In the absence of harm reduction tools, people who use drugs (PWUD) are at increased risk of disease, hospitalization, and death [1-3]. Gaps in the provision of guideline-concordant care to hospitalized PWUD occur, especially among individuals from racially and ethnically minoritized communities [4-6]. Barriers to optimization of health care for hospitalized PWUD include undertreatment of pain and substance use disorders, which have been linked to discharges before medical optimization and higher rates of readmission and mortality [7-9]. Best practices for managing PWUD in a hospitalized setting include addiction care itself as well as treatment and prevention of life-threatening infections [10].

Effective identification of hospitalized PWUD is essential for epidemiological tracking, resource allocation, and evaluating interventions. However, current methodologies often fail to accurately capture this population. The “gold standard” for identifying PWUD hospitalizations is human-guided chart review, a highly regulated and time-intensive process with potential consequences for breach of confidentiality [11,12]. Administrative billing codes (also known as International Classification of Disease codes, ICD codes) have been used for PWUD identification. Unlike several other common conditions such as cardiovascular diseases for which ICD-10 codes are highly accurate [13,14], a systematic review found that for identification of PWUD, ICD-9/10 codes had high specificity but limited sensitivity ranging from 47%‐83% [15,16]. Indicators for substance use tend to be noted in the social history section of the electronic medical record (EMR) rather than a formal diagnosis. Some researchers have used the hepatitis C virus (HCV) codes as a marker of drug use, although there are a substantial number of people with HCV who do not currently use drugs or have ever used drugs [16,17].

The barrier to identifying PWUD can potentially be addressed with natural language processing (NLP), to leverage artificial intelligence (AI) algorithms for interpretation of the written text in a context-relevant manner [18]. NLP has been effectively applied to medical examiners’ reports to increase the accuracy of identifying substance use disorder-related deaths [19], identify substance use disorders in outpatients with HIV [20], and enhance preventive care for hospitalized patients with HIV [17]. In particular, regular expression (RegEx), a rule-based text-matching framework, has been used to identify text patterns [21]. RegEx has recently been used as a tool for identification of encounters with people with opioid use disorder (OUD) [22]. A few studies have examined the application of NLP to identify hospitalized PWUD admitted for bloodstream infections; however, these efforts were single-center evaluations, focused only on injection drug use [23-26]. Despite its innovative capacity to identify PWUD, the field of NLP methodology is nascent. The goal of this study was to evaluate the impact of NLP on the creation of a cohort of hospitalized PWUD and to evaluate disparities in documentation.

The Venn diagram illustrates how 4548 hospitalizations involving PWUD entered the cohort based on inclusion criteria (Figure 1). The study participants’ characteristics are shown in Table 2, along with results of the multivariable logistic regression. People who identified as White or non-Hispanic had higher odds of entering the cohort through NLP alone (adjusted odds ratio [aOR]=2.07; 95% CI 1.54, 2.79). Notably, individuals from the most socioeconomically disadvantaged quartiles (1st and 2nd SVI quartiles) were also significantly more likely to enter the cohort through NLP alone (aOR=1.41; 95% CI 1.06, 1.88). The subcohorts with the highest number of hospitalizations were those with ICD codes only (D-group, n=958), biomarkers only (B-group, n=734), and NLP with all four criteria (B, D, N, M group, n=726). Approximately 10% (n=93) individuals in the D-only group and 35% (n=99) in the N-only group underwent chart review. As shown in Table 3, the positive predictive value (PPV) of the NLP-only cohort was 54%, outperforming the diagnostic codes-only cohort, which had a PPV of 43%. This demonstrates NLP’s ability to enhance identification of PWUD hospitalizations beyond traditional methods.

Our study augments previous work by integrating NLP with diverse identification methods, including urine toxicology and medication records, while simultaneously addressing observed demographic disparities in documentation [23]. NLP has the potential to uncover hospital encounters with PWUD that may have previously been missed. Although NLP had greater PPV than diagnostic codes, its PPV remained low. We found that PWUD from racially and ethnically minoritized communities and those who had low income were more likely to be represented in the minimally documented cohort (ie, entry with NLP-only), rather than the maximally documented cohort.

Largely a result of stigma and racism, PWUD still do not have universal access to evidence-based treatment. Black PWUD tend to enter treatment with a more severe prognosis compared to their White counterparts, partly due to economic barriers in accessing treatment earlier [33]. Black, Latino, and Native American individuals also face additional challenges in accessing treatment for SUD due to geographic barriers, health care access, and potential community characteristics or rapport with clinicians [34]. We found that such a lack of rapport may be represented at the level of documentation for SUD; lack of SUD documentation was strongly associated with racially or ethnically minoritized identity (aOR=2.07).

