Empiric neurocognitive performance profile discovery and interpretation in HIV infection.

The measurement and determinants of HIV-associated neurocognitive disorders (HAND) are under intense debate. We used latent profile analysis (LPA) and machine learning to define neurocognitive performance profiles and identify their associated risk factors in HIV patients receiving antiretroviral therapy (ART). Neurocognitive performance was assessed by a multidomain neuropsychological test battery. LPA was used to define individual neurocognitive profiles. Random forest analyses (RFA) identified the most important factors distinguishing each profile. Three profiles emerged from the LPA: profile 1 (P1, n = 159) achieved the highest performance, while profile 2 (P2, n = 163) had lowered executive functions and verbal memory, and profile 3 (P3, n = 59) was globally impaired. RFA achieved good prediction (area under the curve ≥ 0.80) only for global impairment (P3). Non-North American descent was the dominant predictor of P3, followed by factors coinciding with non-North American descent (female sex and toxoplasma seropositivity). Additional predictors included unemployment, current depressive symptoms, lower nadir CD4, and longstanding HIV. Restricting analyses to North Americans pointed to the additional importance of ART achieving high CSF levels and older age in prediction of P3. HAND diagnoses were most common in the globally impaired profile (P3 = 89.8%), followed by the group with reduced higher-order neurocognitive performance (P2 = 16.6%). Thus, implementation of LPA and RFA empirically distinguished three distinct neurocognitive performance profiles in this HIV-infected cohort while also highlighting potential risk factors and their relative importance to neurocognitive impairment. These data-driven analytical methods pointed to discernible demographic, HIV- and treatment-related risk factor constellations in patients born outside and within North America that might influence diagnostic and therapeutic decisions.