Pancreatic hormones response-generated machine learning model to help distinguish sporadic pancreatic cancer from new-onset diabetes cohort.

Abstract

e16235 Background: Diabetes mellitus (DM), a paraneoplastic phenomenon, can develop earlier than other symptoms in pancreatic cancer (PaC) patients. Enhanced surveillance is encouraged on all elderly patients with new-onset DM. However, it is a challenge to differentiate newly developed PaCDM from type 2 DM (T2DM). Thus, we investigated the differences of pancreatic hormones responses and functions between PaCDM and T2DM patients, and developed discriminative model by machine learning algorithms. Methods: PaC patients with normal blood glucose (PaCNG) or with new-onset DM (PaCDM) were recruited. For each case, age and gender matched newly developed T2DM patients and healthy volunteers were selected as controls. After ≥10 hours fasting, all participants underwent a mixed meal stimulation test (MMTT). Blood samples were collected at 0, 15, 30, 60 and 120 min to measure insulin, C-peptide, glucagon, and pancreatic polypeptide (PP). Indices of insulin sensitivity (HOMA-IS, HOMA-IR) and insulin secretion (HOMA-β, insulinogenic index 30’ and 120’) were calculated. Increases in hormone levels were compared among groups with repeated measure analysis. Four machine learning algorithms (Random Forest, Logistic Regression, Support Vector Machines, Naïve bayes) were used to develop quadri-separated discriminative models of PaCDM based on baseline characteristics, pancreatic hormones and insulin indices listed above. Results: Insulin and C-peptide responses to MMTT were blunted in PaCDM patients compared to T2DM. The AUC of insulin were comparatively lower in PaCDM; between-group differences were observed at the fasting (197.15 ± 16.59 pg/mL to 537.96 ± 118.69 pg/mL; P = 0.040) and 15 min (523.94 ± 81.15 pg/mL to 1182.51 ± 219.35 pg/mL; P = 0.036) time-points. No statistical differences among groups were found for glucagon. The mean peak PP concentration after MMTT in PaCDM group (466.67 ± 79.05 pg/mL) was higher than control group (258.54 ± 31.36 pg/mL, P = 0.034), but not statistically different to T2DM patients (452.34 ± 62.96 pg/mL, P = 0.892). PaCDM patients had lower insulin secretion capacity but better insulin sensitivity compared to T2DM patients. Eight indices (age, HbA1c, CA19-9, peak concentration of glucose, area above basal of PP, HOMA-IR, HOMA-IS, HOMA-β) were recruited for model development. And the discriminative model generated by random forest algorithm obtained best performance (AUC = 1.000, CA = 0.963, F-1 = 0.941, Precision = 0.889, Recall = 1.000, Specificity = 0.947; model verified). Conclusions: PaCDM patients tend to present with lower β-cell function and better insulin resistance compared to T2DM patients. As our model based on machine learning algorithm generates a good result for discrimination, the above findings may help with early screening for sporadic PaC in new-onset DM. Clinical trial information: ChiCTR1800018247.