Background
Central Nervous System (CNS) complications among people living with HIV (PWH) are common, but treatments remain elusive, likely due to underlying heterogeneity. Here, we aim to characterize neurobiological heterogeneity in PWH by leveraging machine learning and neuroimaging data in a large, well-characterized cohort.
Methods
We analyzed structural Magnetic Resonance Imaging (sMRI) data from 610 participants: 244 persons without HIV (PWoH) and 366 PWH. Freesurfer-derived region of interest (ROI) volumes were input to a machine learning technique termed HYDRA (Heterogeneity through Discriminative Analysis) to identify homogeneous patient subgroups. HYDRA clusters cases based on differences from controls, accounting for confounders such as age, sex, and race. The optimal clustering solution was determined by varying the number of clusters from 2 to 8 over 100 realizations of 10-fold cross-validation. The solution with the highest adjusted Rand index (ARI) was selected for further analysis. We then examined subtype differences in demographics, clinical measures, cognition, and imaging features (i.e., ROI volumes, resting state functional connectivity (RSFC), and cortical thickness). Demographic differences were assessed with analysis of variance and t-tests. Clinical, cognitive, and imaging differences were assessed using regression, controlling for age, sex, and race. Global Deficit Scores (GDS) were calculated by averaging cognitive domain deficit scores.
Results
We identified 3 reproducible subtypes (ARI=0.42): Cluster 1 (N=103; 78.6% male), Cluster 2 (N=93; 78.5% male) and Cluster 3 (N=170; 76.5% male). Cluster 1 displayed atrophy in the lateral occipital and paracentral regions, with increased corpus callosum volume (p<0.0001) compared to PWoH. Cluster 2 had widespread cortical volumetric increases, while Cluster 3 exhibited widespread cortical and subcortical atrophy and enlarged ventricles. Similar patterns were observed in cortical thickness. Cluster 3 had significantly reduced RSFC within the dorsal attention network compared to Cluster 2. Importantly, clusters differed in clinical and cognitive aspects. PWH in Cluster 2 had a significantly lower nadir CD4 count, while Cluster 3 had the highest viral load, lowest psychomotor speed, and highest (worst) GDS.
Conclusions
The identification of neurobiologically grounded biotypes of PWH may facilitate clinical trial enrichment and stratification, as well as uncover actionable features to guide development of targeted treatments.