P2090 - Pancreatic Hormones Response-Generated Machine Learning Model Help Distinguish Sporadic Pancreatic Cancer From New-Onset Diabetes Cohort

Tuesday, October 26, 2021

10:30 AM – 4:00 PM PT

Location: Shoreline Exhibit Hall

Has Audio

Award: Presidential Poster Award

Presenting Author(s)

Chenyu Sun, MD, MSc

AMITA Health Saint Joseph Hospital
Chicago, IL, United States

Award: Presidential Poster Award

Jiantong Bao, MD, PhD¹, Ling Li, MD, PhD¹, Chenyu Sun, MD, MSc², Liang Qi, MD, PhD¹, John Pocholo W. Tuason, MD², Na Hyun Kim, MD², Chandur Bhan, MD², Humaed Mohammed Abdul, MD²
¹Zhongda Hospital, School of Medicine, Southeast University, Nanjing, Jiangsu, China; ²AMITA Health Saint Joseph Hospital, Chicago, IL

Introduction: Diabetes mellitus (DM), a precancerous symptom, can develop earlier in pancreatic cancer (PaC) patients. However, it is a challenge to differentiate newly developed PaCDM from type 2 DM (T2DM), which is also developed in elderly patients but with higher morbidity. Thus, we aimed to differentiate PaCDM patients by their pancreatic endocrine function, and develop a discriminative model by machine learning algorithms.

Methods: PaC patients with normal blood glucose (PaCNG) or with DM (PaCDM) were recruited. Newly developed T2DM patients and healthy volunteers were carefully selected as controls. 44 patients were recruited in this study. After ≥10 hours fasting, all participants underwent a mixed meal tolerance 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) were calculated. Four machine learning algorithms (Random Forest, Logistic Regression, Support Vector Machines, Naïve bayes) were used to develop the quadri-separated discriminative model.

Results: Insulin and C-peptide responses to MMTT were blunted in PaCDM patients compared to T2DM; 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 statsticial differences among groups were found for glucagon. PaCDM patients presented the highest AUC of PP as (37250.83 ± 5821.59) pg/mL × min. PaCDM patients had lower insulin secretion capacity but better insulin sensitivity compared to T2DM. PaCDM patients’ HOMA-β level was (38.92 ± 4.21), lowest among groups. Machine learning approach, i.e. naïve bayes supported by leave-one-out cross-validation was used to build and test the classifier. We established a 8 features model (age, HbA1c, CA19-9, peak concentration of glucose, AUC of PP, HOMA-IR, HOMA-IS, HOMA-β) that achieved 0.965 average AUC-ROC, 0.815 average sensitivity and 0.942 average specificity in distinguishing PaCDM patients.

Discussion: PaCDM patients tend to present with reduced early phase insulin secretion, 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 diabetes cohort.

Figure: Nomogram for pancreatic cancer associated diabetes (PaCDM) discriminative model based on naïve bayes algorithm. Total probabilities (%), calculated from indices points listed above, indicates the probability value for patients to be diagnosed as PaCDM.

Models	AUC	CA	F-1	Precision	Recall	Specificiy
Random Forest	0.934	0.815	0.812	0.828	0.815	0.937
Logistic Regression	0.794	0.519	0.519	0.526	0.519	0.828
SVM	0.893	0.741	0.725	0.753	0.741	0.911
Naïve Bayes	0.965	0.815	0.818	0.835	0.815	0.942

Table: Performance evaluation of the PaCDM discriminative model. Performance metrics showing the average model performance over four classes (control participants, pancreatic cancer patients with normal blood glucose (PaCNG), pancreatic cancer associated diabetes (PaCDM) and type 2 diabetes mellitus patients (T2DM)). AUC, area under the receiver operating characteristic curve. CA, classification accuracy. F-1, F-1 score.

Disclosures:

Jiantong Bao indicated no relevant financial relationships.

Ling Li indicated no relevant financial relationships.

Chenyu Sun indicated no relevant financial relationships.

Liang Qi indicated no relevant financial relationships.

John Pocholo Tuason indicated no relevant financial relationships.

Na Hyun Kim indicated no relevant financial relationships.

Chandur Bhan indicated no relevant financial relationships.

Humaed Mohammed Abdul indicated no relevant financial relationships.

Jiantong Bao, MD, PhD¹, Ling Li, MD, PhD¹, Chenyu Sun, MD, MSc², Liang Qi, MD, PhD¹, John Pocholo W. Tuason, MD², Na Hyun Kim, MD², Chandur Bhan, MD², Humaed Mohammed Abdul, MD². P2090 - Pancreatic Hormones Response-Generated Machine Learning Model Help Distinguish Sporadic Pancreatic Cancer From New-Onset Diabetes Cohort, ACG 2021 Annual Scientific Meeting Abstracts. Las Vegas, Nevada: American College of Gastroenterology.