Abstract
Parkinson's disease (PD) is a progressive neurodegenerative disorder characterized by a wide spectrum of motor and non-motor symptoms, often leading to delayed or inaccurate diagnosis. Conventional diagnostic methods frequently suffer from limited sensitivity, scalability, and interpretability, thereby restricting their utility in clinical settings. To address these limitations, this study presents a novel AI-driven diagnostic framework that integrates multimodal data fusion, deep learning-based classification, and generative language modeling to improve diagnostic accuracy and enable personalized reporting. The proposed framework leverages the Parkinson's Progression Marker Initiative (PPMI) dataset, incorporating structural Magnetic resonance imaging (MRI), Single-Photon Emission Computed Tomography (SPECT) imaging, cerebrospinal fluid (CSF) biomarkers, and clinical assessments. Statistical analysis was employed to select 14 key biomarkers-including dopamine transporter SBR values and CSF protein levels-from a total of 21 features identified as clinically relevant. A 1D Convolutional Neural Network (1D-CNN) was developed and trained using 121 engineered features, comprising radiomic descriptors and biologically derived metrics. Preprocessing and extensive feature engineering were conducted prior to a 70:30 train-test split, with data augmentation applied to the training set to enhance model generalization. The classifier achieved an accuracy of 93.7%, surpassing baseline approaches and emphasizing the value of domain-informed feature design. To improve interpretability and clinician usability, a Mini ChatGPT-4.0 Large Language Model (LLM) was fine-tuned using approximately 1,000 domain-specific prompt-response pairs generated from literature, classifier-derived eXplainable AI (XAI) feature scores, and expert annotations. The generated responses were evaluated using a custom scoring metric (0.0-5.0) based on their semantic alignment with ground truth completions. This LLM module produces patient-specific diagnostic summaries and treatment suggestions. Additionally, a cloud-based interface was developed to facilitate real-time MRI uploads, automated inference, and chatbot-driven consultations. Overall, the framework demonstrates high diagnostic performance, transparency, and user accessibility, offering significant potential for real-world clinical deployment in PD diagnosis and decision support.