NeuroInsight-PD: Early Detection and Classification of Parkinson’s Disease

Most people with Parkinson’s Disease are diagnosed too late. By the time a neurologist discovers the condition through standard clinical tests, the brain has already lost 60% or more of its dopamine-producing cells. At this stage, the disease has progressed to a level where slowing its progression becomes challenging. The earliest signs are overlooked and only become apparent during clinical evaluation. There is a gradual decrease in handwriting, along with traces of tremor. Finger dexterity becomes rhythmically inaccurate. Changes in gait patterns are practically unnoticed by the naked eye but can be measured using sensors. The first stage of the NeuroInsight-PD project is devoted to the initial state of the motor dysfunction. The platform records locomotor pattern on a set of wearable sensors, fine motor skill with a series of tremor and finger-tapping tests, and spatial anomalies known of micrographia with a set of handwritten samples. Instead of implementing one of the three modalities, the system combines all three streams of data into the different streams of processing them: a Random Forest classifier is used on the structured sensor data, and an EfficientNetB0 convolutional neural network processes the handwriting images. The resulting scores are then combined through the weighted decision-fusion layer, which generates a single likelihood estimate of the Parkinson disease. Experimental analysis shows that the combined system has a 95% accuracy and 96% precision, which is always higher than the model that is solely dependent on one of the modalities. Ablation analysis indicates that elements such as hard negative mining, threshold calibration, and other elements significantly contribute to the sensitivity of early detection. The solution is deployed to FastAPI back-end and ReactJS front-end offering the ability to screen in real time and be scaled.