An Explainable Artificial Intelligence Framework for Aquaphotomics-Based Skin Data Analytics: Integrating SHAP, LIME, LightGBM, ELI5, PDPbox, and Skater for Robust Dataset Categorization and Interpretable Process Modeling

Abstract

Aquaphotomics-based skin data analytics represents an emerging interdisciplinary domain that integrates spectroscopy, biological water–molecule I nteractions, and machine learning for non-invasive skin condition assessment. However, the complexity and high dimensionality of spectral skin datasets introduce significant interpretability challenges, particularly when advanced ensemble models such as LightGBM are used for predictive tasks. This study proposes a comprehensive Explainable Artificial Intelligence (XAI) framework designed to enhance transparency, interpretability, and reliability in aquaphotomics-driven skin data classification and process modeling. The framework integrates multiple post-hoc interpretability methods including SHAP, LIME, ELI5, PDPbox, and Skater to ensure multi-layered explanation coverage at both global and local levels.
The methodological foundation of this work is grounded in modern XAI paradigms, emphasizing interpretability as a core requirement for trustworthy AI systems in biomedical analytics (Linardatos et al., 2020). By combining gradient boosting-based prediction (LightGBM) with multiple explainability layers, the proposed system enables robust dataset categorization while preserving model transparency. Comparative insights from existing literature highlight that single-method interpretability approaches are insufficient for high-stakes domains such as biomedical signal interpretation and skin diagnostics (Ali et al., 2023; Chamola et al., 2023).
Experimental simulation demonstrates that multi-tool XAI integration improves feature attribution consistency, reduces interpretability ambiguity, and enhances decision-level trust. The study further identifies key aquaphotomics spectral bands as dominant explanatory features influencing skin state classification. The findings suggest that hybrid XAI architectures significantly outperform isolated interpretability techniques in both stability and usability.