Codes for development of algorithms for prediction of viral load suppression in an HIV patients cohort in Conakry, Republic of Guinea

Viral load (VL) suppression represents a key to the end of the global HIV epidemic. It is critical for healthcare providers and people living with HIV (PLHIV) to be able to predict viral suppression. Machine learning (ML) is being used in the medical field, for example, to predict patient outcome. This study was conducted to explore the possibility of predicting viral suppression among HIV patients. Anonymized data were used from a cohort of HIV patients managed in eight health facilities supported by MSF in Conakry (Guinea). The data pre-processing steps included variable recoding, record removal, missing values imputation, grouping small categories, creating dummy variables and oversampling (only applied to the training set) of the smallest target class. Support vector machine (SVM), logistic regression (LR), naïve Bayes (NB), random forest (RF) and four stacked models where developed. The optimal parameters of the algorithms were determined with 03 cross-validation. The 30% of the sample was held as a test set to perform model evaluation. Techniques implemented to determine the most predictive variables were applied on LR, RF, and NB (with analysis of variance, ANOVA). LR was found to be the most optimal model to detect VL suppression and non-suppression. It produced an area under the curve (AUC) of 83%, with 74% and 78% sensitivity and specificity, respectively. In other words, it can correctly detect 74% of suppressed VL and correctly detect 78% of non-suppressed VL. With LR, Gender, Prior antiretroviral therapy (ART), Method into ART, Cotrimoxazole prophylactic therapy (CPT) at ART start, Second Line treatment, Last pre-ART CD4, Last ART CD4, Stage at ART start, Age, and Duration on ART were identified as the most predictive variables for VL suppression. This study demonstrated the capability to predict VL suppression but has some limitations. The results are dependent on the quality of the data and are specific to the Guinea context and thus, there may be limitations with generalizability. The preprocessing technics applied to the training set may have induced an increase in estimated performance of the models and which may also have been affected by their specific limitations. Future studies may be conducting a similar study in a different context and develop the most optimal model into an application that can be tested in a clinical context