Click here to


Are you sure ?

Yes, do it No, cancel

Pancreatic Cancer Risk Prediction Through An Artificial Neural Network

W Muhammad*, G Hart , K Johung , Y Liang , B Nartowt , I Ali , J Deng , Yale university School of Medicine, New Haven, CT


(Sunday, 7/29/2018) 2:05 PM - 3:00 PM

Room: Karl Dean Ballroom A1

Purpose: The 5-year survival rate for pancreatic cancer (PC) is 8% (American Cancer Society, 2017). It would be highly desirable and clinically important to develop screening tools for PC at an early stage with adequate sensitivity. Currently, the most promising modalities (Table 1) for PC screening are often applied after appearance of the symptoms. Our goal is to develop a multi-parameterized artificial neural network (NN) based on personal health data to predict PC risk with high sensitivity prior to symptom onset.

Methods: The NN was trained and validated with the National Health Interview Survey data (from 1997 – 2016). Among a total of 488519 persons interviewed, 100 were diagnosed with PC. A total of 13 personal health parameters (sex, age, race, Hispanic ethnicity, vigorous exercise habits, smoking status, BMI, hypertension, diabetic status, emphysema, asthma, heart diseases and history of stroke) were used in the NN for PC risk prediction. The NN was trained on 70% of the data and validated on 30%. Sensitivity and specificity were calculated for both training and validation sets and compared with EUS, MRI, CT, and PET.

Results: Areas under the curve of the conventional receiver operating characteristic plots for the NN are 0.761 and 0.731 for the training and validation sets, respectively. The corresponding sensitivity and specificity are 84.3% and 53.3% for the training set, and 80% and 53.6% for the validation set, respectively. Compared to current imaging modalities, our results are lower in terms of sensitivity and specificity (Table 1).

Conclusion: The advantages of our NN lie in its non-invasiveness, cost-effectiveness, and most importantly, the ability to predict PC risk based solely on personal health data prior to symptom onset and without knowledge of family history. This method may identify a high-risk PC population that could benefit from further screening.

Funding Support, Disclosures, and Conflict of Interest: Research reported in this publication was supported by the National Institute of Biomedical Imaging and Bioengineering of the National Institutes of Health under Award Number R01EB022589. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.


Not Applicable / None Entered.


IM- Dataset analysis/biomathematics: Machine learning

Contact Email