Evaluation of an Attention-Augmented Neuro-Fuzzy Framework for Multimodal Skin Disease Diagnosis: Performance, Robustness, and Clinical Implications

Abstract

Accurate diagnosis of skin diseases remains challenging due to overlapping visual features and the influence of clinical and environmental factors. The study evaluates the Dynamic Multi-Modal Feature Fusion with Attention-Augmented Adaptive Neuro-Fuzzy Inference System (DMF-ANFIS), a novel framework designed to enhance skin diseases classification, dermatological images, structured clinical attributes, and contextual variables. Using an experimental multimodal dataset comprising of 2,500 balanced cases and external validation on ISIC 2019, ISIC 2020, and HAM10000, DMF-ANFIS achieved 94.7% accuracy internally and 89–91% across benchmarks, outperforming baselines including KNN, SVM, Random Forest, standard ANFIS, and modern deep-learning models (EfficientNet-B0, ResNet-50, Vision Transformer). Robustness is validated under ablation studies, JPEG compression, motion blur, and subgroup analysis across Fitzpatrick skin types I–VI. Statistical reporting includes per-class metrics with 95% confidence intervals, calibration scores (ECE = 0.04, Brier = 0.07), and effect sizes. Interpretability is demonstrated via fuzzy rule bases, attention weight distributions, and case studies. Computational cost analysis shows feasibility for low-resource deployment (training time 2.3h, inference latency 120ms, memory footprint 480MB). This framework will establish DMF-ANFIS as a scalable, interpretable, and generalizable tool for dermatological decision support in telemedicine and climate-affected regions.