Prediction is better than cure: how explainable AI can improve healthcare

Abstract

Machine learning methods have the potential to augment diagnostic capabilities in noninvasive screening techniques like ultrasound and microscopy, which often suffer from noise and lack specificity. This presentation showcases our team's efforts in developing tools for practical clinical applications in areas like sepsis, meningitis, malaria, depression and ageing. Utilizing explainable AI algorithms such as SHAP, LIME, and GradCAM, we prioritize transparency and bias identification in tool development. We provide examples illustrating how deep learning models, when applied to medical images, extend the clinician's vision 'beyond the expert human eye,' complementing doctors' expertise for improved diagnosis. Finally, we explore the transformative impact of foundational models like ChatGPT4 on future healthcare, discussing both limitations and opportunities of these emerging technologies.