Cannabis Use Documentation within the Electronic Health Record: A Use Case for Natural Language Processing Methods

Abstract

Introduction

Recreational and medical cannabis use (CU) information is often available within the electronic health record (EHR) in a format that is impractical for health care provider use. Transformation of free-text EHR documentation in notes to discrete elements is possible using natural language processing (NLP) and has the potential to characterize CU efficiently. The objective of this study was to develop an NLP algorithm to identify CU documentation within unstructured EHR clinical notes.

Methods

We identified EHR notes with cannabis-related terminologies through a keyword search among all Geisinger patients with at least one encounter between January 1, 2013 and June 30, 2022. We trained four NLP models to classify CU documentation within notes into six categories based on time, context, and reliability, as identified through manual annotation. We compared the demographic characteristics of patients with a positive CU classification using the best-performing model to those of the studied sample.

Results

Of the over 1.7 million eligible patients, 150,726 (8.6%) were flagged as cannabis users. Bio-ClinicalBERT, a transformer-based NLP model, achieved close to human performance in classifying CU (weighted precision = 91.4, recall = 93.3, and F score = 92.4). An unadjusted analysis showed that cannabis users had higher body mass index and were at least nine-fold more likely to use tobacco, alcohol, or illicit substances.

Conclusion

Our study evaluated the prevalence of CU documentation across the entire corpus of EHR notes data based on available data without population segmentation over a 9.5-year period. The NLP methodologies used achieved performance close to that of human annotation and laid the foundation for identifying and classifying CU within unstructured data sources, with future applications in research and patient care.

Plain Language Summary

Marijuana, also known as cannabis, may impact the health of patients, yet it is not routinely captured in medical records. When documented, it is often found in unstructured formats (e.g., progress notes) rather than in discrete fields. Incomplete and unstructured capture limits many functional capabilities within the electronic health record (EHR) that enhance patient care (e.g., drug interactions, notifications) and limits researchers from identifying patients routinely exposed to marijuana use. The transformation of free-text documentation of cannabis use (CU) into discrete elements can be performed using natural language processing (NLP). The objective of this study was to develop an NLP model to identify CU in unstructured clinical notes in the EHR. We examined the EHRs of Geisinger patients in Pennsylvania over a 10-year period. Among 1.7 million patients, 9% were identified as CU. One of the NLP models tested, Bio-ClinicalBERT, achieved the best performance. An unadjusted exploratory analysis showed that cannabis users had a higher body mass index and were 10-fold more likely to be tobacco users, 10-fold more likely to use alcohol, and nine-fold more likely to use illicit substances in our studied sample. NLP can be used to better understand the risks and benefits of CU at the population level and may improve patient identification to assist clinical decision-making, but could be limited by the quality of CU documentation within EHR. Future CU epidemiological research should continue to explore other avenues to automate and improve CU documentation by leveraging rapidly evolving technologies such as artificial intelligence-driven tools.

Introduction

Cannabis products, both recreational and medical, are commonly used in the USA [1–3]. The 2024 National Household Survey on Drug Use and Health estimated that 44.3 million people, aged ≥12 years, had past-month marijuana use [4]. The Commonwealth of Pennsylvania (PA) passed its Medical Marijuana (MMJ) Act in 2016 [5]. As of November 2025, more than one out of every 25 PA adults (3.37%) were certified for medical marijuana [6], with anxiety disorders, chronic pain, and posttraumatic stress disorder accounting for the preponderance (85.6%) of certifying conditions (online suppl. Fig. 1; for all online suppl. material, see https://doi.org/10.1159/000553307) [7, 8].

The medical marijuana certification process in PA, like most other states, is typically completed via telemedicine by a physician with whom the patient does not have a long-standing relationship, as they do with their primary care provider [9, 10]. Three-quarters of states, including PA, that have a medical marijuana program do not report any information to their Prescription Drug Monitoring Program [11]. With the executive order from December 2025 facilitating the potential reclassification of cannabis from a schedule I to a schedule III substance, there are a multitude of reasons why physicians and pharmacists need readily available cannabis use (CU) information. These include recognition of cannabis-drug interactions [12, 13], alignment with evidence driven utilization for medical conditions for which CU is indicated [14–16], awareness of CU as a substitute for prescription medications [17], aiding diagnosis of CU-related conditions (e.g., CU disorder, hyperemesis syndrome, etc.) [18], and awareness of CU can assist clinicians in discussing safe use practices (e.g., storage in homes with children) [19]. The examples listed above highlight some of the scenarios that explain the necessity of improving the understanding of an individual’s CU and its importance in making informed health decisions.

