It is unclear whether data-driven machine learning models, which are trained on large epidemiological cohorts, may improve prediction of comorbidities in people living with human immunodeficiency virus (HIV).

In this proof-of-concept study, we included people living with HIV in the prospective Swiss HIV Cohort Study with a first estimated glomerular filtration rate (eGFR) >60 mL/minute/1.73 m2 after 1 January 2002. Our primary outcome was chronic kidney disease (CKD)-defined as confirmed decrease in eGFR ≤60 mL/minute/1.73 m2 over 3 months apart. We split the cohort data into a training set (80%), validation set (10%), and test set (10%), stratified for CKD status and follow-up length.

Of 12 761 eligible individuals (median baseline eGFR, 103 mL/minute/1.73 m2), 1192 (9%) developed a CKD after a median of 8 years. We used 64 static and 502 time-changing variables: Across prediction horizons and algorithms and in contrast to expert-based standard models, most machine learning models achieved state-of-the-art predictive performances with areas under the receiver operating characteristic curve and precision recall curve ranging from 0.926 to 0.996 and from 0.631 to 0.956, respectively.

In people living with HIV, we observed state-of-the-art performances in forecasting individual CKD onsets with different machine learning algorithms.