RAQ-Streams: Real-time pattern recognition with quantization-aware training for substance use episodes in wearable biosensor data streams
Advisor: Dr. Hua (Julia) Fang
Committee members:
- Dr. Firas Khatib
- Dr. Honggang Wang
- Dr. Shiwen Mao, ECE, Auburn University
Title: RAQ-Streams: Real-time pattern recognition with quantization-aware training for substance use episodes in wearable biosensor data streams
Abstract:
The widespread adoption of wearable biosensors in medical and consumer devices such as smartwatches has enabled continuous acquisition of biophysiological data streams for health monitoring. These high-frequency data provide new opportunities to characterize relapse dynamics and other clinically significant events during the treatment of Substance Use Disorder (SUD) in natural environments. In this setting, real-time detection of substance use episodes is critical for enabling timely and individualized interventions. Given the sensitivity of SUD-related data, edge-deployable machine learning models that minimize communication with external servers are particularly advantageous. Quantization-Aware Training (QAT) offers a practical approach by enabling computationally efficient deployment on resource-limited devices while maintaining predictive accuracy comparable to full-precision models. Through on-device inference, such models can achieve privacy-conscious analytics by processing biosensor data locally rather than through centralized cloud systems. Although recent progress in streaming data analytics has improved single-substance detection, accurate polysubstance episode identification remains a significant methodological challenge. To address this limitation, this work presents RAQ-Streams (Real-time Pattern Recognition with Quantization-Aware Training in Wearable Biosensor Data Streams), a framework for detecting substance use episodes from streaming biosensor data. RAQ-Streams integrates QAT with real-time inference to enable deployment on edge devices without compromising model fidelity. Distinct from prior approaches, it accommodates polysubstance detection, performs continuous prediction, and yields higher accuracy in single-substance detection compared to state-of-the-art baselines. Ablation studies demonstrate that QAT markedly reduces the performance gap between quantized and full-precision networks, outperforming post-training quantization. Additional analyses evaluate the contributions of architectural components, with attention enhancing temporal focus and robustness, and L1 regularization reducing overfitting and enabling sparse feature selection. Comparative experiments with alternative classifiers, including multi-layer perceptron, convolutional neural networks, support vector machines, and Bayesian networks, were performed using both simulated data (evaluated via Fréchet Inception Distance and related measures) and real-world biosensor data collected from a clinical SUD trial. Visual analyses of electrodermal activity and related signals reveal characteristic temporal patterns associated with substance use and underscore the complexity of polysubstance detection.
A key challenge addressed in this study is the limited availability and uncertainty of ground-truth labels. To mitigate this, a hybrid labeling strategy combining urinalysis outcomes, prior validated datasets, and domain expertise was implemented to enhance the reliability of training and validation data. Furthermore, novel derived features from biosensor streams were developed to differentiate substance-induced physiological responses from behavioral or environmental artifacts, contributing new insights to SUD signal analytics. The QAT-enabled design allows direct deployment of RAQ-Streams on mobile devices. A prototype mobile application demonstrates its real-time detection capability within a distributed edge–cloud architecture for continuous learning and biosensor data streaming. This research contributes to wearable computing, digital health, and substance use informatics by introducing an efficient, scalable, and privacy-conscious framework for real-time polysubstance use detection. It extends the application of quantization-aware training to continuous health data and advances software infrastructure for biosensor-based behavioral health research. By facilitating near-real-time characterization of relapse-related events, RAQ-Streams represents a technically grounded step toward adaptive, sensor-driven systems for SUD monitoring and intervention.
All EAS students are encouraged to attend and all interested parties invited. For further information, please contact Dr. Hua Fang.
Zoom Meeting: https://umassd.zoom.us/j/95683754446?pwd=WW1paGw5Q29mdXBMb0E3N3dkUTZ2Zz09
Hua (Julia) Fang
hfang2@umassd.edu
https://umassd.zoom.us/j/95683754446?pwd=WW1paGw5Q29mdXBMb0E3N3dkUTZ2Zz09