School of Bio-Medical Sciences (Bio-Medical)

Permanent URI for this community

http://10.29.1.53:4000/handle/10570/6

Browse

Browsing School of Bio-Medical Sciences (Bio-Medical) by Author "Agasi, Herbert"
  • Item
    Predicting major depressive disorder among adults Living with HIV In Uganda: An Artificial Intelligence Approach
    (Makerere University, 2022-08-01) Agasi, Herbert
    Background Major depressive disorder is a leading cause of disability worldwide and a major contributor to the overall global burden of disease. The biggest percentage of the MDD burden pertains to low- and middle-income countries like Uganda. MDD is common among people living with HIV (PLWH) with prevalence ranging from 8% to 60%. In order to bridge the treatment gap for MDD, we need to develop locally relevant and accurate screening tools that will facilitate diagnosis, prevention and early treatment. In this era of personalised medicine with powerful computational tools, it is possible to develop predictive computational algorithms that will enable the early identification of individuals at risk for MDD. Methodology This descriptive nested case control study utilised samples and clinical data that were collected by Prof. Eugene Kinyanda’s Senior EDCTP Fellowship study (2011-2014). The EDCTP Mental Health study undertaken among 1,099 anti-retroviral therapy-naive PLWH. For this study, artificial intelligence was used to predict incident cases of major depressive disorder among adults living with HIV. Classification algorithms were applied and evaluated using classification matrices and area under the curve receiver operating characteristic curves. Results Multiple algorithms were used in the prediction of incident cases of MDD. Random Forest classifier and Xtreme Gradient Boost were the best performing modes. Voting classifier was used to compound the effect of the two algorithms. Area under the receiver operating characteristic curve scores were 0.945, 0.938 and 0.942 for Xtreme gradient Boost, Random Forest classifier and Voting Classifier algorithms respectively. Study site was the biggest predictor of the incident cases of MDD while gender and psychiatric history were also strong predictors of MDD. Conclusion This study demonstrates the applicability of machine learning approaches in the development of predictive models for MDD. The results are consistent with the literature on the predictors of incident MDD where study site, social support, psychiatric history and gender were found to be good predictors of MDD among the same study participants