Application of Basic Epidemiologic Principles and Electronic Health Records in a Deep Learning Prediction Model —Reply

In Reply In our recently published study, we developed a risk prediction model without information on UV exposure or lesional character to predict new-onset nonmelanoma skin cancer in an Asian population. A deep learning convolutional neural network was created using nonimaging diagnostic and medication information to predict nonmelanoma skin cancer. The optimal area under the curve was 0.89. However, owing to random sampling, the group with skin cancer and the control group were not age matched. We agree that age was a dominating factor, but it might not be the most important one. By further building age and gender-matched models, the area under the curve decreases by 4% to 5%. Because this was not a study for independent risk factors, all of the variables, including age, sex, and continuous longitudinal medical data, were used to build the model. Previous studies with large-scale data sets using machine learning to predict hypertension or death during palliative care did not use matched controls either.
Source: JAMA Dermatology - Category: Dermatology Source Type: research