LLM Based XAI Framework for Neurofibromatosis Classification
IEEE Access, cilt.14, ss.66004-66020, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 14
- Basım Tarihi: 2026
- Doi Numarası: 10.1109/access.2026.3688620
- Dergi Adı: IEEE Access
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Directory of Open Access Journals
- Sayfa Sayıları: ss.66004-66020
- Anahtar Kelimeler: AI, LLM, machine learning, XAI
- Ankara Yıldırım Beyazıt Üniversitesi Adresli: Evet
Özet
Half of individuals with neurofibromatosis (NF) inherit the disorder from an affected parent, while the remaining cases arise from sporadic. Neurofibromatosis data is generally tabular that ML algorithms have shown successful results on tabular data. However, insufficient data is a significant factor affecting performance. There is an also another limitations as clinical interpretability—an essential requirement for trust, transparency, and ethical use in medical decision making. Therefore, the proposed frameworks bridges data augmentation, ML interpretability with natural language explanation. Extreme Gradient Boosting (XGBoost) emerges as the best-performing model, achieving approximately 0.80 accuracy and 0.80 F-score, substantially improving upon previously reported results on the same dataset. The integration of Explainable Artificial Intelligence (XAI) techniques such as SHAP (SHapley Additive exPlanations), LIME (Local Interpretable Model-agnostic Explanations), and PFI (Permutation Feature Importance) becomes crucial for uncovering the underlying relationships between genetic, phenotypic, and demographic features. Furthermore, integrating Large Language Models (LLMs) enables the automatic interpretation and textual summarization of XAI outputs, translating complex importance maps and feature interactions into human-readable clinical insights. These XAI artefacts are translated into concise clinical narratives by the Phi-3-mini LLM, and the resulting explanations are quantitatively assessed with global/local faithfulness, hallucination rate, and feature- and text-level consistency metrics, demonstrating high alignment between narrative and XAI evidence after prompt optimization. Consequently, the proposed framework delivers a transparent and auditable decision-support pipeline for health dataset and offers a reusable template for integrating XAI and LLM-based explanation in rare-disease tabular modeling. This hybrid framework bridges the gap between computational reasoning and medical understanding. This framework give us to knowledge-driven decision support system for NF diagnosis and risk assessment. The proposed approach not only improves model explainability but also promotes trustworthy AI practices in rare genetic disease research.