Explainable Artificial Intelligence (XAI) in Medical Education: A Multi- Modal Framework for Enhancing Human-AI Collaboration

Technical Summary: Explainable Artificial Intelligence (XAI) in Medical Education

Problem Statement
The rapid integration of Artificial Intelligence (AI) into clinical diagnostics, particularly in radiography, pathology, and neuroimaging, faces a significant pedagogical barrier: the "black-box" nature of complex machine learning algorithms. While AI systems can match or exceed human diagnostic accuracy, their opacity undermines confidence, restricts interpretability, and hinders substantial pedagogical involvement in medical curricula. Medical students require not only the correct diagnosis but also an understanding of why an AI reaches a specific conclusion to develop robust clinical reasoning skills. Current literature lacks empirical research on XAI as a specific educational intervention to bridge this gap between algorithmic output and human understanding.

Methodology
This study employed a prospective, mixed-methods randomized controlled trial (RCT) with a two-phase sequential explanatory design involving 120 third-year medical students at the University of Liberia.

• Technical Foundation (Phase 1): The researchers developed the CerViD-MultiModal framework (Cognitive enhancement Random Forest with Visual Diagnostics). This system utilized neuroimaging data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (N=100), specifically focusing on fornix atrophy as an early indicator for Alzheimer's disease. The model used FreeSurfer v7.2.0 to segment the fornix and extract morphometric features (anterior/posterior volume, body sphericity, total volume). A Random Forest ensemble (100 decision trees) was trained to distinguish between Early Mild Cognitive Impairment (EMCI) and Late Mild Cognitive Impairment (LMCI).
• Educational Intervention (Phase 2): Participants were randomized into two groups:
  • Control Group (n=60): Received standard AI instruction via static text and typical bar charts.
  • XAI-Enhanced Group (n=60): Received the same core content but utilized the CerViD-MultiModal framework integrated with interactive SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) visualizations. These tools provided dynamic global feature rankings, patient-specific local explanations, and interactive force graphs.
• Assessment: Post-intervention, participants were evaluated using four standardized instruments: AI Literacy Score (0-100), System Usability Scale (SUS, 0-100), NASA Task Load Index (NASA-TLX, 0-100), and Confidence in AI Interpretation (1-5). Statistical analysis utilized independent samples t-tests and Cohen's d for effect sizes.

Key Contributions
The primary contribution of this work is the empirical validation of XAI as a dual-purpose tool: a mechanism for diagnostic transparency and a transformative pedagogical instrument. The study introduces the CerViD-MultiModal architecture as a concrete implementation method for integrating explainable machine learning into medical training, specifically demonstrating how visualizing model decision-making processes (via SHAP and LIME) can demystify AI for learners.

Results
The XAI-enhanced group demonstrated statistically significant improvements across all primary outcomes compared to the control group (p < 0.001):

• AI Literacy: Scores increased by 34.1% (85.0 ± 5.6 vs. 63.4 ± 6.8), with a large effect size (Cohen's d = 3.44).
• System Usability (SUS): Perceived usability improved by 66.4% (36.3 ± 4.3 vs. 21.8 ± 4.8), with Cohen's d = 3.15.
• Cognitive Load (NASA-TLX): Perceived workload was significantly reduced by 37.4% (26.5 ± 6.5 vs. 42.3 ± 7.1), with Cohen's d = -2.31.
• Confidence: Learner confidence in AI interpretation grew by 30.8% (3.4 ± 0.4 vs. 2.6 ± 0.4), with Cohen's d = 1.89.

Additionally, the underlying CerViD-MultiModal model achieved an accuracy of 81.2% (AUC = 0.85) in classifying EMCI vs. LMCI, with anterior fornix volume identified as the most relevant predictor (34%).

Significance and Claims
The authors claim that XAI operates as more than a mere algorithmic transparency mechanism; it functions as a "scaffolding" tool aligned with constructivist learning theories and cognitive load theory. By making expert reasoning apparent, XAI guides learners through the zone of proximal growth, converting opaque algorithms into understandable cognitive tools.

The study concludes that the integration of XAI significantly enhances human-AI collaboration in medical education. It asserts that by demystifying model decision-making, the CerViD-MultiModal framework optimizes diagnostic comprehension and user trust while mitigating the cognitive load typically associated with interpreting "black-box" AI. The authors position these findings as evidence that interpretable machine learning architectures can effectively address the traditional limitations of clinical AI in educational settings, though they acknowledge limitations regarding the single-site sample and the need for longitudinal follow-up to assess retention.