Despite the need, the amount of actionable information documented in electronic health records (EHRs) remains inadequate (e.g., only 4.8% of medical CU is documented in the EHR compared to 35.1% that is implicitly reported) [20, 21]. The most common sources of CU documentation are unstructured clinical notes [22–31] and patient portal messages [32]. A prior report from members of this team in PA discovered that secure patient portal messages between patients and health care providers discussing cannabis from 2012 to 2022 peaked in 2019 after dispensaries opened in 2018 [32]. Natural language processing (NLP) models have been leveraged to identify CU-related documentation from unstructured text [26–31]. Most prior studies applying NLP for CU identification have focused on specific sub-populations by disease states, such as psychosis patients, primary care patients, older adults undergoing surgery [26–28, 30, 31, 33] or limited age ranges like pediatric and young adults [26, 29, 31], thereby limiting their applicability for understanding CU across the entire population captured within the EHR. Despite these limitations in the prior studies, their use of machine learning models such as logistic regression (LR), support vector machine (SVM), linear support vector classifier, and transformer-based models such as bidirectional encoder representations from transformers (BERT), and Bio-ClinicalBERT have laid a strong foundation for applying these models to the current analysis [26, 27, 29, 30, 34]. The purpose of this investigation was to extend prior studies [25, 32] and to use validated NLP methodologies to classify CU from unstructured EHR notes, providing a proof of concept that free-text EHR notes can be extracted for CU using an automated mechanism from a broad, unstratified population [26].

Methods

We adhered to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guidelines for reporting purposes (online suppl. G, [35]). Figure 1 presents the pipeline and phases involved in this study.

Fig. 1.

Pipeline and phases of the development of the NLP algorithm. Diagram of the pipeline and phases involved in this study for the development of an NLP algorithm to assess the prevalence of CU documentation over a period of 9.5 years within EHR notes.

Study Setting

This investigation was conducted at Geisinger, an integrated health system in the USA, spanning 24 counties in central and northeast PA (online suppl. Fig. 2a). Geisinger comprises ten acute care hospitals, 133 primary and specialty clinics, a research institute, and an insurance company, Geisinger Health Plan. Geisinger began implementing its EHR, Epic^® Corporation (Verona, WI), in 1996 at all ambulatory sites and integrated it across inpatient care sites in 2003. This infrastructure provides a comprehensive view, including clinical information, imaging, and administrative claims for rural and urban residents, offering a real-world examination of dynamic changes in this population.

Patient Population

The study included patients with at least one encounter (e.g., inpatient, outpatient, and emergency department) at a Geisinger facility between January 1, 2013 and June 30, 2022. This population will henceforth be referred to as the studied sample. For patients included in the studied sample, we only pulled EHR notes that included one of the cannabis-related terms from the lexicon. Although population declines occurred in 20 of 24 (83.3%) counties (online suppl. Fig. 2b), nativity levels remained high in the Geisinger footprint (online suppl. Fig. 2c), indicating a stable population with low migration, thereby supporting the use of this population for the extended period employed in this investigation.

Lexicon Development and Dataset Construction

This report used snippets of encounter notes curated from the EHR, including nursing notes, appointment notes, problem list-based free-text documentation, visit reason documentation, and follow-up notes (online suppl. A – Data sources). We developed a CU lexicon to narrow the scope of data retrieval. The CU-related terms in the lexicon were finalized based on prior literature [26, 27] and input from the study team’s subject matter experts (M.P.D., C.K.K., and B.J.P.). The lexicon included the following terms: “marijuana,” “cannabis,” “MJ,” “THC,” “CBD,” “weed,” “MMJ,” “indica,” “sativa,” “cannabinoid,” “spice,” “tetrahydrocannabinol,” “pot,” “cannabidiol,” “ganja,” “grass,” “hash,” “hashish,” “bong,” “Mary Jane,” “edibles,” and “joint.” Among the 1.76 million patients seen at Geisinger during the study period, we identified 1.71 million patients (135,858,323 EHR notes) with exact-word matches in their EHR notes using this lexicon. We then trimmed these EHR notes to retain only 300 characters before and after the CU-related term. We will refer to these trimmed snippets of notes simply as EHR notes in the context of this study moving forward. We observed that specific terms were used in contexts unrelated to cannabis; for instance, the term “pot” was used in the context of “neti pot,” “netti pot,” or as a part of the word “hypotension,” “potential,” or “potassium.” In contrast, the term “CBD” was used in an anatomical context for the “Common Bile Duct,” such as “CBD measure” and “CBD stricture.” These phrases and words were removed during data cleaning. “CBD” or “pot,” only when mentioned in the context of marijuana use, was retained in the final analytical sample for classification. The term “joint” was also primarily and extensively used in an anatomical context, such as “joint pain,” “joint space,” “joint deformity,” and “joint disease.” To minimize such misclassifications, we removed EHR notes that contained these specific phrases or words (online suppl. A – Primary data cleaning), resulting in a working sample of 17,789,648 EHR notes. The data cleaning process also involved the removal of notes when the term of interest was negated (e.g., “no cannabidiol use” or “tetrahydrocannabinol [THC]: no”). This led to a more concise final list of terms and a final analytical sample of 2,790,896 EHR notes from 370,885 unique patients with at least one mention of a key term in patient charts. We used SAS Enterprise Guide 8.3 for dataset construction and cleaning.

Manual Annotation Phase

We developed an annotation schema and decision rules for labeling CU in EHR notes. Initially, VS and AP manually reviewed and double-coded a random sample of 100 EHR notes using labeling categories informed by previous literature [26]. Labeling categories were refined following a discussion of the assigned codes, with further input from team members to resolve disagreements (online suppl. B – Labeling guidelines). This review phase eventually led to the creation of six categories:

• (1)True mention – The identified key term in the EHR notes refers to cannabis.
• (2)Patient use – The true mention applies to the patient’s use, not another person.
• (3)Indication of use – The EHR note explicitly mentions or indicates that the patient has used cannabis at some point.
• (4)Medical use – The EHR note explicitly mentions that the patient used it for a medical purpose.
• (5)Current use – The EHR note explicitly mentions that the patient currently uses.
• (6)History of use – The EHR note explicitly mentions that the patient has a history of use.

