DEVELOPMENT OF AN ARTIFICIAL INTELLIGENCE BASED MODEL TO PREDICT EARLY RECURRENCE OF NEUROENDOCRINE LIVER METASTASIS FOLLOWING RESECTION

Background: Recurrence of neuroendocrine tumor liver metastasis (NELM) following curative-intent resection occurs in a large subset of patients. Artificial intelligence (AI) may help predict outcomes among patients with cancer. We sought to develop an AI based model to predict early recurrence following curative-intent resection of NELM.
Methods: Patients who underwent curative-intent resection for NELM were identified from an international multi-institutional database. Early recurrence was defined as recurrence within 12 months of surgical resection. Different machine learning (ML) and deep learning (DL) techniques were employed to develop prediction models for early recurrence.
Results: 473 NELM patients were randomized into training (n= 378, 79.9%) and testing (n= 95, 20.1%) cohorts. Among 284 (60.0%) patients who developed NELM recurrence with a median follow-up period of 55 months (IQR, 23-94), 118 (41.5%) patients developed an early recurrence. Ten clinicopathological factors were identified as most strongly correlated with early recurrence on correlation heatmap and were used to train the AI models. Among the different AI techniques, an ensemble model consisting of Multi-Layer Perceptron (MLP) and Gradient Boosting (GB) classifiers demonstrated the highest accuracy with area under receiver operating characteristic curve (AUC) of 0.763 (95% CI, 0.705-0.821) and 0.716 (95% CI, 0.593-0.828) in the training and testing cohorts, respectively. Specifically, maximum diameter of the primary neuroendocrine tumor (NET), NELM radiologic tumor burden score (TBS), and bilateral liver involvement were the factors most strongly associated with risk of early NELM recurrence, as demonstrated on the SHAP summary plot (Figure 1). Patients predicted to develop early recurrence had worse 5- (21.4% vs. 37.1%), and 10-year (12.8% vs. 37.1%) recurrence free survival (RFS) versus patients not predicted to recur (p=0.002). Similarly, patients predicted to experience early recurrence had worse 5- (61.6% vs. 90.3%) and 10-year (44.4% vs. 75.8%) overall survival (OS) compared with patients not predicted to recur within 12 months (p=0.03) (Figure 2).
Conclusion: An AI based model demonstrated very good discrimination to predict early recurrence following resection of NELM. AI models such as the one proposed here may help identify which patients benefit the most from curative-intent resection, as well as inform treatment decisions around surgical versus non-surgical treatment options for patients with NELM.