Data Driven Strategies for Substance Misuse Identification in Hospitalized Patients

点击翻译按钮获取中文摘要

PROJECT SUMMARY The rate of substance use-related hospital visits in the US continues to increase, and now outpaces visits for heart disease and respiratory failure. The prevalence of substance misuse (nonmedical use of opioids and/or benzodiazepines, illicit drugs, and/or alcohol) in hospitalized patients is estimated to be 15%-25% and far exceeds the prevalence in the general population. With over 35 million hospitalized patients per year, tens of millions of patients are not screened for substance misuse during their stay. Despite the recommendation for self-report questionnaires (single-question universal screens, Alcohol Use Disorders Identification Test [AUDIT], Drug Abuse Screening Tool [DAST]), screening rates remains low in hospitals. Current screening methods are resource-intensive, so a comprehensive and automated approach to substance misuse screening that will augment current clinical workflow would therefore be of great utility. In the advent of Meaningful Use in the electronic health record (EHR), efficiency for substance misuse detection may be improved by leveraging data collected during usual care. Documentation of substance use is common and occurs in 97% of provider admission notes, but their free text format renders them difficult to mine and analyze. Natural Language Processing (NLP) and machine learning are subfields of artificial intelligence (AI) that provide a solution to analyze text data in the EHR to identify substance misuse. Modern NLP has fused with machine learning, another sub-field of AI focused on learning from data. In particular, the most powerful NLP methods rely on supervised learning, a type of machine learning that takes advantage of current reference standards to make predictions about unseen cases In our earlier version of an NLP and machine learning tool, our opioid and alcohol misuse classifiers successfully used data from clinical notes collected in the first 24 hours of hospital admission to reach a sensitivity and specificity above 75% for detecting alcohol or opioid misuse. We will improve the performance of our baseline, individual NLP single-substance classifiers for alcohol and opioid misuse by implementing multi-label and multi-task machine learning methods. These methods will take advantage of information shared across different types of substance misuse and better capture the state of a patient within a single model. The resulting classifier will be capable of jointly inferring all types of substance misuse (alcohol misuse, opioid misuse, and non-opioid illicit misuse) including polysubstance use, and cater to each individual patient’s substance use treatment needs. We aim to train and test our substance misuse classifiers at Rush in a retrospective dataset of over 35,000 hospitalizations that have been manually screened with the universal screen, AUDIT, and DAST. The top performing classifier will then be tested prospectively to: (1) externally validate its screening performance in a hospital without established screening; and (2) test its effectiveness against usual care at a hospital with questionnaire-based substance misuse screening. We hypothesize that a single-model NLP substance misuse classifier will provide a standardized, interoperable, and accurate approach for universal screening in hospitalized patients and guiding interventions.