Design and Multi-level Evaluation of MAP-X: a Medically Aligned, Patient-Centered AI Explanation System

Paper Title

Design and Multi-level Evaluation of MAP-X: a Medically Aligned, Patient-Centered AI Explanation System

Publication Info

• Topic area: Patient-centered AI explanations in healthcare.
• Keywords: Explainable AI, healthcare AI, patient-centered design, large language models, retrieval-augmented generation, post-stroke speech assessment, clinical validation, multi-level evaluation, trust calibration.

Background and Problem

• Problem / challenge: Many healthcare AI systems lack tailored explanations for patients, leading to limited transparency, trust, and adoption. Current evaluation frameworks fail to holistically validate explanations across functional, clinician, and patient levels.
• Significance: Transparent, patient-centered explanations are critical in high-stakes clinical settings to enhance understanding, trust, and shared decision-making, particularly for complex conditions like post-stroke speech disorders.
• Motivation and related work: Prior research highlights the importance of grounding explanations in clinical reasoning and aligning them with patient needs. However, existing systems often fail to balance clinical validity and patient comprehensibility. Multi-level evaluation frameworks are rare, and most studies focus narrowly on accuracy or anecdotal evidence.

Solution

• Proposed approach: MAP-X (Medically Aligned, Patient-centered eXplanation), a system that uses retrieval-augmented generation (RAG) and large language models (LLMs) to generate layered, patient-facing explanations grounded in clinical evidence.
• Novelty:
  1. A Structure–Signal–Source blueprint for designing and validating explanations across clinician-defined structure, model reasoning signal, and clinical data source.
  2. Integration of RAG to reduce hallucinations and ensure explanations are grounded in clinician-authored evidence.
  3. Progressive disclosure interface for layered explanations tailored to patient comprehension.
  4. Multi-phase evaluation framework addressing functional, clinician, and patient perspectives.
• Procedure and key techniques:
  • Prediction Module: Extracts acoustic features from speech tasks and predicts severity using ensemble models with SHAP values for interpretability.
  • Retrieval Module: Matches patient data to clinically validated reference cases using vector search.
  • Generation Module: Synthesizes retrieved evidence and model outputs into structured explanations using GPT-4.
  • Interface Design: Presents explanations through dashboards, feature-specific insights, and conversational summaries, employing progressive disclosure to manage cognitive load.

Results

• Concrete findings:
  • Functional evaluation: Macro F1-score of 84.98% for severity classification, subsystem rating MAE of 0.81, and text quality scores of 75.44 (internal relevance) and 71.78 (external accuracy).
  • Clinician evaluation: High ratings for fluency (4.62/5), relevance (4.49/5), and coherence (4.74/5), with partial agreement (2.89/5) in narrative style consistency.
  • Patient evaluation: Trust scores significantly higher for MAP-X (6.39/7) compared to non-explainable control (5.44/7), with a positive trend in explanation satisfaction.
• Advantage over baselines: MAP-X explanations improved trust and understanding compared to non-explainable controls, while clinicians valued its structured, evidence-grounded outputs as collaborative tools.
• Experiments / evaluation:
  • Phase 1: Functional evaluation of faithfulness using clinician-defined metrics and rubric-based text scoring.
  • Phase 2: Application-grounded evaluation with 10 speech-language pathologists assessing clinical relevance and workflow fit.
  • Phase 3: End-user evaluation with 15 post-stroke patients comparing MAP-X explanations to non-explainable controls.
• Limitations and future work:
  • Risk of patient over-trust; need for harm-mitigation protocols.
  • Limited dataset size and reliance on audio-based perceptual ratings.
  • Generalizability restricted to South Korea and chronic stroke survivors.
  • Lack of empirical testing of clinician-mediated workflows and longitudinal impact.

Summary

MAP-X is an AI system designed to generate medically aligned, patient-centered explanations for post-stroke speech assessments. Using a Structure–Signal–Source blueprint and RAG, it produces layered, transparent explanations that align with clinical reasoning and patient needs. A three-phase evaluation demonstrated functional faithfulness, clinician acceptance as a collaborative tool, and improved patient trust and understanding. Limitations include risks of over-trust, dataset constraints, and restricted generalizability. Future work should focus on adaptive explanations, harm mitigation, and longitudinal studies to assess real-world impact.