Bhanu Gullapalli , Yunfei Luo , Tauhidur Rahman , Eric L. Garland
{"title":"基于认知和生理数据的阿片类药物滥用检测","authors":"Bhanu Gullapalli , Yunfei Luo , Tauhidur Rahman , Eric L. Garland","doi":"10.1016/j.drugalcdep.2025.112774","DOIUrl":null,"url":null,"abstract":"<div><h3>Introduction</h3><div>Machine learning may enable detection of opioid misuse to prevent opioid-related risks including overdose and opioid use disorder.</div></div><div><h3>Methods</h3><div>Here, we collected 9238 datapoints from on-body sensors and cognitive tasks in a sample of 169 patients who were prescribed opioid analgesics to manage chronic pain. We categorized patients into one of two groups using the Current Opioid Misuse Measure (COMM): those showing signs of opioid misuse (MISUSE+, n = 116) and those without signs of opioid misuse (MISUSE-, n = 53). Heart rate variability and respiration rate were assessed while participants completed a Dot Probe task involving shifting attention towards and away from opioid-related and emotional cues, and a Go/No-Go task involving inhibition of automatic responses. Cross-sectional data (e.g., physiological responses, task reaction times, task accuracy) were analyzed with a temporal fusion transformer machine learning (ML) model to predict COMM opioid misuse status. We employed Leave-One-Group-Out (LOGO) cross-validation with the participants divided into 10 groups. Each cycle, one group was held out for testing, ensuring robust, unbiased model validation across different subsets of participants.</div></div><div><h3>Results</h3><div>The ML model showed good predictive performance for identifying opioid misuse (AUC, 0.81; specificity, 0.78; sensitivity, 0.78). Behavioral responses were stronger predictors of misuse status than physiological signals.</div></div><div><h3>Conclusions</h3><div>ML models using data from cognitive tasks and on-body sensors detected opioid misuse with an accuracy comparable to gold-standard self-reported opioid misuse assessments. Wearable sensors may provide only incremental predictive power over behavioral responses. Our ML model should be benchmarked against objective measures of opioid misuse.</div></div>","PeriodicalId":11322,"journal":{"name":"Drug and alcohol dependence","volume":"274 ","pages":"Article 112774"},"PeriodicalIF":3.6000,"publicationDate":"2025-07-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"Opioid misuse detection from cognitive and physiological data with temporal fusion deep learning\",\"authors\":\"Bhanu Gullapalli , Yunfei Luo , Tauhidur Rahman , Eric L. Garland\",\"doi\":\"10.1016/j.drugalcdep.2025.112774\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<div><h3>Introduction</h3><div>Machine learning may enable detection of opioid misuse to prevent opioid-related risks including overdose and opioid use disorder.</div></div><div><h3>Methods</h3><div>Here, we collected 9238 datapoints from on-body sensors and cognitive tasks in a sample of 169 patients who were prescribed opioid analgesics to manage chronic pain. We categorized patients into one of two groups using the Current Opioid Misuse Measure (COMM): those showing signs of opioid misuse (MISUSE+, n = 116) and those without signs of opioid misuse (MISUSE-, n = 53). Heart rate variability and respiration rate were assessed while participants completed a Dot Probe task involving shifting attention towards and away from opioid-related and emotional cues, and a Go/No-Go task involving inhibition of automatic responses. Cross-sectional data (e.g., physiological responses, task reaction times, task accuracy) were analyzed with a temporal fusion transformer machine learning (ML) model to predict COMM opioid misuse status. We employed Leave-One-Group-Out (LOGO) cross-validation with the participants divided into 10 groups. Each cycle, one group was held out for testing, ensuring robust, unbiased model validation across different subsets of participants.</div></div><div><h3>Results</h3><div>The ML model showed good predictive performance for identifying opioid misuse (AUC, 0.81; specificity, 0.78; sensitivity, 0.78). Behavioral responses were stronger predictors of misuse status than physiological signals.</div></div><div><h3>Conclusions</h3><div>ML models using data from cognitive tasks and on-body sensors detected opioid misuse with an accuracy comparable to gold-standard self-reported opioid misuse assessments. Wearable sensors may provide only incremental predictive power over behavioral responses. Our ML model should be benchmarked against objective measures of opioid misuse.</div></div>\",\"PeriodicalId\":11322,\"journal\":{\"name\":\"Drug and alcohol dependence\",\"volume\":\"274 \",\"pages\":\"Article 112774\"},\"PeriodicalIF\":3.6000,\"publicationDate\":\"2025-07-08\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Drug and alcohol dependence\",\"FirstCategoryId\":\"3\",\"ListUrlMain\":\"https://www.sciencedirect.com/science/article/pii/S0376871625002273\",\"RegionNum\":2,\"RegionCategory\":\"医学\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q1\",\"JCRName\":\"PSYCHIATRY\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Drug and alcohol dependence","FirstCategoryId":"3","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S0376871625002273","RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"PSYCHIATRY","Score":null,"Total":0}
Opioid misuse detection from cognitive and physiological data with temporal fusion deep learning
Introduction
Machine learning may enable detection of opioid misuse to prevent opioid-related risks including overdose and opioid use disorder.
Methods
Here, we collected 9238 datapoints from on-body sensors and cognitive tasks in a sample of 169 patients who were prescribed opioid analgesics to manage chronic pain. We categorized patients into one of two groups using the Current Opioid Misuse Measure (COMM): those showing signs of opioid misuse (MISUSE+, n = 116) and those without signs of opioid misuse (MISUSE-, n = 53). Heart rate variability and respiration rate were assessed while participants completed a Dot Probe task involving shifting attention towards and away from opioid-related and emotional cues, and a Go/No-Go task involving inhibition of automatic responses. Cross-sectional data (e.g., physiological responses, task reaction times, task accuracy) were analyzed with a temporal fusion transformer machine learning (ML) model to predict COMM opioid misuse status. We employed Leave-One-Group-Out (LOGO) cross-validation with the participants divided into 10 groups. Each cycle, one group was held out for testing, ensuring robust, unbiased model validation across different subsets of participants.
Results
The ML model showed good predictive performance for identifying opioid misuse (AUC, 0.81; specificity, 0.78; sensitivity, 0.78). Behavioral responses were stronger predictors of misuse status than physiological signals.
Conclusions
ML models using data from cognitive tasks and on-body sensors detected opioid misuse with an accuracy comparable to gold-standard self-reported opioid misuse assessments. Wearable sensors may provide only incremental predictive power over behavioral responses. Our ML model should be benchmarked against objective measures of opioid misuse.
期刊介绍:
Drug and Alcohol Dependence is an international journal devoted to publishing original research, scholarly reviews, commentaries, and policy analyses in the area of drug, alcohol and tobacco use and dependence. Articles range from studies of the chemistry of substances of abuse, their actions at molecular and cellular sites, in vitro and in vivo investigations of their biochemical, pharmacological and behavioural actions, laboratory-based and clinical research in humans, substance abuse treatment and prevention research, and studies employing methods from epidemiology, sociology, and economics.