A machine learning approach for predicting probability of death or disease progression in an early-treated pediatric African cohort
Authors: Domínguez-Rodríguez S, Tagarro A, Serna Pascual M, Otwombe K, Violari A, Fernández S, Nhampossa T, Lain M, Vaz P , Behuhuma NO, Danaviah S, Dobbels E, Barnabas S, Cotugno N, Zangari P, Palma P, Oletto A, Nardone A, Nastouli E, Spyer M, Kuhn L, Rossi P, Giaquinto C, Rojo P on behalf of EPIICAL consortium.
Published in: 23rd international AIDS Conference, July 6th-10th 2020
Background In perinatally HIV infected children, mortality and morbidity are highest in the first months after ART initiation and is linked to advanced disease and late diagnosis. The random forest approach can deal with more predictors than classical models and has no model assumptions such as normality, linearity or hazard proportionality. The aim of this study was to predict the probability of death or clinical progression at a specific time of follow-up.
Methods EARTH (EPIICAL consortium) is an African multi-centre cohort enrolling HIV-infected infants treated within 3 months of life (n=151). A total of 134 infants with >1 follow-up visit were included in this analysis. The primary endpoint was the right-censored time to death or progression to AIDS. To predict the outcome, a log-rank random survival forest with imbalance correction was performed in a training subset (n=95, 70%). The algorithm was validated on the remaining 30% (n=39).
Results A total of 22 infants reached the primary endpoint with 13 (10%) patients dead and 9 (7%) with an AIDS defining condition. A total of 10000 trees were built with an error rate of 20%. The most important predictors of reaching the primary endpoint were baseline HIV viral load, age at diagnosis, weight-for-age, gender, age at ART initiation, and baseline CD4 count. In the validation, the model predicted a higher probability of reaching the primary endpoint among children who did indeed die or progress to AIDS, as compared to the group of children who did well (1-month: 14% vs. 0.01%, p-value=0.045; 6-months: 62% vs. 0.03%, p-value=0.019; 12-months: 76% vs. 16%, p-value=0.012). The AUC for predicting survival or progression was 0.83, 0.84, and 0.72 for 1-month, 6-months, and 1-year respectively.
Conclusions This model helps clinicians individualize the probability of death or progression to AIDS at time of diagnosis and may be useful for the early identification of high-risk patients.