JMIR Publications

ABSTRACT

Background:

Hepatocellular carcinoma (HCC) usually known as hepatoma which is the third leading cause of cancer mortality globally. Since the progression of HCC is preventable.

Objective:

To develop a deep learning model, utilizing the trend and severity of each medical event from the electronic health record (EHR), to accurately predict HCC patient’s one-year earlier.

Methods:

HCC patients were selected from the National Health Insurance Research Database (NHIRD) between 1999 and 2013. To be included, all the individuals had to register as cancer patients in the catastrophic illness file and to diagnosis as a cancer patient in an inpatient admission. Controls (Non-HCC patients) were randomly selected from the same database. We used age, gender, diagnosis code, medication code, and time information as the input variables of the convolution neural network (CNN) model to predict HCC patients. We also calculate odd ratios to understand the relationship between potential variables and HCC risk.

Results:

We included 47,945 individuals, of whom 9,553 patients had HCC. The area under the receiver operating curve (AUROC) of predicting HCC risk one year earlier was 0.94 (95%CI: 0.937-0.943) with sensitivity 0.869 and specificity 0.865. The AUROC for predicting HCC patients 7 days, 6 months, 1 year, 2 years, and 3 years earlier were 0.96, 0.94, 0.94, 0.91, and 0.91, respectively.

Conclusions:

The findings of this study show that the CNN model has immense potential to predict the risk of HCC one-year earlier with minimal features available in the EHRs.