The predictive accuracy of machine learning for the risk of death in HIV patients: a systematic review and meta-analysis

Abstract

Abstract Background Early prediction of mortality in individuals with HIV (PWH) has perpetually posed a formidable challenge. With the widespread integration of machine learning into clinical practice, some researchers endeavor to formulate models predicting the mortality risk for PWH. Nevertheless, the diverse timeframes of mortality among PWH and the potential multitude of modeling variables have cast doubt on the efficacy of the current predictive model for HIV-related deaths. To address this, we undertook a systematic review and meta-analysis, aiming to comprehensively assess the utilization of machine learning in the early prediction of HIV-related deaths and furnish evidence-based support for the advancement of artificial intelligence in this domain. Methods We systematically combed through the PubMed, Cochrane, Embase, and Web of Science databases on November 25, 2023. To evaluate the bias risk in the original studies included, we employed the Predictive Model Bias Risk Assessment Tool (PROBAST). During the meta-analysis, we conducted subgroup analysis based on survival and non-survival models. Additionally, we utilized meta-regression to explore the influence of death time on the predictive value of the model for HIV-related deaths. Results After our comprehensive review, we analyzed a total of 24 pieces of literature, encompassing data from 401,389 individuals diagnosed with HIV. Within this dataset, 23 articles specifically delved into deaths during long-term follow-ups outside hospital settings. The machine learning models applied for predicting these deaths comprised survival models (COX regression) and other non-survival models. The outcomes of the meta-analysis unveiled that within the training set, the c-index for predicting deaths among people with HIV (PWH) using predictive models stands at 0.83 (95% CI: 0.75–0.91). In the validation set, the c-index is slightly lower at 0.81 (95% CI: 0.78–0.85). Notably, the meta-regression analysis demonstrated that neither follow-up time nor the occurrence of death events significantly impacted the performance of the machine learning models. Conclusions The study suggests that machine learning is a viable approach for developing non-time-based predictions regarding HIV deaths. Nevertheless, the limited inclusion of original studies necessitates additional multicenter studies for thorough validation.

Keywords

• HIV
• Predictive model
• Risk of death
• Systematic review
• Meta-analysis