Assisting clinical diagnosis with interpretable fuzzy probabilistic modelling

The need for transparency and interpretability is a fundamental theme to be addressed by Artificial Intelligence (AI) research, especially in high-risk applications such as healthcare. In this work, we propose Fuzzy Sets in Probability Trees (FPT), a novel method that combines probabilistic trees and fuzzy logic. This approach is fully interpretable, providing clinicians with a tool generate and verify the entire clinical decision process. FPT extends the existing framework of Probabilistic Decision Trees by incorporating the uncertainty in the data, allowing for a flexible description of vague variables. Thus, FPTs enable the incorporation of domain knowledge in the form of fuzzy membership functions within the framework of probabilistic trees. Furthermore, FPTs can represent circumstances or explanations that cannot be represented with other techniques (e.g., Bayesian networks), paving the way to a novel form of interpretable AI that allows clinicians to generate, control and verify the entire diagnosis procedure; one of the strengths of our methodology is the capability to decrease the frequency of misdiagnoses by providing an estimate of uncertainties and counterfactuals. We applied FPT to two real medical scenarios: classifying malignant thyroid nodules, and predicting the risk of progression in chronic kidney disease patients. Our results show that FPTs can provide interpretable support to clinicians. We also show that FPT and its predictions can assist clinical practice in an intuitive manner, with the use of a user-friendly interface specifically designed for this purpose. The integration of probabilistic trees and fuzzy reasoning preserves the nuances that are generally lost in (probabilistic) decision trees due to the adoption of crisp thresholds, leading to hybrid trees that provide an AI system better aligned with human reasoning processes and that can effectively support clinicians in the diagnosis decision process.