INTELLIGENT CLINICAL DECISION SUPPORT FOR HEART DISEASE DIAGNOSIS USING MACHINE LEARNING IN NIGERIAN HOSPITALS

There is a cardiac crisis in Nigeria - less than 100 cardiologists for a population of greater than 220 million and more than 200 000 out of every 300 000 people in the country die from cardiovascular disease. Although progress has been made in cardiovascular AI research in the past decade, there is currently no machine learning model that can diagnose heart disease from data of Nigerian patients. This paper outlines an ML-CDSS that is specifically targeted to the Nigerian tertiary care hospitals. The CDSS has six main structured clinical features and raw 12-lead electrocardiogram (ECG) waveforms and adopts a multi-modal stacked ensemble approach with XGBoost, feedforward neural network, as well as a CNN-LSTM deep learning sub-model. In addition to each prediction, TreeSHAP provides an explanation layer for the tree-based approach, which is easy to understand, and DeepSHAP yields an explanation layer for the neural network; LIME is used to pass on secondary explanations, and GradCAM to visualise the activations for the ECG waveform. The process is done through five phases, HEART (Harvest, Engineer, Architect, Refine, and Transform) including prospective data collection from three Nigerian federal tertiary care hospitals, pre-processing, modelling building, validation with a holdout set blinded data, and deployment. Architecture designed to run intermittent connections, commodity compute, energy constrained. Since the inception of this project, there have been implicit collaboration between governance and NAFDAC Software to develop a Medical Device guide and Nigeria Data Protection Regulation guidelines. This paper tackles seven known gaps within the cardiac AI literature in Africa head-on.