CAN-STRESS: A Real-World Multimodal Dataset for Understanding Cannabis Use, Stress, and Physiological Responses

This paper introduces CAN-STRESS, a publicly available multimodal dataset comprising physiological data from Empatica E4 wristbands and self-reported surveys from 82 participants, designed to investigate the distinct physiological stress responses of cannabis users compared to non-users in real-world settings.

The Gist

Imagine you're trying to understand why some people reach for a cannabis plant when they feel stressed, while others don't. For years, scientists have studied this in sterile, quiet laboratories—like watching fish in a tiny, artificial tank. But real life is messy, loud, and full of unexpected events. To truly understand how cannabis affects our bodies in the wild, we needed to let the fish swim in the ocean.

That's exactly what this paper does. It introduces CAN-STRESS, a new, massive "ocean" of data designed to help us understand the real-world relationship between stress, cannabis, and our body's physical reactions.

Here is the story of how they did it, explained simply:

1. The "Black Box" Wristbands

The researchers gave 82 volunteers (about half cannabis users, half non-users) a special smartwatch called the Empatica E4. Think of this watch as a 24-hour body spy.

While the volunteers went about their normal lives—working, sleeping, exercising, and yes, using cannabis—the watch secretly recorded their internal "vital signs" every second:

• The Heartbeat: How fast their heart was racing.
• The Sweat: Even if you didn't feel sweaty, the watch measured tiny electrical changes in the skin (like a lie detector test) that happen when we get stressed or excited.
• The Temperature: How their skin temperature changed.
• The Movement: Whether they were running, sitting, or sleeping.

2. The "Diary" Connection

A watch alone is just a bunch of numbers. To make sense of them, the volunteers also kept a digital diary. They tapped their phones to say:

• "I just used cannabis."
• "I just woke up."
• "I just finished a workout."
• "I'm feeling super stressed right now (rate 1 to 10)."

By stitching the watch data and the diary together, the researchers created a perfect map. Now, they could look at a specific moment in the diary (e.g., "Cannabis use at 2:00 PM") and instantly see what the body was doing at that exact second.

3. The Big Discovery: The "Stress Alarm"

When the researchers looked at the data, they found some interesting patterns.

• The Baseline: Cannabis users generally reported feeling more stressed and had higher heart rates and skin reactions (EDA) compared to non-users.
• The AI Detective: The team built a computer program (a machine learning model) to act like a detective. They fed it the body data and asked, "Is this person a cannabis user or not?"

The result? The computer was shockingly good at guessing. It got it right 96% of the time.

It's as if the computer learned to recognize the unique "fingerprint" of a cannabis user's body. The most important clues? Heart rate and skin conductance (the sweat sensor). It turns out, the way these users' bodies react to stress is distinct enough that a computer can spot it just by looking at their wrist data.

4. Why This Matters

Before this, we only had data from people sitting in a lab, which is like trying to learn how to drive by sitting in a parked car. CAN-STRESS is the first time we've let people drive the car in real traffic.

• For Doctors: It helps them understand how cannabis changes the body's stress system over time.
• For Researchers: It's a free, public library of data that anyone can use to build better health apps or study addiction.
• For Everyone: It moves us away from guesswork and toward real, scientific evidence about how cannabis interacts with our daily lives.

The Bottom Line

Think of this paper as opening a window into the real world. By combining high-tech wristbands with honest self-reports, the researchers created a powerful tool that shows us exactly how cannabis users' bodies handle stress differently than non-users. It's a giant leap forward in understanding the "why" and "how" of cannabis use, moving us from the lab bench to the living room.

Technical Summary

Here is a detailed technical summary of the paper "CAN-STRESS: A Real-World Multimodal Dataset for Understanding Cannabis Use, Stress, and Physiological Responses."

1. Problem Statement

• Context: Chronic cannabis use is frequently linked to stress coping. It is hypothesized that cannabis users exhibit distinct physiological stress responses compared to non-users, particularly during consumption.
• Gap: Existing research is predominantly conducted in controlled laboratory settings, which lack ecological validity (real-world context). There is a scarcity of publicly available, high-quality datasets that capture the interplay between cannabis use, stress, and physiological responses in naturalistic daily environments.
• Objective: To bridge this gap by creating and releasing a comprehensive, real-world multimodal dataset to test hypotheses regarding physiological stress regulation in cannabis users versus non-users.

