Identifying the characteristics of late HIV diagnosis using optimized machine learning algorithm

Background. Early detection of HIV infection is essential for clinical diagnosis, preventing transmission, and ensuring the safety of blood products. Individuals diagnosed late with HIV may unknowingly transmit the virus, and once diagnosed, they may experience worse health outcomes. Therefore, this study aims to identify the characteristics associated with late diagnosis of HIV patients. Materials and methods. In this retrospective cohort study, the information of 236 patients with HIV infection in Hamadan, the West of Iran, was collected by recording the CD4 count during 2011 to 2022 years. Late HIV diagnosis was considered with a CD4 ≤ 350/mm3. Initially, Extreme Gradient Boosting (XGBoost) and Random Forest (RF) algorithms identified important variables. Subsequently, models such as Logistic Model Tree (LMT), Classification and Regression Tree (CART), Deep Neural Network (DNN), and Support Vector Machine (SVM) were developed using a 70/30 training/test dataset split for clinical and demographic variables. Finally, the optimal model was selected based on accuracy and F1-score using Python software version 3.10. Results. The age, logarithm of Viral Load (LVL), Wight Blood Cell (WBC), Red Blood Cell (RBC), Lymphocyte (Lym), Hematocrit (Hct), Platelet (PLT), Hemoglobin (Hb), and clinical stage variables had relative importance above 6%. Among the developed models for the importance variables, the CART with F1-score and Accuracy values of 0.887 and 0.801 and 0.897 and 0.822 for training data, respectively. The AUC value obtained for the CART was equal to 0.918. Conclusions. Late diagnosis of HIV infection is a substantial problem, particularly in developing an algorithm that can accurately and interpretably detect disease characteristics, such as the CART, which could be essential for identifying characteristics that influence late HIV diagnosis and clinical and therapeutic decisions.