A New Approach to Discovering HIV Symptom and Patient Clusters Using CNICS Data and Topic Modeling
Keywords: Unsupervised Learning, Dimensionality Reduction, Latent Dirichlet Allocation, Survival Analysis, Symptom Index
TL;DR: This study leveraged and adapted LDA topic modeling to identify clinically meaningful and prognostic symptom clusters among people living with HIV.
Abstract: The identification of symptom patterns and assessment of their impacts on relevant health outcomes are important to symptom management and caring for people living with HIV (PWH). Research on HIV symptom clusters has been hampered by small sample sizes, conventional statistical methods not adequately capturing the intricate relationships among symptoms, and not associating symptom patterns with health outcomes. In this study, we proposed a new approach leveraging and adapting the Latent Dirichlet Allocation (LDA) topic modeling method to discover latent symptom clusters in one of the largest cohorts of PWH in the United States (US), sourced from the Centers for AIDS Research Network of Integrated Clinical Systems. Based on the reduced symptom space, patient clusters were then derived and analyzed for time to virological failure. The results showed that LDA outperformed traditional symptom clustering methods in identifying clinically meaningful symptom clusters. It included a novel systemic inflammatory response cluster among PWH in the US as a significant prognostic marker of virological failure. Moreover, the uncovered patient clusters were significantly distinguished in experiencing virological failure and could be characterized by distinct symptom clusters. The findings suggested a strong association between symptom patterns and subsequent virological failure among PWH. The study demonstrated the power of topic modeling as a new direction in symptom research to reveal complex symptom patterns, toward development of personalized symptom management and targeted interventions to improve the life span and quality of life in PWH.
Track: 11. General Track
Registration Id: YXNDK8FQCB8
Submission Number: 192
Loading