PP01-09: Depression Among People Living With HIV/AIDS: Real-World Application of Deep Neural Networks
Gilead Sciences United States
The objective of this experiment is to demonstrate that deep neural networks with multi-autoencoders and skip-connection architectures can help identify those prone to depression among people living with HIV/AIDS.
For this study we use health claims data of people living with HIV/AIDS and identified those suffering from depression. We researched and selected two very specific architectures to solve the intermediate problem of sparseness in the data and the information loss in deep neural networks.
Among the patients living with HIV/AIDS, sourced from the patient claims database, those suffering from depression were identified based on the relevant ICD codes and associated medications. For the control group we selected patients who were also living with HIV/AIDS and were administered with similar medications. The prepared cohort was validated to have similar age distribution across the target and control group. However, patients in control group were never diagnosed with depression. A set of predictors used for the analysis belonged to the following broad categories: Patient demographics, disease codes (ICD 9/10), General Product Identifier (GPI), procedure codes and patient characteristics. In total there were 181 predictors and approximately 180k records for the study, out of which the prevalence of depression was 8%. The preliminary results of the experiments indicate that the deep neural networks can be used to predict the occurrence of depression among people living with HIV/AIDS. Multiple modeling experiments present the effect of shallow vs deep neural network and how the use of auto-encoders and skip-connections can improve upon the performance efficiency of the model while retaining the prediction accuracy. So far, we use the evaluation metric of AUROC and have the best value of 0.77 compared to baseline of Logistic regression with AUROC as 0.75.
There is a rising trend of people living with HIV/AIDS getting depressed which effects them adversely as it increases their likelihood of worse adherence to medication. Finding innovative ways to proactively identify those who are prone to depression while under HIV/AIDS medication, can help the physicians to do a timely intervention. Deep learning is a very promising approach which has solved some of the complex problems in real world. This experiment aims at presenting the effectiveness of custom deep learning architectures to predict depression and the people living with HIV/AIDS. The key contribution of this work is to highlight that with such custom architectures, we can handle the data sparsity problems and further research in representation learning of the medical codes can improve the prediction of disease progression in patients.