Privacy-preserving assessment of depression using Speech signal processing
Open Access
- Author:
- Bettapalli Nagaraj, Suhas
- Graduate Program:
- Electrical Engineering
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- April 02, 2021
- Committee Members:
- Saeed M Abdullah, Thesis Advisor/Co-Advisor
David Jonathan Miller, Committee Member
Jing Yang, Committee Member
Kultegin Aydin, Program Head/Chair - Keywords:
- Speech Classification
Data Privacy
Depression Detection
Paralinguistics - Abstract:
- Machine learning has been used for classification experiments in various domains such as healthcare, banking, and sports. Due to the lack of time and large training datasets, it could be a good idea to outsource data collection to third parties, for example, through cloud-based aggregation. This outsourcing, however, could lead to data breaches. When datasets include immutable biometric data, the situation becomes much direr, necessitating data confidentiality during the testing phase. In this work, as a proof of concept that could be applied to other healthcare and commercial analytics tasks, we focus primarily on using privacy-preserving systems to detect depression. We perform two experiments, a two-class depression detection task (has depression/ no depression) and a two-class depression severity classification (none-mild/ moderate-severe). Our goal is to demonstrate that the implementation of privacy-preserving speech mining systems is feasible and comparable to the performance of regular speech features in challenging tasks involving paralinguistic components while crowd-sourcing both the data collection and training phases on mobile devices. We use two federated learning schemes, Federated Averaging (FedAvg) and Federated Matching Averaging (FedMA), for the privacy-preserving schemes. All the three algorithms (centralized, FedAvg, and FedMA) are trained on three pre-trained networks (ResNet-18, GoogleNet, and MobileNet v2). These models are suitable for mobile devices due to their small size and efficient processing. The two federated approaches ensure data privacy with a slight degradation regarding accuracy compared to centralized architectures. This thesis work could pave the way for a robust, privacy-sensitive, edge device-based detection of psychological distress assessments for use in a clinical setting.