Machine learning models reveal neurocognitive impairment type and prevalence are associated with distinct variables in HIV/AIDS

Abstract

Neurocognitive impairment (NCI) among HIV-infected patients is heterogeneous in its reported presentations and frequencies. To determine the prevalence of NCI and its associated subtypes as well as predictive variables, we investigated patients with HIV/AIDS receiving universal health care. Recruited adult HIV-infected subjects underwent a neuropsychological (NP) test battery with established normative (sex-, age-, and education-matched) values together with assessment of their demographic and clinical variables. Three patient groups were identified including neurocognitively normal (NN, n = 246), HIV-associated neurocognitive disorders (HAND, n = 78), and neurocognitively impaired-other disorders (NCI-OD, n = 46). Univariate, multiple logistic regression and machine learning analyses were applied. Univariate analyses showed variables differed significantly between groups including birth continent, quality of life, substance use, and PHQ-9. Multiple logistic regression models revealed groups again differed significantly for substance use, PHQ-9 score, VACS index, and head injury. Random forest (RF) models disclosed that classification algorithms distinguished HAND from NN and NCI-OD from NN with area under the curve (AUC) values of 0.87 and 0.77, respectively. Relative importance plots derived from the RF model exhibited distinct variable rankings that were predictive of NCI status for both NN versus HAND and NN versus NCI-OD comparisons. Thus, NCI was frequently detected (33.5%) although HAND prevalence (21%) was lower than in several earlier reports underscoring the potential contribution of other factors to NCI. Machine learning models uncovered variables related to individual NCI types that were not identified by univariate or multiple logistic regression analyses, highlighting the value of other approaches to understanding NCI in HIV/AIDS.