Supplement: SCH: Enabling Data Outsourcing and Sharing for AI-powered Parkinson's Research

Supplement Project Summary What does Project R01 LM014027-01 do? Artificial intelligence holds the promise of transforming data-driven biomedical research and computational health informatics for more accurate diagnosis and better treatment at lower cost. In the meantime, modern digital and mobile technologies make it much

easier to collect information from patients in large scale. While “big” medical data offers unprecedented opportunities of building deep-learning artificial neural network (ANN) models to advance the research of complex diseases such as Parkinson's disease (PD), it also presents unique challenges to patient data privacy.

This project will develop novel data masking technologies based on randomized orthogonal transformation to enable AI-computation outsourcing and data sharing, with the following two aims: 1) Perform two experimental studies of training ANN models with data masking in the HiperGator cloud for PD prediction

and Parkinsonism diagnosis; 2) establish the theoretical foundation on data privacy, inference accuracy, and training performance of the ANN models used in the experimental studies. Why do we make this supplement request? Dr. Aidong Adam Ding from Northeastern University visited us in Summer 2022. Together we produced a manuscript that expanded our data masking method with noise

addition to achieve differential privacy (DiP) when we outsource medical data to the cloud for AI model training. This is a significant advance that goes beyond the originally proposed technical approaches; yet it remains in the scope of the research plan. Therefore, we request a supplement project that utilizes our new

DiP method to transform two PD data sets for guaranteed differential privacy and make them AI-ready for cloud-based machine learning studies. Our analysis has showed that the new DiP method could be improved with much less noise addition, which would result in much better model accuracy. We plan to bring Dr. Ge

Han from Towson University into the team. His expertise in random forests and perturbations could help us reduce the noise. The proposed supplement tasks for the two aims of Project R01 LM014027-01 are below. Supplement Task to Aim 1: Produce two sharable PD data sets with differential privacy and perform a

machine learning case study on the DiP-protected data for AI-readiness evaluation. We will process our two PD data sets with the new DiP method to ensure differential privacy. We will perform an experimental study over the two data sets to evaluate PD-diagnosis models learned from the DiP-protected data and to

quantify the tradeoff between model accuracy and privacy protection, which helps us determine the best configuration of the privacy-protected data that we will share with the community. Supplement Task to Aim 2: Improve the DiP Method and Enhance the Quality of AI-Ready, DiP- Protected Data. We will refine the DiP method for less noise addition, which helps improve the accuracy

of ML/AI models trained from the DiP-protected data.