From Detection to Behavioral Change: A Scoping Review of Natural Language Processing in Substance Use Research
Abstract
Background: Natural language processing (NLP) has emerged as a promising approach for extracting clinically relevant information from unstructured text that would otherwise be difficult to capture through manual review alone. Although NLP in medical settings began in the 1960s, applications have expanded rapidly in the 2000s, with broader use of digital narrative data such as electronic health records (EHRs), clinical notes, psychotherapy transcripts, and other language-based data sources. In substance use research, however, the literature on NLP applications remains fragmented across data sources, methods, and clinical domains, making it difficult to identify areas of convergence and important gaps. Accordingly, this scoping review examined NLP applications in substance use research, focusing on clinical screening and phenotyping, risk prediction, treatment outcomes, and psychotherapy process. Method: A systematic search of nine clinical and technical databases identified studies published between 2016 and 2026 to examine NLP applications related to substance use behaviors, treatments, and clinically meaningful outcomes. Following screening and eligibility review, 28 studies were included. Extracted data included substance use focus, setting, data source, NLP methods, outcome domain, and key findings. Results: Most studies (78.6%) utilized EHR clinical notes from U.S. healthcare systems, focusing primarily on opioid outcomes. NLP methods consistently improved substance use detection compared to structured diagnostic coding, contributing to nuanced models of overdose risk, relapse, treatment attrition, and opioid prescribing. Recent work focused on detection and prediction to improve multidimensional characterization and computational phenotyping of substance use populations. Across psychotherapy studies, semantic, syntactic, contextual, and affective language features captured clinically meaningful behavior processes and, in recent work, predicted subsequent alcohol use 3 outcomes. However, external validation performance often declined across datasets. Few studies examined non-opioid substances, spoken language, or community samples. Discussion: NLP methods demonstrate potential for improving substance use detection, characterization, and modeling of behaviors across clinical and behavioral domains. Current work suggests a shift from simple classification towards richer representations of behavioral and psychosocial processes with notable limitations in generalizability, data diversity, and external validation remaining challenging for substance use research.
Related
- Lung cancer research: how natural language processing enhances clinical data extraction
- Using Natural Language Processing (NLP) to assess changes in transdisciplinary understandings across a large research consortium
- Sociodemographic characteristics, psychoactive substance use, and an operational marker of recurrent and complex behavioral patterns in patients hospitalized for bipolar disorder: a cross-sectional study
- Volumetric Integrity of Frontotemporal and Perisylvian Regions for Predicting Behavioral and Language Impairment in Behavioral FTD: A Structural MRI Study
- Predicting One-Year Psychiatric Rehospitalization Using Natural Language Processing and Machine Learning Methods Based on Clinical Texts
BibTeX
@misc{prince2026from,
title = {{From Detection to Behavioral Change: A Scoping Review of Natural Language Processing in Substance Use Research}},
author = {Erin Prince and Sheridan Fine and Francesca Filbey},
howpublished = {PsyArXiv},
year = {2026},
month = sep,
doi = {10.31234/osf.io/k6s57_v1},
url = {https://osf.io/preprints/psyarxiv/k6s57_v1/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - From Detection to Behavioral Change: A Scoping Review of Natural Language Processing in Substance Use Research
AU - Erin Prince
AU - Sheridan Fine
AU - Francesca Filbey
PY - 2026
DA - 2026/09/29
PB - PsyArXiv
DO - 10.31234/osf.io/k6s57_v1
UR - https://osf.io/preprints/psyarxiv/k6s57_v1/
AB - Background: Natural language processing (NLP) has emerged as a promising approach for extracting clinically relevant information from unstructured text that would otherwise be difficult to capture through manual review alone. Although NLP in medical settings began in the 1960s, applications have expanded rapidly in the 2000s, with broader use of digital narrative data such as electronic health records (EHRs), clinical notes, psychotherapy transcripts, and other language-based data sources. In substance use research, however, the literature on NLP applications remains fragmented across data sources, methods, and clinical domains, making it difficult to identify areas of convergence and important gaps. Accordingly, this scoping review examined NLP applications in substance use research, focusing on clinical screening and phenotyping, risk prediction, treatment outcomes, and psychotherapy process. Method: A systematic search of nine clinical and technical databases identified studies published between 2016 and 2026 to examine NLP applications related to substance use behaviors, treatments, and clinically meaningful outcomes. Following screening and eligibility review, 28 studies were included. Extracted data included substance use focus, setting, data source, NLP methods, outcome domain, and key findings. Results: Most studies (78.6%) utilized EHR clinical notes from U.S. healthcare systems, focusing primarily on opioid outcomes. NLP methods consistently improved substance use detection compared to structured diagnostic coding, contributing to nuanced models of overdose risk, relapse, treatment attrition, and opioid prescribing. Recent work focused on detection and prediction to improve multidimensional characterization and computational phenotyping of substance use populations. Across psychotherapy studies, semantic, syntactic, contextual, and affective language features captured clinically meaningful behavior processes and, in recent work, predicted subsequent alcohol use 3 outcomes. However, external validation performance often declined across datasets. Few studies examined non-opioid substances, spoken language, or community samples. Discussion: NLP methods demonstrate potential for improving substance use detection, characterization, and modeling of behaviors across clinical and behavioral domains. Current work suggests a shift from simple classification towards richer representations of behavioral and psychosocial processes with notable limitations in generalizability, data diversity, and external validation remaining challenging for substance use research.
ER -