Three annotators were subsequently trained using the decision rules and the initial set of 100 annotated EHR notes (online suppl. C – Decision rules). Following training, all three annotators independently reviewed and labeled the same random sample of 50 additional EHR notes (300 potential labels across six categories), which was used to calculate inter-rater reliability kappa’s for each labeling category (online suppl. D – Examples; online suppl. E − Inter-rater reliability kappa). Following completion of the inter-rater reliability phase, the annotators independently single-coded an additional 3,500 unique EHR notes without overlap between annotators, resulting in 3,650 total annotated EHR notes used for model development. These notes were randomly sampled based on cannabis-related terminology from the lexicon, with each cannabis-related term represented in at least 50 notes. Throughout the annotation phase, ad hoc situations when existing labeling guidelines seemed insufficient arose, the notes were reviewed and discussed with study team experts to further refine category definitions and decision rules, and labeling decisions for those cases were made using a consensus-based approach. The annotated sample was drawn based on cannabis-related terms in the lexicon and was not adjusted for any patient characteristics. All annotation work was conducted using Microsoft Excel™.

Model Development and Training

We trained two traditional models and two transformer-based models to identify CU in EHR notes: (1) LR, (2) SVM, (3) BERT [36], and (4) Bio-ClinicalBERT [37]. LR and SVM were included as traditional baseline machine learning approaches for text classification tasks. BERT and Bio-ClinicalBERT were selected based on prior studies demonstrating strong performance for cannabis-related and clinical NLP classification tasks using EHR notes [26]. Bio-ClinicalBERT was considered particularly suitable for this application because it was pretrained on clinical text corpora, which may improve representation of clinical terminology and documentation patterns commonly observed in EHR notes. Prediction of each CU category was treated as a separate binary classification task for each model. Model development and training were conducted using Python 3.10.13. A detailed breakdown of the training and testing sample is included in Table 1.

Table 1.

Frequency breakdown of cannabis-related terminology used

Term	Training dataa		Total analytical sampleb
	notes, n (n = 3,650), %	patients, n (n = 3,316c), %	notes, n (n = 2,790,896), %	patients, n (n = 601,687c), %
Marijuana	17.8	19.4	56.7	35.5
CBD	6.8	7.5	13.2	17.2
Cannabis	6.8	7.4	7.9	5.0
THC	6.8	7.5	7.5	13.2
Pot	6.8	7.5	5.1	11.0
MJ	6.8	7.4	3.1	5.0
Grass	6.8	7.5	3.0	6.7
Weed	6.8	7.3	1.8	3.8
Spice	6.8	7.1	1.2	1.3
Hash	6.8	6.5	0.3	0.7
Edibles	6.8	5.9	0.1	0.2
Indica	6.8	6.1	0.1	0.2
Sativa	6.8	3.0	0.04	0.03
Bong	–	–	0.02	0.04
Ganja	–	–	0.001	0.002

Frequency breakdown of cannabis-related keyword search terms that were used for developing classification NLP models post data cleaning within a 2.7 million analytical sample at Geisinger over the 9.5-year study period.

CBD, cannabidiol; CU, cannabis use; EHR, electronic health records; MJ, marijuana; NLP, natural language processing; THC, tetrahydrocannabinol.

^aThe training data were random sample of notes with keyword search terms that were used for manual annotation and developing the NLP models.

^bThe total analytical sample included the 2,790,896 EHR notes that the NLP models were subsequently applied to for classifying notes with CU.

^cThe number of patients listed here does not represent the number of unique patients, considering that a patient may have more than one term of interest documented within their EHR notes. The number of unique patients in the training/annotated data sample was 3,173, while in the analytical sample was 370,885.

For the LR and SVM models, we used a bag-of-words approach using unigrams, bigrams, and trigrams to represent each EHR note. We removed stop words and built bag-of-words representations using the sci-kit learn library [38]. We used elastic net regularization for dimensionality reduction in the LR models, with an L1 ratio of 0.5, which applies both L1 and L2 penalties equally. For SVM, a linear kernel was applied. We tested regularization weights of 0.01, 0.1, 1, 10, and 100 for LR and SVM. We used a regularization weight of 0.01 for LR models and 10 for SVMs, as model performance was highest with these weights (based on F score across all CU categories). We conducted 5-fold cross-validation to train and evaluate models, averaging performance metrics across each test fold. Precision, recall, and F score were used to assess model performance (online suppl. F: Classifier performance).

