Predictive variables for peripheral neuropathy in treated HIV type 1 infection revealed by machine learning.

Abstract

Peripheral neuropathies (PNPs) in HIV-infected patients are highly debilitating because of neuropathic pain and physical disabilities. We defined prevalence and associated predictive variables for PNP subtypes in a cohort of persons living with HIV. Adult persons living with HIV in clinical care were recruited to a longitudinal study examining neurological complications. Each patient was assessed for symptoms and signs of PNP with demographic, laboratory, and clinical variables. Univariate, multiple logistic regression and machine learning analyses were performed by comparing patients with and without PNP. Three patient groups were identified: PNP (n = 111) that included HIV-associated distal sensory polyneuropathy (n = 90) or mononeuropathy (n = 21), and non-neuropathy (n = 408). Univariate analyses showed multiple variables differed significantly between the non-neuropathy and PNP groups including age, estimated HIV type 1 (HIV-1) duration, education, employment, neuropathic pain, peak viral load, polypharmacy, diabetes, cardiovascular disorders, AIDS, and prior neurotoxic nucleoside antiretroviral drug exposure. Classification algorithms distinguished those with PNP, all with area under the receiver operating characteristic curve values of more than 0.80. Random forest models showed greater accuracy and area under the receiver operating characteristic curve values compared with the multiple logistic regression analysis. Relative importance plots showed that the foremost predictive variables of PNP were HIV-1 duration, peak plasma viral load, age, and low CD4+ T-cell levels. PNP in HIV-1 infection remains common affecting 21.4% of patients in care. Machine-learning models uncovered variables related to PNP that were undetected by conventional analyses, emphasizing the importance of statistical algorithmic approaches to understanding complex neurological syndromes.