Knowledge-guided counterfactual explanations for diabetes risk decision support: A directional intervention taxonomy

Counterfactual (CF) explanations help patients and clinicians understand which features could be changed to lower a predicted diabetes-risk score, supporting risk communication in ML-based decision support. Existing CF methods, however, distinguish features only as mutable or immutable; they do not encode the direction of clinically appropriate intervention (for example, physical activity should always be recommended to increase, never to decrease), and observational health-survey data such as BRFSS contain selection biases that CF generators can inadvertently exploit, producing guideline-inconsistent recommendations such as "reduce healthcare access to lower predicted diabetes risk".

To produce counterfactual explanations that clinicians can review as candidate intervention recommendations aligned with clinical guidelines, by encoding clinical domain knowledge as a per-query constraint layer over an off-the-shelf CF generator. The approach is a practical alternative to structural causal modelling when SCMs are not identifiable from observational health-survey data.

A five-class intervention-direction taxonomy (immutable, monotonic_up, monotonic_down, bidirectional, conditional) is encoded as a lightweight per-feature knowledge tuple and translated per query by a rule-based reasoner into the DiCE constraint API (features_to_vary, permitted_range). Evaluation uses BRFSS 2021 (n=236,378) with an XGBoost classifier (test AUC =0.8233), 200 high-risk patients, six ablations, and external validation on BRFSS 2015 (n=253,680).

On a taxonomy-consistency measure of actionability, the per-query mode raises the score from 0.6655 to 0.9880 (+48.5% relative) by eliminating wrong-direction and immutable violations as defined by the taxonomy; validity, an independent metric defined on the classifier rather than on the taxonomy, also improves from 0.752 to 0.808 (+7.5% relative). A per-feature breakdown attributes approximately 64% of the eliminated wrong-direction violations to suppression of three healthcare-utilisation indicators and behavioural self-report features, and the remaining 36% to directional redirection. External validation on BRFSS 2015 confirms both classifier transfer (AUC 0.827, Brier 0.0975) and per-query mechanism transfer (wrong-direction violations 0.000, validity 0.829, validity gain +0.031).

A clinical intervention taxonomy translated per query into the constraint interface of an off-the-shelf CF generator closes a substantial fraction of the actionability gap in current CF tooling and makes generated counterfactuals reviewable by clinicians as candidate intervention recommendations consistent with clinical guideline directions. The contribution is methodological, and the actionability score is reported as a taxonomy-consistency proxy rather than a clinically validated endpoint. The counterfactuals are intended as decision-support candidates for clinician review, not as personalised medical advice. Implications are discussed for future clinician-facing risk-communication tools and EHR-integrated decision-support research in chronic-disease screening and prevention, subject to clinician review and deployment-population validation.