For BERT and Bio-ClinicalBERT, we used each model’s pretrained transformer layers and tokenizers available through Hugging Face’s transformer library [37, 39]. All pretrained transformer weights were fully fine-tuned on the corpus of labeled EHR notes. Training was conducted for ten epochs with a batch size of 8 using the AdamW optimizer with a learning rate of 2 × 10⁻⁵. Models were trained to minimize binary cross-entropy loss using sigmoid activation functions for each binary classification task. We conducted 5-fold cross-validation for model training and evaluation, wherein the annotated dataset was randomly partitioned into five mutually exclusive folds. During each iteration, approximately 80% of the data were used for training and 20% were held out for testing, with each fold serving once as the test set across the five iterations. The number of epochs was selected a priori to maintain a consistent training framework across all transformer models and cross-validation folds. Because model evaluation relied on 5-fold cross-validation without a separate validation set, epoch selection was not based on validation performance or early stopping procedures. Precision, or more commonly known as positive predictive value, is the proportion of positive results that are true positives in the predicted sample [40]. It answers the question, “if the patient is classified as a CU, what is the probability they actually engage in CU?” [40] Recall, also known as sensitivity or true-positive rate, quantifies how well a test identifies true positives (i.e., correctly identifies subjects who truly have the condition of interest) [40]. An F score combines precision and recall into a single score and is used to evaluate the performance of the machine learning model. All BERT and Bio-ClinicalBERT models were implemented using Pytorch 2.1.0. Our model training and testing scripts are available at [41].

CU Prevalence and Patient Characteristics

After evaluating each trained model, we used the highest-performing model (in terms of weighted F score) to assign “True mention” and “Indication of Use” classifications to all unlabeled EHR notes. Patients who had an EHR note positively classified for both the “True mention” and “Indication of Use” categories were flagged as cannabis users. Any patient with at least one EHR note being flagged positive for CU was subsequently identified as a cannabis user for the remainder of the study period for descriptive analysis purposes.

We assessed the prevalence of CU documentation in the EHR using the CU flag. We computed descriptive statistics in patients flagged for CU at the time when the patient was first identified as a cannabis user, and in the studied sample at the time of their first encounter during the study period; this date was termed as the index date. We calculated means or medians with SDs or interquartile ranges for continuous variables and frequencies and percentages for categorical variables, when appropriate, to evaluate the distribution of patients’ demographic and clinical characteristics, including current substance use (smoking, alcohol, and illicit drug use) and body mass index (BMI) in the annotated sample, studied and CU populations. Demographics and clinical characteristics were collected from the EHR as of the index date. Substance use data was collected using a combination of the social history documentation of drug use and the presence of ICD 9/10 codes. We further assessed the magnitude of difference in the demographic and clinical characteristics between the annotated sample, CU, and the studied sample using t-tests and chi-square tests. However, we did not perform any adjusted analysis accounting for confounders as a part of this study, since the intent of assessing patient characteristics was purely exploratory in nature to better understand preliminary distributions. Of note, most recent BMI data were missing at a varying level across the three samples; the distributions were computed based on data available, and as a result, should be interpreted with increased caution. Future analysis should focus on accounting for these confounders as well as missingness in the case of BMI data. All analyses were performed using SAS statistical software (SAS Enterprise Guide 8.3, Cary, NC).

Results

We retrieved 2,790,896 EHR note snippets from 370,885 unique patients; 98.7% were captured within nursing notes. “Marijuana,” along with its misspelling (“Marjiuana”), was the most used terminology (56.7%), followed by “CBD” (13.3%), “Cannabis” (7.9%), “THC” (7.5%), and “Pot” (5.1%) (Table 1). From the 3,650 EHR notes in the annotation set, 1,945 (53.3%) had a “True mention” of CU. Among the “True mentions,” 1,911 (98.3%) were explicitly discussing the patient’s personal use at the time of the encounter. 1,476 (77.2%) had documentation about the “Indication of use,” 359 (18.8%) indicated medically recommended use, 973 (50.9%) had documentation indicating “Current Use,” and 551 (28.8%) had a match indicating “Past Use.” A high kappa (>0.83) was observed between the three annotators for all the labeling categories in the training sample, except for “History of use” (online suppl. C – Labeling categories). Online supplementary Table D includes example snippets for each labeling category.

Classification of CU was close to human performance for the “true mention” and “indication of use” categories across all models, with an F score range of 93.2%–95.7% and 84.2%–88.0%, respectively (Table 2). Bio-ClinicalBERT achieved the highest performance for the “true mention” and “indication of use” categories with a weighted average F score of 92.4%. Except for the “patient use” category, the performance thresholds were comparatively lower for all models for the “medical use,” “current use,” and “history of use” categories. For these remaining categories, SVM performed slightly better than Bio-ClinicalBERT (weighted average F score 81.6% vs. F score 80.8%). SVM also performed the best for identifying medical use (P-82.8%, R-58.7%, F-68.4%) and current use (P-78.1%, R-68.1%, F-72.7%), while Bio-ClinicalBERT performed the best for identifying patient use (P-95.4%, R-95.1%, F-95.3%) and past use (P-57.9%, R-74.1%, F-64.5%) (Table 3).

Table 2.

