463P - Plasma protein fingerprinting and machine learning for the diagnosis of small intestinal neuroendocrine tumors: The nordic NET biomarker group EXPLAIN study

Abstract

Background: Small intestinal neuroendocrine tumors (siNETs) are notoriously difficult to diagnose, especially in an early stage. The EXPLAIN study aimed to investigate 92 plasma proteins (PP), previously shown to be cancer related, in an attempt to improve the accuracy in diagnosis of siNETs. Methods: This non-interventional exploratory study in the nordic countries analysed 136 patients with siNET from 17 hospitals and 144 age and sex matched controls (all with written consent). Exclusion criteria: NET not confirmed, previously treated for NET, other malignant diseases, chronic inflammatory disease, kidney or liver failure. Blood samples (4 ml) were obtained at first visit. Samples analysis used the Proseek Multiplex Oncology II assay (OLink) to measure relative levels of the 92-cancer related PP. In addition, chromogranin A (CgA) was analyzed centrally (Akademiska Lab. Uppsala). Data was subjected to statistical supervised learning techniques (SSLT): random forest and support vector machine. Results: This is the first interim analysis. Patient characteristics: age 65±10 (mean±SD), 58% male, 48% G1 and 52% G2, 88% N1 and 65% M1, 23% >3 bowel mov/d and 11.5% >3 flushes/d. CgA (mean (SD), nmol/L) in 115 patients free from proton pump inhibitor treatment (PPI): 42.37 (86.62), in 21 NET patients with PPI: 68.41 (74.21), in 132 controls free from PPI: 3.67 (3.57) and in 12 controls treated with PPI: 11.83 (8.97). Several PP (>20) showed significant p<.005 different mean levels compared with controls (t-test with Satterthwaite correction). Ten valuable PP in the model: CgA, LYN, ABL1, DKN1A, TXLNA, MUC-16, EGF, MetAP 2, VIM and MK.Table463P Comparison of SSLT modelsSVM – RadialSVM – LinearRandom ForestAccuracy (95% CI)0.8429 (0.7362, 0.9189)0.8714 (0.7699, 0.9395)0.8857(0.7872, 0.9493)Sensitivity0.76470.79410.8824Specificity0.91670.94440.8889AUC0.91910.94280.9404 Open table in a new tab Conclusions: Both a high level of sensitivity and specificity (0.9) were obtained using our multi plasma protein strategy combined with SSLT for the diagnosis of siNET. Further development of the machine learning model is ongoing. Legal entity responsible for the study: Peter Myrenfors Ipsen Funding: Ipsen Disclosure: All authors have declared no conflicts of interest.