Abstract

Unfortunately, users find it challenging to comprehend how ML models generate predictions due to their intricate parameter sets. In this work, ML models that predicted HIV positive patients with viral rebound using clinical, demographic, laboratory, and healthcare facility data were developed. LIME (Local Interpretable Model-agnostic Explanations), which explained specific predictions provided by ML models was focused on. Ensemble learning techniques such as boosting, bagging, stacking and hard voting were applied to reduce errors, improve performance, and boost the overall robustness of predictions. While individual ML models may perform poorly due to their high variation or bias, their integration strengthens each other’s shortcomings and yields more precise predictions. Hard voting had the best performance on test data achieved with the f1-score of 0.72 and 0.6 recall which was high compared to other ensemble learning methods. Furthermore, the viral rebound feature interaction for patients was highlighted, which will help health workers in model debugging, building trust and validating predictions.

Find out more by clicking this Springer link : https://link.springer.com/chapter/10.1007/978-3-031-69197-3_31

DOI: 10.1007/978-3-031-69197-3_31

Contributors: Ssenoga Badru, Rose Nakibuule, Joyce Nakatumba-Nabende, Ggaliwango Marvin

Ssenoga Badru presenting at Uganda National Digital Health Conference at Serena Hotel, Kampala-Uganda.