Performance metrics for models used to classify cannabis users

​	Logistic regression	BERT	SVM	Bio-Clinical BERT
True mention (n = 1,945)
Precision	0.95	0.91	0.97	0.96
Recall	0.96	0.96	0.94	0.96
F1 score	0.95	0.93	0.96	0.96
Indication of use (n = 1,496)
Precision	0.85	0.81	0.87	0.86
Recall	0.88	0.88	0.86	0.90
F1 score	0.87	0.84	0.86	0.88
Weighted average
Precision	0.90	0.86	0.93	0.91
Recall	0.92	0.92	0.91	0.93
F1 score	0.91	0.89	0.92	0.92

Performance metrics of algorithms for identifying cannabis users from preprocessed EHR notes data from Geisinger. The models were run on a randomly selected and manually annotated sample of 3,650 notes.

BERT, bidirectional encoder representations from transformers; SVM, support vector machine.

Table 3.

Performance metrics of models used for identifying CU-related characteristics

​	Logistic regression	BERT	SVM	Bio-Clinical BERT
Patient use (n = 1,911)
Precision	0.94	0.91	0.96	0.95
Recall	0.95	0.94	0.94	0.95
F1 score	0.94	0.92	0.95	0.95
Medical use (n = 359)
Precision	0.82	0.58	0.83	0.52
Recall	0.33	0.71	0.59	0.86
F1 score	0.46	0.61	0.68	0.65
Current use (n = 973)
Precision	0.76	0.60	0.78	0.60
Recall	0.60	0.74	0.68	0.80
F1 score	0.67	0.66	0.73	0.68
Past use (n = 551)
Precision	0.73	0.61	0.74	0.58
Recall	0.41	0.69	0.51	0.74
F1 score	0.52	0.64	0.60	0.65
Weighted average
Precision	0.85	0.75	0.87	0.77
Recall	0.72	0.83	0.78	0.87
F1 score	0.77	0.79	0.82	0.81

Performance metrics of algorithms for identifying characteristics of CU context are documented in Geisinger EHR notes. The models were run on a randomly selected and manually annotated sample of 3,650 notes.

BERT, bidirectional encoder representations from transformers; SVM, support vector machine.

The trained Bio-ClinicalBERT models for “True Mention” and “Indication of use” subsequently classified over two-fifths (40.6%, 150,726) of the patients from the analytical sample as positive for CU. This represented 8.6% of the studied sample. The distributions of all demographic variables were statistically different between the annotated sample (training dataset), the CU-documented population, and the studied sample based on an unadjusted analysis (p < 0.0001). The average age of patients with CU documentation was similar to that of the studied sample (CU-37.6 [SD-18.8] vs. studied-38.1 [SD-24.6]). The proportion of patients over 65 (CU-9.3% vs. studied-17%) and under 18 (CU-11.1% vs. studied-24.6%) was higher in the studied sample than in patients with CU documented in their EHR. Additionally, just over half of patients with CU were male (52.5%), compared with the studied sample, where just over half were female (52.3%). Patients with CU were less likely to be Asian and more likely to be African American than the studied sample. In the subset of patients for whom BMI data at the index date were available, more than half of cannabis users (55.3%) had a BMI of >25, and of these, 32.4% had a BMI of ≥30 (i.e., obese). The proportion of obese patients was lower in the studied sample (23.9%). The average BMI of cannabis users was also greater than the studied sample (CU mean = 28.5, SD = 8.6 vs. studied mean = 26.9, SD = 9.0; Table 4). The distribution of some patient characteristics also differed between the annotated sample and those for CU and the studied sample, as the random sample used for annotation was based on CU-related terminology of interest (Table 4). Upon comparing substance use behaviors between cannabis users and studied sample without adjusting for any demographic or clinical covariates, we observed that CU were unlike the studied sample in their current or history of use of various drugs in that they were ten-fold more likely to report tobacco smoking (49.3% vs. 5.1%), ten-fold more likely to report alcohol use (48.2% vs. 4.9%), and nine-fold more likely to report illicit drug use (4.7% vs. 0.5%, Fig. 2).

Table 4.

Descriptive breakdown of patient characteristics

Total (n) –	Annotated sample	Cannabis use in EHR	Studied sample
Total patient (N)	3,173	150,726	1,762,548
Demographics**
Age, years
Mean (SD)	40.2 (21.8)	37.6 (18.8)	38.1 (24.6)
<18, N (%)	534 (16.8)	16,786 (11.1)	432,621 (24.6)
18–30, N (%)	640 (20.2)	46,324 (30.7)	314,511 (17.8)
31–45, N (%)	640 (20.2)	36,295 (24.1)	297,333 (16.9)
46–64, N (%)	897 (28.3)	37,276 (24.7)	418,268 (23.7)
>65, N (%)	462 (14.5)	14,045 (9.3)	299,815 (17.0)
Sex, N (%)
Male	1,534 (48.4)	79,180 (52.5)	840,193 (47.7)
Female	1,639 (51.6)	71,534 (47.5)	922,199 (52.3)
Unknown	–	12 (0.01)	156 (0.01)
Race,aN (%)
White	2,884 (90.9)	132,942 (88.2)	1,587,810 (90.1)
African American	208 (6.6)	13,896 (9.2)	100,816 (5.7)
Asian	25 (0.8)	687 (0.5)	25,961 (1.5)
Other	56 (1.8)	3,201 (2.5)	47,961 (2.7)
Ethnicity, N (%)
Non-Hispanic	2,929 (92.3)	139,157 (92.3)	1,617,259 (91.8)
Hispanic	216 (6.8)	9,835 (6.5)	112,801 (6.4)
Unknown	28 (0.9)	1,734 (1.2)	32,488 (1.8)
Type of other insurance at the most recent encounter, N members (%)
Geisinger Health Plan	1,122 (35.4)	49,240 (32.7)	439,333 (24.9)
Commercial	679 (21.4)	38,008 (25.2)	645,661 (36.6)
Medicare	702 (22.1)	21,420 (14.2)	321,307 (18.2)
Medicaid	438 (13.8)	24,585 (16.3)	141,901 (8.1)
Self-pay	161 (5.1)	13,573 (9.0)	158,200 (9.0)
Other/Unknown	71 (2.2)	3,900 (2.6)	56,146 (3.2)
BMIb
Mean (SD)	28.9 (9.1)	28.5 (8.6)	26.9 (9.0)
Median (IQR)	28 (10.4)	27.3 (10.2)	26.4 (10.8)
<25.0, N (%)	1,000 (31.5)	49,272 (32.7)	539,818 (30.6)
25.0–29.9, N (%)	794 (25.0)	34,503 (22.9)	316,728 (18.0)
≥30.0–34.9, N (%)	615 (19.4)	23,132 (15.3)	215,337 (12.2)
>35.0, N (%)	630 (19.9)	25,734 (17.1)	206,634 (11.7)

