Diagrammatization and Abduction to Improve AI Interpretability With Domain-Aligned Explanations for Medical Diagnosis

TL;DR

The paper tackles the interpretability gap in AI for high-stakes medical diagnosis by introducing DiagramNet, an ante-hoc model that uses diagrammatic murmur representations and abductive reasoning to align AI explanations with domain knowledge. It formalizes diagrammatization, abductive hypotheses, and an eight-stage architecture to produce perceptual and explained predictions for cardiac diagnoses from phonocardiograms, with murmur diagrams serving as the explanatory medium. Through demonstration, modeling, and qualitative user studies, the approach shows improved explanation faithfulness and predictive performance relative to baselines, and greater clinician trust for domain-aligned diagrammatic explanations over saliency maps. The work provides a framework for domain-aligned XAI in complex domains and demonstrates how explicit domain ontology and reasoning conventions can be embedded into AI to support human-AI collaboration and trust.

Abstract

Many visualizations have been developed for explainable AI (XAI), but they often require further reasoning by users to interpret. Investigating XAI for high-stakes medical diagnosis, we propose improving domain alignment with diagrammatic and abductive reasoning to reduce the interpretability gap. We developed DiagramNet to predict cardiac diagnoses from heart auscultation, select the best-fitting hypothesis based on criteria evaluation, and explain with clinically-relevant murmur diagrams. The ante-hoc interpretable model leverages domain-relevant ontology, representation, and reasoning process to increase trust in expert users. In modeling studies, we found that DiagramNet not only provides faithful murmur shape explanations, but also has better performance than baseline models. We demonstrate the interpretability and trustworthiness of diagrammatic, abductive explanations in a qualitative user study with medical students, showing that clinically-relevant, diagrammatic explanations are preferred over technical saliency map explanations. This work contributes insights into providing domain-aligned explanations for user-centric XAI in complex domains.

Figures (22)

• Figure 1: Reasoning processes between the user and AI. a) Current XAI: the user views an explanation based on deductive reasoning, and post-hoc has to make sense of it in context of the domain and evaluate its plausibility. This leaves an interpretability gap. b) Domain-aligned XAI: the ante-hoc interpretable AI encodes domain ontology and conventions to provide explanations diagrammatically to convey domain concepts and relational structures, and abductively to convey how it evaluates multiple hypotheses. The user simply verifies the AI explanation and prediction.
• Figure 2: Four steps of abductive reasoning demonstrated on a pedagogical scenario of inferring cats and dogs: I) observe the instance, II) hypothesize causes and retrieve associated rules that determine the consequent evidence (e.g., cat has vertical eye pupils and claws, dog has round pupils and nails), III) evaluate the evidence on the observation then deductively backchain on each rule to determine the plausibility of its hypothesis (e.g., Cheshire's vertical pupils and claws imply it is a cat, but no round pupils or nails does not imply it is a dog), and IV) resolve the best explanation with highest plausibility, i.e., that Cheshire is a cat and Doge is a dog. Image credits: "https://thenounproject.com/icon/cat-1736115", "https://thenounproject.com/icon/dog-1609923", "https://thenounproject.com/icon/paw-1738438", and "https://thenounproject.com/icon/paw-1738440" by https://thenounproject.com/maxim221 under the https://creativecommons.org/licenses/by/3.0/deed.en.
• Figure 3: Upper-left: Anatomy and physiology of the heart to see how blood flows through the left atrium and ventricle via the aortic and mitral valves, which make lub-dub sounds on their closure. Lower-left: Aortic valve with normal or aortic stenosis pathology showing calcification (in red) that leads to stiffness, resulting the a crescendo-decresendo murmur sound. See lilly2012pathophysiology for details on the physiological dynamics of other valvular diseases. Right: Example murmur diagrams showing typical murmurs for a) aortic stenosis (AS), b) mitral regurgitation (MR), c) mitral valve prolapse (MVP), d) mitral stenosis (MS), and their more severe variants (e--h) with slightly different shapes. Black rectangles indicate normal "lub" (S1) and "dub" (S2) sounds. Red areas indicate abnormal murmur sounds. See Figs. \ref{['fig:demo-explanations']}--\ref{['fig:demo-as']} for diagrams on real PCGs. Image credits: heart drawing adapted from https://commons.wikimedia.org/wiki/File:Diagram_of_the_human_heart_(valves_improved).svg by https://en.wikipedia.org/wiki/de:User:Ungebeten, and valve drawings adapted from https://commons.wikimedia.org/wiki/File:Aortic_valve_pathology_(CardioNetworks_ECHOpedia).svg. All images under the https://creativecommons.org/licenses/by-sa/3.0/deed.en.
• Figure 4: Abductive reasoning for cardiac diagnosis to I) observe an abnormal murmur, II) hypothesize possible diagnoses (N, AS, MR, MVP, MS) and retrieve corresponding rules that relate to evident symptoms of murmur heart phase and murmur shape, III) evaluate the coherence of all symptoms with respect to the observation and deduce via backward chaining on each hypothesis rule to determine the plausibility of each hypothesis, and IV) resolve to select the plausible explanation with simplest murmur shape function to infer the best fit to the observation (AS in this case). Although the murmur shapes for MVP and MS could fit the murmur equally well as that for AS, these two shapes are increasingly complex, and thus less preferred based on the principle of parsimony. False premises are indicated with a cross negation line, and less preferred premises with dashed lines.
• Figure 5: Modular architecture of DiagramNet with 7 stages corresponding to the steps of abductive reasoning (I to IV) in Section \ref{['subsection:abductive-reasoning']}, and the 8th stage for ensemble prediction to combine perceptual and abductive predictions. Black arrows indicate feedforward activations, the blue arrow indicates an iterative nonlinear optimization at inference time to estimate the final murmur shape parameters, and red downward arrows indicate which variables are trained with supervised labels. Variables are annotated with bold for vectors or tensors, $\hat{\space}$ for predictions from trained modules, $\Breve{\space}$ for heuristically-calculated values, and $\tilde{\space}$ for those optimized at inference time. $\circ$ is the Hardamard operator for element-wise multiplication. Narrow rectangles indicate an input or predicted variable. Other shapes indicate processes: trainable neural network blocks as rectangle $F_{y_0}$ or trapezoid $M_m$ (capital letters), non-trainable heuristic processes as rounded squares (script letters), and vector operators (circles).