OpiTrack: A Wearable-based Clinical Opioid Use Tracker with Temporal Convolutional Attention Networks

Area of Science:

• Digital Health
• Biomedical Engineering
• Data Science

Background:

• Opioid use disorder presents significant societal and economic challenges.
• Wearable physiological sensing technologies are underutilized for detecting drug use in real-world settings.
• Developing systems to identify high-risk opioid events is crucial for mitigating overdose risks.

Purpose of the Study:

• To investigate the feasibility of detecting opioid administration using physiological data from wrist-worn sensors.
• To establish a foundation for a mobile technology system for real-time opioid use monitoring and intervention.

Main Methods:

• Thirty-six hospitalized patients receiving opioid analgesics were enrolled.
• A noninvasive wrist sensor continuously collected physiological data (heart rate, skin temperature, accelerometry, electrodermal activity, interbeat interval) over 1-14 days.
• A Channel-Temporal Attention TCN (CTA-TCN) model was employed to analyze 2070 hours of data, correlating with 339 opioid administrations.

Main Results:

• The CTA-TCN model detected opioid administration within a time window with an F1-score of 0.80, sensitivity of 0.80, specificity of 0.77, and AUC of 0.77.
• The model predicted the exact moment of administration with a normalized mean absolute error of 8.6%.
• An R-squared coefficient of 0.85 was achieved for predicting the timing of opioid administration.

Conclusions:

• Physiological signals captured by wrist-worn sensors can effectively detect opioid administration.
• The developed CTA-TCN model demonstrates significant potential for real-time opioid use detection.
• This research supports the development of mobile health interventions for opioid use disorder management and overdose prevention.