Breakdown of demographic characteristics in Geisinger patients from the annotated sample (training dataset) of 3,173 unique patients, 150,726 patients positively classified as cannabis users from a total analytical sample of 370,885 unique patients, and the studied sample of more than 1.7 million unique patient at the time of either first CU documentation, or first encounter documented within the electronic health records (EHRs) during the study period (January, 2013–June, 2022).

BMI, body mass index; CU, cannabis use; SD, standard deviation.

**p value < 0.0001.

^aPatients could select multiple racial categories at the time of self-reporting.

^bMost recent BMI data at the time when the patient was first flagged to be included in the above categories was not available for all patients. Annotated sample 134 (4.2%), CU 18085 (12%), studied sample 488031 (28%).

Fig. 2.

Unadjusted prevalence of tobacco, alcohol, and illicit substances among cannabis users and the studied sample. Increased unadjusted prevalence of smoking tobacco (9.7-fold), drinking alcohol (9.8-fold), and illicit substance use (8.7-fold) among cannabis users, relative to the studied sample, identified with NLP of EHRs among Geisinger patients seen between 2013 and 2022.

Discussion

This study used NLP algorithms to identify patients within an integrated health system over a 10-year period who had CU documented in their EHR notes. This novel NLP application builds on prior reports [26–32] and demonstrates that CU can be identified with high precision and recall from unstructured data sources using advanced NLP models. While individual model performance varied across cannabis-related categories, the models generally performed well. Bio-ClinicalBERT, a transformer-based model, and a more traditional machine learning model, the SVM, performed nearly equally well for CU identification. Despite variations in each model’s performance for the other classification tasks, such as medical, past, current, and patient use, Bio-ClinicalBERT, and SVM consistently outperformed BERT and LR. Based on these models, the prevalence of CU-related documentation was 8.6% in the studied cohort curated using only the keyword search strategy.

EHR notes and other non-discrete fields are a highly sought option for documenting cannabis and non-allopathic substances and medicines, thereby making them a key source of information for understanding utilization patterns [27, 42–44]. The current analysis provides the first NLP-driven, overarching insight into unsegmented EHR-based unstructured notes data on CU documentation within a single health system. This study also includes a diverse cross-section of racial breakdown in rural (e.g., Sullivan County, population 5,845) and urban (e.g., Luzerne County, population 326,496) populations (online suppl. Fig. 2d). The use of traditional and transformer-based NLP models to identify CU documentation within different types of EHR notes and other unstructured data sources is rapidly increasing [22, 23, 26, 29, 45, 46], and models like ours could be applied and adopted across different health systems. This investigation's rigorous manual annotation phase, with a high inter-rater reliability (kappa > 0.83), also ensured consistency in the classification of CU. Furthermore, because the models were trained on various EHR notes spanning different specialties and disease states, as well as documenting providers, the algorithms could be easily applied to further targeted clinical and epidemiological research across the system and across different disease states.

