Health and Behavior

Deep Learning for Biomedical Health Data

REU Mentor: Behnaz Ghoraani, Ph.D.

REU Scholar: Maria Cardei

REU Scholar Home Institution: University of Florida

Project: Domain Adaptation for Human Activity Recognition of Parkinson’s Disease Patients


Human activity recognition (HAR) is the act of identifying and naming activities using machine learning from gathered data collected from wearable sensors. HAR is revolutionary for the healthcare field. Parkinson’s disease (PD) is a neural disorder that affects patients’ movements, balance, and coordination, which significantly get worse as the disease progresses. With HAR, medical specialists can better understand PD, how it progresses, and improve overall diagnosis and prognosis. The biggest challenge of performing HAR on PD patients is that there is not sufficient labeled sensor data from PD patients. Machine learning models can classify human activities, however these models need a large amount of labeled data to provide significant, robust results.

We have a machine learning model that can successfully classify sensor acceleration data of two public healthy patient datasets (pamap2 and mhealth) with 83.98% and 91.70% accuracy, respectively. When cross testing this model on our privately acquired PD dataset, the accuracy decreases. Domain adaptation (DA) is a method that can be used to improve these results. DA involves training a machine learning model on a source dataset and securing a good accuracy on the target dataset, which is significantly different from the source dataset. The aim of this project is to apply DA to our machine learning model in order to improve HAR classification results of PD data. We ran a few state-of-the-art DA methods, such as Discriminative Adversarial Neural Network (DANN) and Margin Disparity Discrepancy (MDD) with a healthy dataset (pamap2) as the source and the PD dataset as the target dataset. The results so far are comparable to literature standards and our cross testing results. More work is currently being carried out to refine our DA implementations and improve overall results.