Authors: N Jeevana Jyothi, M. Priyatharshini
Abstract: Combination of multimodal healthcare data such as genomic profiles, EHRs, medical imaging, and wearable sensor data brings in new opportunities for early disease prediction and precision medicine. Nevertheless, the complexity and black-box nature of state-of-the-art machine learning models bring many challenges towards their clinical deployment. This paper introduces a novel explainable artificial intelligence (XAI) architecture for early disease prediction using multi-modal healthcare analytics. The proposed framework combines various data streams using cross-modal generative transformers, interprets results by means of feature attribution and attention heatmaps computed using SHAP values, and utilizes federated learning techniques to preserve the privacy of the participating institutions. Evaluation using real-world multimodal datasets confirms high effectiveness of our solution with accuracy of 97% and AUC of 0.971, which is much better compared to unimodal and black box models.
DOI: http://doi.org/10.5281/zenodo.21338356