A key report from an integrated health system in Washington state determined that for every person with past-year marijuana use documented in their EHR, seven others used marijuana, but who would only disclose this on a survey [20]. CU identified in our study over 9.5 years (8.6%) was less than half that reported in PA in the 2021–2022 National Survey of Drug Use and Health among those aged ≥12 years for marijuana use in the past year (19.2%) [47], affirming similar incomplete documentation as noted in Washington State. Despite the recent push for cannabis being reclassified as a schedule III substance [48] and potentially soon to be reported through the state’s prescription drug monitoring program, self-reported CU information will continue to be recorded in the EHR. This is because medical use accounts for only a small portion of the total use among high schoolers [1, 2] as well as adults [20], and only 24 states, three territories, and Washington, D.C. have legalized adult use, thereby resulting in a continued reliance on self-reported CU information for assessing prevalence [5, 49]. A hesitancy to disclose CU is perhaps understandable as patients might perceive that this documentation could negatively impact one’s ability to possess a firearm [50], employment [51], or even child custody [52]. The authors of an EHR report in California from 2013 to 2017 (i.e., before adult use passage) of primary care patients interpreted their social history CU documentation (0.38%) as “surprisingly low” [53] and at least ten-fold below what might be anticipated based on past-year use. Even within this context of persistent cannabis stigma [54] and under-reporting of tobacco [53] and alcohol use (online suppl. Figs. 3, 4), the pronounced higher rates of tobacco, alcohol, and other illicit drug use among CU relative to the studied sample are clear. The higher prevalence of CU tobacco, alcohol, and other illicit drug users aligns with findings from the literature [55, 56], there is also some evidence that indicates that CU is more likely to be documented in the EHR for substance users compared to nonusers, indicating potential biased approach towards not just who is asked or tested for CU, but also when it is documented [29]. Prior works have found that patient characteristics such as age, race, ethnicity, gender, history of illicit drug use, and mental health condition affect whether the patient is screened for substance use. Similarly, provider characteristics such as comfort addressing substance use, years of practice, specialty, and preconceived biases all affect who and how a patient is screened as well as their level of comfort in discussing substance use with a clinician [57–60]. We could potentially also hypothesize that CU is higher among those who report illicit drug use, in part, because individuals who already disclose illicit drug use may feel less need to hide or underreport their CU. However, the prevalence estimates reported here are unadjusted for any demographic or clinical covariates, and hence more research is needed to validate these findings and assumptions.

Demographics and clinical characteristics differed between our annotated sample, the studied sample, and CU patients. However, the clinical interpretation of these findings may not be clear due to potential confounding, given the large sample sizes, unadjusted analysis, and focus on having an equitable distribution of CU-related terminologies in the annotated sample over patient characteristics [61]. For example, in our study, we found a higher proportion of obese patients in the CU cohort, which is in contrast with other studies, which found that compared to nonusers, the prevalence of obesity and being overweight was lower among cannabis users in the US compared to nonusers due to the increase in metabolic rates and consequent reduction in BMIs [62–65]. However, in our analysis, we did not stratify patients based on length and frequency of CU, nor did we adjust for the higher prevalence of alcohol intake in this group or potential clinical comorbidities that may be associated with high obesity prevalence, all of which may have influenced this difference in prevalence observed [64, 66]. Additionally, most recent BMI data at the index date were not available for all patients across the board in the annotated (4.2%), CU (12%), and studied samples (28%). This varying level of missingness could also have further impacted the findings, and hence the findings should be interpreted with caution. Furthermore, we observed a higher prevalence of CU documentation among African Americans and a lower prevalence among Asians. While these findings align with prior research, there is existing evidence that indicates that preconceived biases, stigma, and racial profiling and discrimination surrounding who is asked about or tested for CU may be a reason behind the higher prevalence observed [29, 67]. It is also pertinent to note that for the purpose of this analysis, we did not delve into how patient characteristics varied based on the type of cannabis that was used, for example, CBD users versus THC users, or marijuana users versus spice users, since that was beyond the scope of this work, but should be explored in subsequent research initiatives.

The EHR landscape is rapidly evolving, with the diffusion of agentic and generative artificial intelligence-based tools implemented across multiple platforms to collect and transform data, predict potential health outcomes, and improve documentation. As the landscape continues to evolve, healthcare systems will need to develop internal capacity to tailor and fine-tune universal models to meet local and system-specific needs. In the context of CU, models developed as part of this study raise the potential for this technology to improve the current capture of CU within EHR systems and to fine-tune newer models in the future. However, it is also noteworthy that the NLP models in the current analysis performed comparatively poorly while classifying medical use of cannabis, with F1 scores ranging from 45% to 68%, indicating that the algorithms had difficulty consistently distinguishing medical from nonmedical use. One possible explanation is that documentation of medical CU within EHR notes was often not explicit or standardized and may instead have relied on subtle contextual indicators, such as references to pain relief, anxiety, sleep, or symptom management. These contextual distinctions may be more readily interpreted by trained human annotators than by NLP models trained on relatively limited annotated data. Additionally, the distinction between medical and nonmedical CU may not exist as a strict dichotomy in practice. Among primary care patients in Washington state who indicated CU, 40.1% categorized their use as medical, 31.8% as nonmedical, and 28.2% as both medical and nonmedical [20]. Similarly, among medical patients in New England surveyed before the passage of laws for legalizing medical use, about half described their use as on a continuum that involved at least some recreational use (7, online suppl. Fig. 4) [17]. Among adults living in a state that condones medical use, the most common source of cannabis (57.5%) was from a non-dispensary source [68]. Together, these findings suggest that overlap between medical and nonmedical use patterns, combined with variability in clinical documentation, may have limited the models’ ability to reliably classify medical CU.

