Parkinson’s Disease (PD) is a progressive neurodegenerative disorder that impairs both motor and cognitive functions. Accurate detection of PD remains a major challenge, particularly at early stages when clinical symptoms are subtle. This study presents the first multimodal machine learning framework integrating functional near-infrared spectroscopy (fNIRS) and inertial measurement unit (IMU) data for early-stage PD detection during dual-task mobility assessments. Data were collected from 62 participants, including 28 people with PD and 34 age-matched controls, who performed the clinically recommended Timed Up and Go (TUG), Cognitive Dual-Task TUG (CDTUG), and Motor Dual-Task TUG (MDTUG) tests. This complex multimodal experimental design simultaneously captured brain activation and body motion under motor and cognitive dual-task conditions. Four machine learning models combined with two feature selection techniques were applied to unimodal and multimodal datasets. The multimodal approach achieved superior classification accuracy (96%) compared to fNIRS-only (87%) and IMU-only (95%) models. Key brain–body biomarkers were identified, including dorsolateral prefrontal and frontopolar cortex activations during dual tasks, alongside motor features such as turn, sit-to-stand, and stand-to-sit durations. These findings highlight the promise of combining brain and motion measures and complex functional mobility tests for early-stage PD detection and advance the development of non-invasive, AI-driven biomarker discovery frameworks.