An Interpretable Multi-Modal Diagnostic Framework for Parkinson’s Disease Using Auxiliary Kolmogorov–Arnold Networks and Probabilistic Decision Fusion

Abstract

Parkinson’s disease (PD) is a progressive neurodegenerative disease that causes a variety of both motor and non-motor impairments, creating a significant clinical challenge in terms of diagnosing the disorder early and reliably. Historically, doctors diagnosed PD using neurological assessments and rating scales based solely on subjective observations, which lack sensitivity for detecting early signs of PD. Over the past few years, there has been some success developing diagnostic systems based on machine learning, though these systems to date are almost entirely unimodal except for a few exceptions. Most current approaches do not allow for interpretability or accommodate the uncertainty associated with clinical data. In this paper, we propose a multi-modal diagnostic framework for PD that is interpretable and combines heterogeneous biomedical signals using deep representation learning (DRL), cross-modal modelling using functional relationships, and Bayesian probabilistic decision fusion (DPDF). Each of the modalities will first have its discriminative latent representations extracted using each modality's deep encoder. An Auxiliary Kolmogorov–Arnold Network (AKAN) will be used to model the non-linear functional dependencies between modalities. Finally, the Diagnostic Inference will be performed using DPDF to fuse modality-specific and cross-modal evidence while providing a measure of uncertainty for each diagnostic inference made. The research on the proposed methodology was evaluated via publicly available datasets from Kaggle and through the use of a subject-independent cross-validation scheme. The results obtained from the experimental test cases have shown that fusing multimodal data sources gives significantly better performance than using a single modal data source alone and results in substantially increased rates of accuracy, sensitivity and AUC. With all the capabilities outlined in this paper, our new approach represents a new state of the art for AI-supported diagnosis of PD. In particular, it blends deep learning methods, functional interpretive capabilities, and probabilistic reasoning into a single methodology that meets the requirements of a clinically relevant solution.