Machine learning algorithm to predict >5% weight gain in PWH switching to INSTI

Background: Weight gain (WG) is a well-described phenomenon in PWH starting or switching ART. Machine learning (ML) methods is a tool of P4 medicine (Predictive, Preventive, Personalized & Participatory) and can generate models to identify patients at risk of WG. The objective was to develop an ML algorithm that predicts a 9-month WG≥5% in PLWH switching to InSTI with/without TAF. Methods: This was an observational study that comprised ART-experienced PWH attending Modena HIV metabolic clinic from 2004 to 2020. The patients' medical, HIV and ART data were partitioned in an 80/20 training/test set to generate predictive models. A ML model was used to leverage a hybrid approach where clinical expertise is applied along with data-driven analysis. The study outcome was the prediction at 9 months of weight change with a cut of 5%: at any patient visit (model 1) and in the subset of PWH switching to InSTI with/without TAF (model 2). 9-month prediction was chosen as being the minimum time occurring between any two given visits in the 95% of the cases. A robust implementation of linear regressor algorithms were able to predict weight gain/loss while tolerating missing data. Intelligible explanations were obtained through Shapley Additive exPlanations values (SHAP), which quantified the positive or negative impact of each variable included in each model on the predicted outcome. A measure of effectiveness (E-measure) was chosen as a performance metric, because unlike accuracy it can penalize errors, particularly underestimation ones. Results: A total of 2817 patients contributed to generate 10877 observations, which allowed construction of 2 predictive models based on 44-variables including anthropometric, HIV and laboratory biomarkers. At last observation median age was 51 years (IQR 11); 70% were male. Median CD4 nadir was 200 cells/μL (IQR 217), current CD4 was 659 cells/μL (IQR 372), 97% had undetectable VL and time since HIV diagnosis was 20 years (IQR 13). Median BMI was 23.4 (IQR 4.5) and 5.8% had obesity. The highest ranked variables used to train the models were weight at time of prediction and the ones depicted in the figure. Model 1 had accuracy of 84.4% and 83.9% E-measure; model 2 had accuracy of 84.4% and 86.4% E-measure. Conclusion: We developed a ML tool with a remarkable E-measure that may assist clinicians in decision-making and shift HIV care towards a P4 medicine. Immune-metabolic variables were more relevant than ART switching in the prediction of WG.