Click here to


Are you sure ?

Yes, do it No, cancel

Liver Cancer Risk Quantification with Smoking Through Artificial Neural Network

W Muhammad1*, G Hart1 , B Nartowt1 , Y Liang2, Deng1 , (1) Yale School of Medicine, Yale University, New Haven, CT, (2) Medical College of Wisconsin, Milwaukee, WI,


(Tuesday, 7/16/2019) 10:30 AM - 11:00 AM

Room: Exhibit Hall | Forum 6

Purpose: Current smokers are at higher risk for liver cancer compared to former or non-smokers. This study models a reduction in liver cancer risk of an individual using a multi-parameterized neural network (NN) in response to quitting smoking.

Methods: A basic health data of total 795,163 persons including 478 liver cancer were acquired from National Health Interview Survey (NHIS) data and the Prostate, Lung, Colorectal, Ovarian cancer screening trail data. A 70 % and 30 % random split of the data was used to respectively train and test the NN. The testing population was stratified into low-, medium- and high-risk groups. From the high-risk group in the testing population, 17 non-cancer respondents who were obese (BMI > 0.30), current smokers, and relatively heavy drinkers (drinks > 100 days/year) were selected. The smoking status of these 17 respondents was changed from “current� to “former�, and their risk for liver cancer was recalculated.

Results: Our NN has an area under the curve (AUC) concordance of the conventional receiver operating characteristic plots of 0.83 and 0.85 for the training and testing datasets, respectively. By recalculating liver cancer risk for the selected group whose smoking status was changed as a model of quitting smoking, reductions in the liver cancer risk score as high as 78% were observed. Six of the 17 respondents (35%) reduced their risk category from high to medium.

Conclusion: A NN trained with personal health information can stratify a large population into risk categories and model a drop-in risk in response to modifications in life style such as quitting smoking.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email