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

This study demonstrates that a transformer-based natural language processing model (Bio-ClinicalBERT) can effectively identify and classify cannabis use documentation within unstructured electronic health record notes with near-human accuracy, revealing that identified users have significantly higher rates of comorbid substance use and BMI compared to the general population.

The Gist

Imagine you are trying to find out how many people in a massive city are using cannabis. You have a giant library containing millions of medical records (Electronic Health Records, or EHRs) for every person who has visited a doctor in that city.

The Problem: The "Needle in a Haystack" with a Twist
In this library, doctors don't always fill out neat, tick-box forms saying "Patient uses cannabis: Yes/No." Instead, they write free-flowing notes, like a diary entry: "Patient mentions they use weed for back pain," or "No history of marijuana use," or even "Patient's joint pain is worse."

The word "joint" could mean a marijuana cigarette or a knee joint. The word "pot" could mean a cooking pot or a neti pot (for sinuses), or it could mean marijuana. Finding the real stories about cannabis use hidden inside these millions of messy, handwritten-style notes is like trying to find a specific needle in a haystack, but the needles are disguised as other objects, and the haystack is the size of a mountain.

The Solution: The "Super-Reader" Robot
The researchers in this paper built a team of "Super-Reader" robots (called Natural Language Processing, or NLP, models). Think of these robots as incredibly fast, super-smart interns who have read every book in the library.

1. Training the Interns: First, the researchers showed a small group of human experts how to read these notes and label them. They taught the experts to spot the difference between "I have a joint in my knee" (medical) and "I smoke a joint" (cannabis).
2. The Robot School: They then taught four different types of robots using these human labels.
   • Two robots were "traditional" (like a very organized librarian).
   • Two robots were "modern AI" (specifically Bio-ClinicalBERT). Think of Bio-ClinicalBERT as a robot that has read millions of medical textbooks and understands the context of words better than anyone else. It knows that "CBD" usually means a chemical in a bottle, but "CBD stricture" means a problem in the bile duct.

The Results: Who Found What?
The robots went to work scanning the entire library of 1.7 million patients.

• The Winner: The modern AI robot (Bio-ClinicalBERT) was the star student. It performed almost as well as the human experts, correctly identifying cannabis use about 92% of the time.
• The Discovery: The robot found that about 8.6% of the patients had cannabis use mentioned in their notes.
• The Pattern: When the researchers looked at the people the robot flagged as cannabis users, they noticed a pattern. These patients were more likely to:
  • Have a higher body weight (BMI).
  • Also use tobacco, alcohol, or other drugs (about 9 to 10 times more likely than the average person).

Why Does This Matter?
Think of the hospital system as a big ship. If the captain (the doctor) doesn't know the passengers are smoking cannabis, they might give them medicine that interacts badly with it, or miss a diagnosis.

Before this study, the ship's log (the computer system) was full of messy notes that the computer couldn't read. Now, thanks to this "Super-Reader" robot, the hospital can:

• Spot Risks: Automatically flag patients who might have dangerous drug interactions.
• Help Research: Understand how many people are actually using cannabis in the real world, not just those who fill out a survey.
• Improve Care: Help doctors have better conversations with patients about safe storage (like keeping it away from kids) or how it affects their other health conditions.

The Catch
The robot isn't perfect yet. It sometimes struggles to tell the difference between "using it for medical reasons" and "using it for fun," and it can't always tell if someone used it yesterday or ten years ago. Also, because it only looked at one health system (Geisinger in Pennsylvania), it might not work exactly the same way in a different city or country.

The Bottom Line
This paper is like a proof-of-concept that says: "We built a smart tool that can read messy doctor's notes and find hidden information about cannabis use." It turns a chaotic pile of paper into a clear, usable map, helping doctors and researchers understand the health landscape much better.

Technical Summary

1. Problem Statement

Despite the increasing prevalence of recreational and medical cannabis use (CU) in the United States, actionable data regarding CU is often missing or inaccessible within Electronic Health Records (EHRs).

• Data Fragmentation: CU information is frequently documented in unstructured free-text clinical notes (e.g., progress notes, nursing notes) rather than discrete, queryable fields.
• Clinical Impact: The lack of structured data hinders critical healthcare functions, including identifying drug-drug interactions, managing conditions like cannabis hyperemesis syndrome, and facilitating safe storage discussions.
• Research Gap: Prior Natural Language Processing (NLP) studies on CU have been limited to specific sub-populations (e.g., pediatric, surgical, or psychiatric patients), lacking a comprehensive approach applicable to a broad, unstratified population across an entire health system.

2. Methodology

The study utilized a multi-phase pipeline to develop and validate NLP models for extracting CU data from the EHR of Geisinger Health System (Pennsylvania) covering the period from January 1, 2013, to June 30, 2022.

