ABSTRACT:
Universal health coverage (UHC) is a global health initiative captured under the Sustainable Development Goal 3 (SDG 3). Specifically, SDG 3.8 aims at attaining universal health coverage which includes access to quality essential health-care services, financial risk protection, and access to effective, quality, safe, and affordable vaccines and essential medicines for all. Among others, SDG 3 targets significant increase in health financing as well as places emphasis on recruiting, developing, training and retaining health workforce in developing countries, particularly in small island developing States and least developed countries. This is measured by health worker density and distribution as contained in SDG Indicator 3.c.1. In Nigeria, doctor to patient’s ratio is one doctor to six thousand patients as against World Health Organization (WHO) standard of 1 doctor to six hundred patients. Shortages in medical personnel can threaten the attainment of UHC. Partly responsible for the deficit is the poor student retention and progression in medical schools in developing economies. In this paper, we formulate the problem of high failure rate and low student progression as a machine learning problem using data from the College of Medicine of a Nigerian university and artificial neural networks as classifier. Our aim is to solve the problem using an artificial neural network-based learning analytics (ANN-based-LA) system. The intelligent system is trained using historical data of medical and dental students to understand students learning behaviour as well as predict academic performance. The factual and unbiased information generated will empower relevant stakeholders in medical education in formulating adaptive learning techniques that takes into cognisance the unique learning abilities of respective students with a view to promoting better learning experience, reducing student failure rate and enhancing student progression. Since challenges like high oscillations, presence of poor local optima, sparse gradient, and inappropriate (vanishing or extremely large) learning rate could trigger prolonged neural network training time and low convergence rate, we proposed a new stochastic gradient descent algorithm called Adum-Aiona. A modification of Adam – the popular algorithm for training neural networks, Adum-Aiona is designed to tackled afore-mentioned challenges. In future work, Adum-Aiona will be benchmarked against popular stochastic gradient descent algorithms like Adadelta, RMSprop, and Adam to ascertain gains it has made.
Keywords:
Adum-Aiona, Artificial Neural Networks, Stochastic Gradient Descent, Sustainable Development Goal 3, Universal Health Coverage
