Objective: This study aims to use machine learning methods combined with heat shock protein 90 alpha (HSP90α) to identify patients with unresectable hepatocellular carcinoma (HCC) who are likely to benefit from transarterial chemoembolization (TACE) treatment and to evaluate the prognosis of patients receiving TACE therapy. Methods: A total of 2555 unresectable HCC patients were initially enrolled at seven Chinese tertiary hospitals between 2016 and 2021. Residual methods were used to identify the TACE benefit population. Eight machine learning models were applied to select predictive factors for TACE benefit. Based on these factors, an online nomogram was constructed to predict which patients are likely to benefit from TACE. Subsequently, 101 machine learning models were developed to assess 1-, 2-, and 3-year overall survival (OS) in TACE-treated patients. Results: Before (TACE, 1429; No TACE, 1126) and after propensity score matching (TACE, 666; No TACE, 666), patients in the TACE group showed significantly better outcomes compared to the No TACE group (all p < 0.05). Residual methods successfully identified the TACE benefit group. Subsequently, eight machine learning models were constructed, and the intersection of these models was identified using an Upset plot, which revealed three key predictive factors for TACE benefit: HSP90α, barcelona clinic liver Cancer (BCLC), and tumor size. An online nomogram based on these three factors was constructed in the training set. In the validation set, the model achieved an area under the receiver operating characteristic curve (AUC-ROC) value of 0.901 for predicting TACE benefit. To further assess OS in unresectable HCC patients receiving TACE, 101 machine learning models were developed. Among these models, the StepCox [forward] + RSF model exhibited the highest C-index, with values of 0.84, 0.80, and 0.82 in the training set, internal validation set, and external validation set, respectively. The top five variables most strongly associated with OS were identified as HSP90α, BCLC, number of tumors, tumor size, and alkaline phosphatase (ALP). The AUC-ROC values for predicting 1-, 2-, and 3-year OS in the internal validation set were 0.943, 0.951, and 0.888, respectively. In the external validation set, the AUC-ROC values were 0.953, 0.903, and 0.883. Conclusion: Machine learning techniques combined with HSP90α expression effectively identify patients who are likely to benefit from TACE and provide accurate prognostic assessments for TACE treatment.

Machine Learning, HSP90α, Hepatocellular Carcinoma, TACE, Nomogram