Identification of PWUD who access medical care is important for several reasons. Best practice guidelines for hospitalized PWUD include management of substance use disorder, pain, and acute infection, testing and management for HIV and HCV, vaccinations for hepatitis or other relevant infections, and prevention of HIV with medications [10,35]. In this study, we applied NLP retrospectively. Following previous studies that identified low HIV testing rates, we plan to use NLP to augment PWUD cohort creation in a study examining patterns of HIV testing [27,36]. NLP could indeed become a valuable tool for identifying PWUD before discharge, facilitating intervention during hospitalization if EMRs could use NLP to trigger clinical decision support tools that trigger clinicians to consider SUD treatment, prescribe overdose prevention medications at discharge, order labs to prepare for pre-exposure prophylaxis, or offer vaccine services.

As we consider this study in the larger context of improving health equity, we believe that the next step would be refining the NLP system by adding more keywords, including and excluding certain conditions and medications, and conducting analyses on false positives and false negative cases. This study should be replicated in other medical centers across the United States; its wider application across various hospitals, encapsulating diverse populations and regions, will be instrumental. This study also has multifaceted applications, spanning epidemiological tracking, optimizing hospital resource utilization, and influencing the design of specific interventional studies. This study’s findings could serve as a launchpad for integrated care for PWUD with less prejudice and inequity. ReGex is a relatively fundamental AI technology, and as more advanced NLP tools become available, we envision our methodology being expanded alongside these too. Regardless of type and complexity of NLP technology, the cohorting and comparative analysis outlined in this paper can be used as a framework to assess the NLP’s performance against conventional ways of locating PWUD.

This study is not without its limitations. The NLP system, despite its effectiveness, occasionally misidentifies certain keywords. The constant calibration of the algorithm and frequent addition of keywords is needed to optimize and sustain accuracy. There are potential flaws in our characterization of domains; limitations include false positives from using ’amphetamine’ as a keyword, which unintentionally classified patients prescribed amphetamines for attention-deficit/hyperactivity disorder as PWUD. Similarly, methadone prescribed for pain management in conditions such as sickle cell disease was misclassified as OUD treatment. Achieving a balance between NLP’s inclusivity and exclusivity presents a significant challenge for this purpose. Future steps should include evaluating the NLP system’s sensitivity and specificity and iterating on the model to enhance these metrics. This will involve refining the keyword list for PWUD, enhancing the NLP algorithm to better account for common confounding variables. The field of addiction medicine is innovative and adaptive; to make NLP a meaningful clinical or research tool in this field, the NLP systems need to receive extensive training and constant input of nuanced decision-making that clinicians partake in daily. Thus, a feedback mechanism and fine-tuning to train the NLP model based on clinician feedback would be critical, fully leveraging repetitive learning, which is one of AI’s biggest strengths. Furthermore, the single-cohort design of the study may limit generalizability; therefore, future studies with streamlined cross-institutional protocols, allowing simultaneous data collection from diverse locations, would improve external validity. This study had a particular focus on comparing diagnostic codes and NLP as single identifiers of PWUD. While NLP identified PWUD with higher PPV than the diagnostic codes, it must be noted that diagnostic criteria still exceeded NLP in the actual number of PWUD cases identified. One major purpose of NLP in PWUD identification is to identify cases that are otherwise missed in conventional screenings; thus, the fact that NLP alone identified a comparable number of PWUD to diagnostic code, with a higher predictive rate, is still remarkable. Future investigation should include a more robust performance comparison between a combination of two or more PWUD clinical identification tools.

The ethics of improving identification of PWUD requires careful consideration. Medical records indicating drug use may become a source of discrimination, compromise job security, housing, and ability to care for family. To mitigate these risks, institutions should implement strict policies ensuring that NLP findings are used solely for improving patient care. Members of this research team collaborated with a broad group of experts including people with lived experience of SUD on a study outlining some of the potential pros and cons of improving systems to identify PWUD with the creation of an additional ICD-10 code for injection drug use [37]. Future work should proactively incorporate the perspectives of individuals with lived experience of SUD. Furthermore, broader discussion regarding AI’s role in health care is needed for effective, ethical, and productive clinical implementation: “Should NLP be a “wide net” or “precision tool” when locating PWUD and connecting them to the care they need?”

Despite these limitations, we believe that this study helps frame the future of systems for measuring health care delivery to PWUD. Hospitalization represents a crucial opportunity when nonjudgmental, trauma-informed, culturally competent care can be offered to PWUD. This presents many potential applications for NLP to be built into systems that track epidemiology and inform quality improvement and implementation science. By integrating NLP, we can advance equitable PWUD care.