Domain-specific Explainable AI Solutions for Healthcare

Ravisankar PRIYADHARSINI, Guru Subbiah DEVI LAKSHMI, Jayavarthanaraj Soundharavalli MALAVIKA

Abstract

Explainable AI (XAI) has emerged as a pivotal solution for addressing different concerns such as lack of transparency, interpretability, trust, accountability, and clinical decision-making., particularly in the healthcare sector. XAI strives to bridge the gap between AI’s high-performance predictive capabilities and the need for understandable, trustworthy, and accountable AI systems. The dataset selected for this study includes chest X-ray images of patients affected by Covid-19 along with the corresponding patient metadata. The primary goal of constructing the proposed system is twofold: firstly, to determine whether an individual is affected by Covid-19 or pneumonia and secondly, to provide accountability for the AI models responsible for such classification decisions. The Deep Learning architectures, namely CNN+LSTM, EfficientNetB0, and ResNet18 are employed for classification tasks and the CNN+LSTM model outperforms the other two models, achieving a training accuracy of 99.42% and a testing accuracy of 83.66%. The preprocessed metadata and image-derived features were provided as inputs to the XGBoost and Random Forest classifiers. XGBoost and Random Forest achieved a validation accuracy of 98.69%, while their training accuracy reached 99.26% and 99.67%, respectively. The explainability of the deep learning models is visually illustrated by using the XAI technique Grad-CAM. Similarly, the machine learning models are interpreted by using two XAI techniques: SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME). This comprehensive approach combines accurate prediction with an improved transparency, enabling the stakeholders to better understand the underlying decision-making process through multimodal data fusion.

Keywords

EfficientNetB0, Explainable AI, CNN+LSTM, Grad-CAM, LIME, Random Forest Classifier, ResNet18, SHAP, XGBoost Classifier.

View full article