Our investigation had several limitations. First, the results from this study are based on data collected from a single health system. CU documentation within a system could be influenced by the legalization status of cannabis at the time of data collection (i.e., currently condoning medical but not recreational in PA) [44], health system-level regulations, equity concerns about cannabis documentation attributable to racial diversity (online suppl. Fig. 2d) [29, 69], and clinician-level characteristics, including but not limited to personal biases. Second, while using fivefold cross-validation improved the internal validity of these models, they were not externally validated, thereby limiting their generalizability for use at external institutions. Third, we identified the EHR notes sample using a keyword search for terminologies. There is still a possibility that some terminologies, variations, or misspellings could have been missed (e.g., “marijuana” can be misspelled as “marihuana” or “marjiuana” or “marajuana”), impacting the number of notes identified and restricting the sample. We curated literature and sought stakeholder input to assemble a comprehensive lexicon that would minimize the impact of this limitation. Additionally, based on our models, we could not conclude medical use or distinguish between current and past use with certainty due to poor model performance. This could be attributed to semantic variability in the documentation of this information, low coder agreement during the manual annotation phase when classifying these factors, and potential loss of additional contextual knowledge at the time of note trimming. Prior research has encountered similar difficulties when leveraging NLP algorithms [26, 27]. Consistency in documenting practices across the system may help address ambiguity in reported information and improve model performance. Lastly, data captured in the EHR remains susceptible to the general limitations of EHR data, such as inaccuracy, incompleteness, and insufficient granularity [70]. System-level initiatives that ensure regular data quality checks could help overcome this.

Conclusion

Our study successfully leveraged NLP methodologies to classify CU documentation, when recorded, from available EHR notes with high precision. System-level initiatives supporting automated extraction of this information using NLP and its potential transformation into data that can be discretely queried, as well as leveraging the rapidly evolving artificial intelligence capabilities within the EHR, may lead to an improved understanding of patients’ health and facilitate future clinical and epidemiological research surrounding CU prevalence and patterns among different patient populations.

Acknowledgments

We would like to thank our Summer Research Immersion Program student, Ms. Alivia Roberts, and Ms. Faith Garasich, a Summer Undergraduate Research Program student, for their assistance in the manual annotation work on this study, and Wei Wei, PhD, and Joseph J. Dewalle, for fruitful discussions, as well as Megan Gunther and Iris Johnston for technical support. Vanessa Troiani is currently the director of Geisinger’s Academic Clinical Research Center, which operates through the Center for Substance Use Research and Education.

Statement of Ethics

This study was performed in accordance with the Declaration of Helsinki. This human study was approved by Geisinger Institutional Review Board – Approval No: 2022-0498. Parent, guardian or next of kin consent was not required for the minors because consent was waived due to the following conditions being met: waiver approved under 45 CFR 46.117 (c) 1;waiver of HIPAA authorization for research approved under 45 CFR164.512 (i) (2) (ii); and waiver/alteration approved 46.116(c) or (d). Adult participant consent was not required because consent was waived due to the following conditions being met: (a) waiver approved under 45 CFR 46.117 (c) 1; (b) waiver of HIPAA authorization for research approved under 45 CFR164.512 (i) (2) (ii); and (c) waiver/alteration approved 46.116(c) or (d).

Conflict of Interest Statement

The authors had no conflicts of interest to report during the study period.

Funding Sources

The study is funded by Ascend Wellness Holdings, Clinical Registrant, through the Geisinger Academic Clinical Research Center. The funder was not involved in the study or the decision to submit for publication.

Author Contributions

Apoorva M. Pradhana, Vishal A. Shetty, Christina M. Gregor, Jove H. Graham, Lorraine Tusing, Annemarie G. Hirsch, Eric Hall, Vanessa Troiani, Mellar P. Davis, Donielle L. Beiler, Katrina M. Romagnoli, Chadd K. Kraus, Brian J. Piper, Eric A. Wright contributed to the study design and conceptualization of the manuscript. A.M.P., B.J.P., and V.A.S. drafted the initial manuscript. C.M.G., B.J.P., and A.M.P. drafted the supplemental material. A.M.P. developed the initial lexicon that was revised, edited, and approved by J.H.G., B.J.P., V.A.S., M.P.D., and E.A.W. A.M.P., V.A.S., and J.H.G. performed the data preprocessing. A.M.P. and J.H.G. performed the data extraction. V.A.S. drafted the initial annotation guidelines, which were revised, approved, and finalized by A.M.P., B.J.P., L.T., E.A.W., and C.M.G. C.M.G., A.M.P., and V.A.S. performed manual annotation. V.A.S. performed inter-rater reliability analysis. A.M.P. and V.A.S. performed the final data analysis. A.M.P., V.A.S., C.M.G., L.T., B.J.P., J.H.G., and E.A.W. revised and approved the results and conclusions. A.M.P., V.A.S., C.M.G., J.H.G., L.T., A.G.H., E.H., V.T., M.P.D., D.L.B., K.M.R., C.K.K., B.J.P., and E.A.W. edited, revised, and approved the final manuscript.

Funding Statement

The study is funded by Ascend Wellness Holdings, Clinical Registrant, through the Geisinger Academic Clinical Research Center. The funder was not involved in the study or the decision to submit for publication.

Data Availability Statement

The data that supports the findings of this study are not publicly available due to privacy reasons. However, data and models may be shared upon reasonable request in the future, contingent on IRB approval. Specific requests for access to the data should be directed to the corresponding author, who will facilitate the process, contingent upon approval and adherence to the necessary ethical and institutional guidelines.