Integrating Machine Learning Pipelines for Multimodal Biomarker Prediction in Alzheimer and Parkinson Disease: A Component of the Neurodiagnoses Framework
Authors
Affiliations (1)
Affiliations (1)
- University of Grenoble Alpes
Abstract
Alzheimers and Parkinsons diseases are age-related neurodegenerative diseases that often require invasive procedures for diagnosis. Traditional diagnostic methods may fail to capture the interplay between genetic, molecular, and neuroanatomical markers. This manuscript aims to develop interpretable machine learning models that can predict key biomarkers, such as pTau, tTau, A{beta} positivity, and motor symptom severity, using non-invasive data. Machine learning models (Random Forest, XGBoost) were trained using ADNI and PPMI baseline data. Using the APOE4 genotype, MRI volumes, cognitive scores, and demographics as inputs, SHAP was employed to enhance model interpretability. Models achieved AUCs of 0.859 (tTau) and 0.852 (pTau) with recall > 80%. The PD motor severity yielded an MAE of 5.72 and an R2 of 0.586. SHAP confirmed the contributions of APOE4 status, hippocampal atrophy, and dopaminergic asymmetries. The pipelines provide clinically meaningful predictions of biomarker status and motor symptoms, supporting interpretable, multi-axis neurodiagnostic tools within the neurodiagnoses framework.