Interpretable AI in Healthcare: Enhancing Fairness, Safety, and Trust

Samual James MacDonald, Kaiah Steven, Maciej Trzaskowski

Abstract: The value and future potentials of AI in healthcare are becoming self-evident, presenting an escalating body of evidence. However, the adoption into clinical practice is still significantly impacted by the lack of transparency, stemming from inadequate focus on human-comprehensive information from AI, i.e. Interpretable AI. AI interpretations of uncertainty, significance, and causality translate to a more fair, safe, and reliable AI. This is especially pertinent to safer clinical decision making and for minimising any risk to the patient. In this chapter we aim to elucidate what interpretability means and why most machine learning (i.e. AI) models fail to satisfy these definitions. We lay this phenomenon out through what we believe to be its canonical components: predictions, uncertainty, significance, and causality; explaining how these different types of interpretations can support various explanations and how overcoming this barrier can permit the adoption of AI into healthcare.