Efficacy of Postoperative Adjuvant PD-1 Inhibitors Guided by a Deep Learning Model: a Multicenter, Prospective Cohort Study
Efficacy of Postoperative Adjuvant PD-1 Inhibitors Guided by a Deep Learning Model: a Multicenter, Prospective Cohort Study
Early hepatocellular carcinoma (HCC) recurrence (driven by residual tumors) and late recurrence (driven by de novo tumors) exhibit distinct biological behaviors, suggesting differential therapeutic vulnerabilities. The beneficiaries of adjuvant PD-1 inhibitors (aPD-1) and their efficacy across these temporally divergent recurrence patterns remains unestablished.
Hepatocellular carcinoma (HCC), a leading cause of global cancer-related mortality, continues to rise in incidence and lethality despite advancements in early detection and surgical techniques. Curative liver resection, while the cornerstone of therapeutic management, is frequently undermined by postoperative recurrence, a phenomenon observed in up to 70% of patients within five years, with early (≤2 years) and late (>2 years) recurrences reflecting distinct biological origins. Early recurrences predominantly stem from residual micro-metastases of the primary tumor, strongly associated with aggressive histopathological features such as microvascular invasion (MVI), multifocality, and satellite nodules. In contrast, late recurrences often arise de novo from the cirrhotic liver microenvironment, driven by persistent viral activity or chronic hepatic inflammation rather than the index tumor's biological behavior. Despite decades of research, postoperative adjuvant strategies, including antiviral therapy, transarterial chemoembolization, and traditional agents like Huaier granules, have yielded inconsistent results or lack robust evidence for standardization. The emergence of immune checkpoint inhibitors (ICIs) has reignited hope, yet recent randomized controlled trials (RCT) underscore unresolved challenges. The IMbrave050 trial initially demonstrated reduced recurrence with adjuvant Atezolizumab-Bevacizumab (median follow-up of 17 months). However, with longer follow-up (35 months), results shifted to negative. Another RCT has shown promising outcomes for patients with MVI-positive HCC who received adjuvant therapy with Sintilimab. Nevertheless, the median follow-up was only 23 months, which does not provide adequate resolution of late recurrence. Similarly, a recent prospective cohort study reported positive results of adjuvant immunotherapy in high-risk patients. These studies, limited by follow-up durations insufficient to capture late-recurrence dynamics, leave critical questions unanswered: Do adjuvant ICIs durably suppress recurrence, or merely delay its onset? In addition, there is currently no gold standard for defining high recurrence risk. Common pathological factors include MVI and satellite nodules13, but the same patient may have multiple high-risk factors simultaneously. Machine learning (ML) is increasingly being used in the construction of predictive models, and its performance often exceeds that of models based on standard statistical methods and traditional staging systems14, 15. Therefore, there is great potential for using ML to integrate clinical and pathological characteristics and quantify these risk factors to accurately identify high-risk populations and guide postoperative management strategies.
In order to fill these research gaps, we constructed an ML model to predict the risk of HCC recurrence through previous studies, and found that high-risk groups were more likely to be the potential benefit population of HCC. Therefore, this study aimed to verify the value of ML model in guiding postoperative adjuvant PD-1 inhibitors.
Inclusion Criteria:
Exclusion Criteria:
wgzhang@tjh.tjmu.edu.cn+8613886195965