Determining neuroendocrine tumor (NET) primary sites is pivotal for patient care as pancreatic NETs (pNETs) and small bowel NETs (sbNETs) have distinct treatment approaches. The diagnostic power and prioritization of fluorescence in situ hybridization (FISH) assay biomarkers for establishing primary sites has not been thoroughly investigated using machine learning (ML) techniques. We trained ML models on FISH assay metrics from 85 sbNET and 59 pNET samples for primary site prediction. Exploring multiple methods for imputing missing data, the impute-by-median dataset coupled with a support vector machine model achieved the highest classification accuracy of 93.1% on a held-out test set, with the top importance variables originating from the ERBB2 FISH probe. Due to the greater interpretability of decision tree (DT) models, we fit DT models to ten dataset splits, achieving optimal performance with k-nearest neighbor (KNN) imputed data and a transformation to single categorical biomarker probe variables, with a mean accuracy of 81.4%, on held-out test sets. ERBB2 and MET variables ranked as top-performing features in 9 of 10 DT models and the full dataset model. These findings offer probabilistic guidance for FISH testing, emphasizing the prioritization of the ERBB2, SMAD4, and CDKN2A FISH probes in diagnosing NET primary sites.
Keywords: biomarker; fluorescence in situ hybridization; imputation; machine learning; model; neuroendocrine tumor.