Machine Learning for Predicting Survival in Poorly Differentiated Neuroendocrine Carcinoma: Pooled Analysis of Four Retrospective Studies

Background: Lung and Digestive system represent the most frequent occurrence sites for neuroendocrine neoplasms (NENs). NEN clinical outcome can be predicted by the integrated evaluation of morphology and proliferation index, allowing NENs distinction into well differentiated neuroendocrine tumors (WD-NET) and poorly differentiated neuroendocrine carcinoma (PD-NEC). We recently published four studies addressed to PD-NECs in order to identify PD-NECs subgroups with better prognosis and/or to predict their therapy outcome. Design: A pooled analysis of our four retrospective studies (3 on digestive [PMID: 26943788; 29592868; 31557757] and 1 on lung NENs [PMID: 32365350]) was performed to evaluate for PD-NEC: i) Best Ki-67 cut-off in predicting Overall Survival (OS); ii) OS according to cancer primary site iii) best common predictive model using different variable selection methods in Cox proportional hazards model (Cox) and machine learning Random Survival Forest (RSF). Results: Overall, 422 PDNECs were analyzed distributed in colorectal (n = 156, 37%), lung (n = 111, 26.3%), gastroesophageal (n = 83, 19.7%), pancreas (n = 42, 10%), ileum-cecum-duodenum (n = 18, 4.2%) and gallbladder/biliary (n = 12, 2.8%). Pure neuroendocrine morphology (n = 227, 53.8%) was slightly more represented than combined (n = 195, 46.2%) although is more frequent in colorectal (59%), gastroesophageal (53%) and gallbladder/biliary (83.3%) compared to lung (31.5%) and pancreas (33.3%). Interestingly, all PDNEC in ileum- cecum-duodenum had pure morphology. Ki-67 at 55% was confirmed as the best value in predicting OS. Colorectal and gastroesophageal PD-NEC showed worse OS than pancreatic and lung. The most predictive Cox model included Ki67 (HR 1.03 - CI95% 1.03-1.04), Morphology (Pure vs Combined - HR 1.44 - CI95% 1.16-1.78), Stage III-IV (HR 1.47 - CI 95% 1.06-2.04) and Age (HR 1.01 - CI 95% 1.00-1.02). In Pancreas PD-NEC HR decreased by 0.58 (CI 95% 0.40-0.83) if compared to colorectal PD-NEC. RSF, which is dependent on the complete combination of all risk factors, confirmed that Ki-67 was the most relevant predictor followed by morphology, stage and site. The predicted response for survival at 1 or 3 years of the forest showed decreasing survival with increasing Ki-67, pure morphology, stage III-IV, and colorectal and gastroesophageal NEC disease. Furthermore, patients with Ki-67<55% and combined morphology had better survival than those with pure morphology. Morphology became irrelevant in NENs when Ki-67 resulted ≥ 55%. RSF had the best predictive accuracy (AUC > 80) compare to other models computed at 6, 12, 24 and 36 months specific time points. Conclusions: The present pooled analysis showed that most predictive parameters to predict OS for PD-NEC patients included Ki67, Morphology, Stage and Site. RSF outperformed all other models in OS prediction performance.