2. Methodology

A. Data Collection & Participants

• Dataset Name: CAN-STRESS.
• Participants: 82 adults (39 chronic cannabis users, 43 non-users).
  • Users: Daily/near-daily use (≥4 times/week for ≥1 year).
  • Non-users: <10 lifetime uses, none in the past year.
  • Exclusion: Neurological disorders, psychosis, heavy alcohol use, recent illicit drug use, nicotine, or corticosteroid use.
• Hardware: Empatica E4 wristbands (medical-grade wearables).
• Duration: 24-hour continuous monitoring during daily routines.
• Data Modalities:
  1. Physiological Signals (Sensor Data):
     • EDA (Electrodermal Activity): 4 Hz (Skin conductance/stress).
     • BVP (Blood Volume Pulse): 64 Hz (Raw signal for HR/IBI).
     • HR (Heart Rate): 1 Hz (Derived from BVP).
     • ACC (Accelerometer): 32 Hz (Triaxial movement).
     • TEMP (Body Temperature): 4 Hz.
  2. Self-Reported Data (Questionnaire):
     • Timestamps for cannabis use, sleep, and exercise.
     • Perceived stress ratings (1–10 scale) at key moments.
     • Sleep patterns (bedtime/wake time).

B. Data Processing & Preprocessing

• Segmentation: Continuous signals were segmented into 15-minute sliding windows with 50% overlap.
• Filtering: Analysis was restricted to waking hours to minimize confounding effects from sleep-related physiological changes.
• Feature Extraction: 31 physiological features were extracted per window using the NeuroKit2 package:
  • EDA: Tonic/phasic decomposition, peak counts, amplitude statistics.
  • BVP/HR: Heart Rate Variability (HRV) metrics (RMSSD, pNN50, SDNN) and basic HR stats.
  • TEMP: Mean, std, range, and linear slope.
  • ACC: Vector magnitude, activity intensity, and direction statistics.
• Normalization: Subject-wise standardization (mean=0, std=1) to account for baseline physiological differences.

C. Machine Learning Pipeline

• Task: Binary classification (Cannabis User vs. Non-User).
• Model Architecture: Multi-Layer Perceptron (MLP) with three hidden layers (128, 64, 32 neurons), Batch Normalization, ReLU activation, and Dropout (0.2).
• Evaluation Strategy: Leave-One-Subject-Out (LOSO) cross-validation.
  • Personalization: A transfer learning approach was used where the model was fine-tuned on 50% of the test subject's data before evaluating on the remaining 50%.
• Explainability: SHAP (Shapley Additive Explanations) values were used to determine feature importance.

3. Key Contributions

1. Dataset Release: Public release of the CAN-STRESS dataset, one of the largest publicly available multimodal datasets capturing physiological characteristics of cannabis users in naturalistic settings.
2. Comprehensive Documentation: Detailed description of data modalities, collection protocols, and file structures to facilitate reproducibility.
3. Baseline Benchmark: Establishment of a machine learning baseline demonstrating the feasibility of distinguishing users from non-users using passive sensing.
4. Comparative Analysis: Statistical analysis highlighting differences in stress ratings, EDA, and heart rate between the two groups.

4. Results

A. Descriptive Statistics

• Stress & Physiology: Cannabis users reported higher self-reported stress ratings compared to non-users.
• Physiological Markers: Users exhibited elevated Electrodermal Activity (EDA) and higher Heart Rate (HR) values on average compared to non-users.

B. Machine Learning Performance

The model successfully classified users vs. non-users with high accuracy:

• Test Accuracy: 95.96% (±0.06)
• Test F1-Score: 95.92% (±0.06)
• Precision: 97.82%
• Recall: 100.00%
• Note: Training accuracy was near perfect (99.93%), indicating the model learned the patterns effectively without severe overfitting given the regularization and LOSO strategy.

C. Feature Importance (SHAP Analysis)

• Top Predictors: Heart Rate (HR) and Electrodermal Activity (EDA) were the most influential features.
• Specific Features:
  1. hr_max (Maximum Heart Rate)
  2. eda_min, eda_mean, eda_phasic_mean, eda_max
  3. HRV metrics (hrv_sdnn, hrv_rmssd)
• Interpretation: The results confirm that stress-related physiological markers (skin conductance and cardiovascular dynamics) are the primary differentiators between chronic users and non-users in real-world settings.

5. Significance

• Ecological Validity: Moves beyond lab constraints to study cannabis effects in real-life contexts, capturing the nuance of stressors and daily activities.
• Clinical & Policy Impact: Provides empirical data to inform healthcare policies and individual decision-making regarding cannabis use and stress management.
• Research Utility: Serves as a foundational resource for:
  • Stress Detection: Validating wearable-based stress algorithms.
  • Activity Recognition: Understanding how cannabis use correlates with physical activity and sleep.
  • Medical Diagnostics: Developing passive monitoring tools for substance use disorders.
• Community Resource: The dataset is openly available (via Zenodo), encouraging further computational health research and the development of more sophisticated models for behavioral science.