[Paper] Enhancing Framingham Cardiovascular Risk Score Transparency through Logic-Based XAI

Source: arXiv - 2602.22149v1

Overview

The paper proposes a logic‑based explainable‑AI (XAI) layer for the classic Framingham Risk Score (FRS), turning a “black‑box” risk number into a transparent, actionable diagnosis. By expressing the FRS in first‑order logic, the authors can pinpoint the smallest set of patient attributes that justify a given risk category and suggest concrete lifestyle or clinical changes that would move the patient into a lower‑risk bucket.

Key Contributions

• Logical Formalisation of FRS: Re‑writes the entire Framingham algorithm as a set of first‑order logic rules, enabling formal reasoning about each decision.
• Minimal‑Sufficient Explanations: Introduces an algorithm that extracts the smallest subset of input features (e.g., age, cholesterol, smoking status) that alone guarantee the observed risk classification.
• Prescriptive “What‑If” Scenarios: Generates actionable counterfactuals showing which modifiable factors (e.g., blood pressure, smoking) need to change—and by how much—to lower the risk tier.
• Exhaustive Evaluation: Tests the explainer on every possible FRS input combination (≈22 k distinct profiles), confirming that explanations are sound and that suggested interventions are feasible.
• Open‑Source Prototype: Provides a lightweight Python implementation that can be plugged into existing clinical decision‑support pipelines.

Methodology

1. Rule Extraction: The authors translate the numeric FRS equations (points for age, cholesterol, etc.) into a deterministic set of logical predicates.
2. Explanation Engine: Using a SAT‑based minimal hitting‑set solver, the engine searches for the smallest group of predicates that, when true, force the overall risk classification.
3. Counterfactual Generation: For each identified minimal set, the system flips the values of modifiable predicates (e.g., “smoker → non‑smoker”) and recomputes the risk to see if the category drops.
4. Validation Loop: All 22,000+ possible input vectors are fed through the pipeline; the resulting explanations are compared against the original FRS calculation to ensure correctness.

Results & Findings

• Correctness: The logical explainer reproduced the exact FRS classification for every tested profile (100 % fidelity).
• Explanation Size: On average, only 3–4 attributes were needed to fully justify a risk tier, dramatically simplifying the clinician’s mental model.
• Actionable Insights: In >85 % of high‑risk cases, the system identified at least one modifiable factor (e.g., systolic BP, smoking, cholesterol) whose improvement would shift the patient to a lower risk category.
• Performance: The entire exhaustive test suite completed in under 2 minutes on a standard laptop, demonstrating that the approach is computationally lightweight enough for real‑time use.

Practical Implications

• Enhanced Trust: Clinicians can see why a patient is flagged as high risk, reducing reliance on opaque scores and encouraging adoption in primary‑care settings.
• Decision‑Support Integration: The explainer can be wrapped as a micro‑service that augments existing EHR‑embedded FRS calculators, delivering explanations alongside the risk number.
• Patient‑Facing Tools: Mobile health apps could surface personalized “risk‑reduction tips” derived from the counterfactuals, empowering patients to act on specific lifestyle changes.
• Resource‑Constrained Environments: In regions lacking cardiology specialists, the prescriptive output can guide community health workers to prioritize interventions (e.g., smoking cessation programs).
• Regulatory Alignment: By providing traceable, rule‑based rationales, the method helps satisfy emerging AI transparency requirements in healthcare (e.g., EU AI Act).

Limitations & Future Work

• Static Model Scope: The explainer is tied to the original Framingham equations; it does not yet handle newer, machine‑learning‑based CVD risk models that incorporate additional biomarkers or imaging data.
• Binary Counterfactuals: Current scenarios flip attributes to a single alternative (e.g., smoker → non‑smoker) without exploring graded changes (e.g., reducing systolic BP by 10 mmHg).
• Clinical Validation Needed: While the logical consistency is proven, prospective studies are required to confirm that the suggested interventions improve outcomes in real patient populations.
• Scalability to Multi‑Risk Scores: Extending the approach to simultaneously explain composite scores (e.g., combining Framingham with ASCVD or CHA₂DS₂‑VASc) is an open research direction.

Bottom line: By marrying classic epidemiological risk equations with modern XAI techniques, this work offers a practical pathway to make cardiovascular risk assessment both transparent and prescriptive—an advance that could accelerate the adoption of risk‑based prevention strategies across the healthcare ecosystem.

Authors

• Emannuel L. de A. Bezerra
• Luiz H. T. Viana
• Vinícius P. Chagas
• Diogo E. Rolim
• Thiago Alves Rocha
• Carlos H. L. Cavalcante

Paper Information

• arXiv ID: 2602.22149v1
• Categories: cs.LO, cs.AI
• Published: February 25, 2026
• PDF: Download PDF