, Child B represented 44% compared with 21.6% in the study by Sato and 0% in the study by Tong.…”

“…The author used a transformer based on a neural network structure that achieved a Harrel's c-index of 0.69 for the training sets and 0.60 for the validation set. 15 The second study was conducted by Tong et al 16 It included 578 patients from a publicly available data set that stored data from University of Tokyo, which was the same hospital as in the first study. The authors reported an AUC of 0.72.…”

Alpha-fetoprotein: Past, present, and future

“…Future studies should attempt to leverage machine-learning algorithms to refine the diagnostic, predictive, prognostic, and therapeutic capacity of AFP, which will likely further enhance the role of AFP in personalized patient care of HCC. 143 …”