A. Data Collection and Preprocessing

• Population: 1.76 million patients with at least one encounter.
• Lexicon Development: A curated list of 22 terms (e.g., "marijuana," "CBD," "THC," "weed," "joint") was used to filter notes.
• Data Cleaning:
  • Contextual Filtering: Notes were trimmed to 300 characters surrounding the keyword. Ambiguous terms were removed based on context (e.g., "pot" in "neti pot" or "hypotension," "CBD" in "Common Bile Duct," "joint" in "joint pain").
  • Negation Handling: Notes containing negations (e.g., "no cannabis use") were excluded.
  • Final Sample: The process yielded 2,790,896 EHR note snippets from 370,885 patients.

B. Manual Annotation (Ground Truth)

• Schema: Three annotators labeled notes into six categories based on time, context, and reliability:
  1. True mention (refers to cannabis)
  2. Patient use (refers to the specific patient)
  3. Indication of use (explicit mention of use)
  4. Medical use (for medical purposes)
  5. Current use
  6. History of use
• Reliability: Inter-rater reliability (Kappa) was >0.83 for most categories, except "History of use."
• Training Set: 3,650 unique annotated examples were generated.

C. Model Development

Four machine learning models were trained to perform binary classification for each of the six categories:

1. Logistic Regression (LR): Bag-of-words (unigrams, bigrams, trigrams) with Elastic Net regularization.
2. Support Vector Machine (SVM): Linear kernel with bag-of-words.
3. BERT: Pre-trained transformer model.
4. Bio-ClinicalBERT: A transformer model pre-trained specifically on biomedical and clinical text.

• Training Strategy: 5-fold cross-validation was used. Performance was evaluated using Precision, Recall, and F-score.
• Deployment: The best-performing model (Bio-ClinicalBERT) was applied to the entire unlabeled dataset to flag patients as "Cannabis Users" if they had positive classifications for both "True Mention" and "Indication of Use."

3. Key Contributions

• Population-Scale NLP Application: This is one of the first studies to apply NLP to identify CU across a massive, unstratified population (1.7M+ patients) rather than a specific disease cohort.
• Hybrid Model Evaluation: The study rigorously compared traditional machine learning (LR, SVM) against modern transformer-based models (BERT, Bio-ClinicalBERT) in a clinical setting.
• Comprehensive Annotation Schema: Developed a robust, multi-dimensional labeling framework that distinguishes between the mere mention of cannabis and the actual indication of patient use, current vs. past use, and medical vs. non-medical intent.
• Demographic and Clinical Profiling: Provided a detailed epidemiological profile of CU patients identified via NLP, comparing them against the general population.

4. Results

Model Performance

• Best Performer: Bio-ClinicalBERT achieved the highest overall performance, closely matching human annotation levels.
  • Weighted F-score: 92.4%
  • Precision: 91.4%
  • Recall: 93.3%
• Category Specifics:
  • Bio-ClinicalBERT excelled at "True Mention" and "Indication of Use."
  • SVM slightly outperformed Bio-ClinicalBERT in distinguishing "Medical Use" and "Current Use" (likely due to the complexity and semantic variability of these specific contexts).
  • All models struggled with "History of Use" and "Medical Use" (F-scores ranging 45–68%), highlighting the difficulty in distinguishing these nuances in free text.

Prevalence and Patient Characteristics

• Prevalence: The NLP model identified 150,726 patients (8.6%) as cannabis users. This is significantly lower than national survey estimates (e.g., 19.2% in PA), confirming the "under-documentation" bias in EHRs.
• Demographics:
  • Age: Similar mean age to the general population (37.6 vs. 38.1), but lower proportions of patients >65 and <18.
  • Gender: Slightly more male (52.5%) compared to the general population (47.7% male).
  • Race: Higher prevalence of African American patients and lower prevalence of Asian patients compared to the general population.
• Clinical Correlates:
  • BMI: Cannabis users had a higher mean BMI (28.5 vs. 26.9) and a higher rate of obesity (36.8% vs. 23.9%).
  • Substance Co-use: Cannabis users were significantly more likely to use other substances:
    • Tobacco: 10x higher (49.3% vs. 5.1%)
    • Alcohol: 10x higher (48.2% vs. 4.9%)
    • Illicit Drugs: 9x higher (4.7% vs. 0.5%)

5. Significance and Implications

• Clinical Utility: The study demonstrates that NLP can transform unstructured clinical notes into discrete, queryable data, enabling real-time clinical decision support (e.g., drug interaction alerts) and better patient identification.
• Epidemiological Insight: By capturing a broader dataset than surveys, the study reveals that while EHR documentation is incomplete, the patients who are documented show strong correlations with poly-substance use and higher BMI, offering new avenues for targeted interventions.
• Future Directions: The authors suggest that while current models are effective for general identification, future work must address the semantic ambiguity of "medical" vs. "recreational" use and "current" vs. "past" use. The study advocates for integrating these NLP tools into EHR workflows to improve data quality and support the evolving landscape of cannabis regulation (e.g., potential reclassification to Schedule III).

Limitations Noted: The study was single-system (Geisinger), limiting external generalizability. The models were not externally validated. The reliance on keyword search may have missed misspellings or slang not in the lexicon. The distinction between medical and recreational use remains semantically challenging for current algorithms.