Noninvasive and Objective Near Real-Time Detection of Pain Changes During Tonic Fluctuating Noxious Heat Stimulation

This study demonstrates that deep learning models utilizing easily obtainable physiological signals, particularly electrodermal activity and heart rate, can objectively detect spontaneous decreases in fluctuating tonic pain with high accuracy and low latency, thereby enabling noninvasive, near real-time closed-loop interventions for chronic pain management.

The Gist

Imagine you are in a room where the temperature is constantly changing. Sometimes it gets uncomfortably hot, and sometimes it cools down a little bit. For someone with chronic pain, life feels a bit like this room: the pain isn't a constant, unchanging wall; it's a wave that goes up and down. The problem is, patients often feel helpless because they can't control when the pain spikes or when it dips. They feel like they are just passengers on a bumpy ride.

This study is like a team of engineers trying to build a "Smart Thermostat for Pain." Their goal? To create a device that can instantly tell when the pain wave is going down, so a patient can press a button and feel like they caused the relief. This creates a "magic trick" of control: even if the pain was going to go down anyway, the patient feels like their action made it happen, which helps them feel less helpless.

Here is how they built this "Smart Thermostat," explained in simple terms:

1. The Experiment: The "Hot Plate" Game

The researchers didn't test this on people with chronic pain yet; they started with healthy volunteers. They put a special heating pad on the volunteers' arms.

• The Setup: The pad would heat up to a "ouch" level, then cool down, then heat up again, all in a random, unpredictable pattern for 3 minutes at a time.
• The Goal: They wanted to see if a computer could look at the volunteers' bodies and say, "Aha! The pain is about to get better!" before the person even realized it.

2. The Sensors: The Body's "Secret Messengers"

To detect these changes, the researchers hooked the volunteers up to various sensors. Think of these sensors as spies reporting on what the body is doing:

• Skin Conductance (EDA): This measures how sweaty your palms get. When you feel pain or stress, your body's "fight or flight" system turns on, and your skin gets slightly more conductive (like a wet sponge). This was the star player of the team.
• Heart Rate: How fast your heart is beating.
• Pupil Size: How big your pupils get (they often dilate when you are in pain or stressed).
• Brain Waves (EEG): A cap with electrodes to listen to the brain.
• Facial Expressions: A camera watching for frowns or grimaces.

3. The Brain: The "AI Detective"

The researchers fed all this data into a super-smart computer program (Deep Learning). They taught the AI to look for patterns.

• The Challenge: The AI had to distinguish between "pain is getting worse" and "pain is getting better."
• The Result: The AI got really good at this! It achieved an accuracy of about 85% in detecting when pain was dropping.

4. The Big Surprise: Simplicity Wins

You might think the most expensive, high-tech sensor (the brain cap/EEG) would be the best. But the study found something surprising: The brain waves were actually the worst at predicting pain changes. They were too noisy and varied from person to person.

Instead, the winning team was the "Low-Tech Trio":

1. Skin Conductance (Sweat)
2. Heart Rate
3. Pupil Size

The AI realized that when pain starts to fade, your skin conductance and heart rate drop in a very specific, predictable way. It's like the body's "relief signal" is much clearer in your sweat and heart than in your brain waves.

5. The Speed: "Near Real-Time"

The AI didn't just work; it worked fast. It could detect a drop in pain about 5 to 6 seconds after it started.

• Why this matters: In the real world, if a patient feels pain dropping, they want to press a button immediately to feel like they did it. If the machine takes 30 seconds to say, "Hey, pain is dropping," the moment has passed. A 5-second delay is fast enough to make the "magic trick" feel real.

6. The Catch: Everyone is Different

The study found that while the "Low-Tech Trio" worked well on average, every person's body is unique.

• For some people, their heart rate was a great indicator.
• For others, their pupils were the best clue.
• Some people's faces didn't show any pain at all (they are "stoic"), while others grimaced a lot.

This means that in the future, these devices won't be "one size fits all." They will need to be personalized, learning specifically how your body reacts to pain.

The Bottom Line

This paper proves that we can build a simple, non-invasive device (using just a wristband for heart rate and skin sensors) that acts like a pain radar. It can spot when your pain is naturally taking a break.

If we can give this to chronic pain patients, they can use it to trigger treatments (like electrical stimulation) right at that moment. Even if the treatment didn't cause the pain to stop, the patient will feel like they are the hero who stopped the pain. This feeling of control is a powerful medicine in itself, helping to break the cycle of helplessness that comes with chronic pain.

In short: They built a digital "pain radar" using sweat and heart rate that can tell you when your pain is about to get better, giving you the power to feel in control of your own body.

Technical Summary

1. Problem Statement

Chronic pain is characterized by spontaneous, unpredictable fluctuations in intensity that patients cannot control, often leading to learned helplessness and functional impairment. Current objective pain assessment methods primarily focus on estimating static pain intensity from short, discrete, and predictable stimuli. There is a significant gap in the ability to objectively detect moment-to-moment changes (specifically decreases) in ongoing, fluctuating tonic pain in real-time.

The authors propose that detecting spontaneous pain decreases could enable "closed-loop" interventions (e.g., triggering TENS stimulation) that reinforce a patient's sense of control. However, this requires a system capable of:

• Operating noninvasively.
• Detecting pain dynamics (first derivative of intensity) rather than absolute levels.
• Functioning with minimal latency suitable for real-time application.
• Utilizing easily obtainable physiological signals.

2. Methodology

Experimental Design

• Participants: 42 healthy right-handed adults (mean age 26.2 years).
• Stimulus: A custom-designed tonic noxious heat stimulus applied to the left forearm.
  • Duration: 12 trials of 3 minutes each.
  • Pattern: Unpredictable sinusoidal temperature fluctuations between the individual's pain threshold and a "strong pain" level (VAS 70).
  • Key Feature: The stimulus included uniform "major decreasing intervals" (5–20 seconds) to serve as ground truth for pain decreases, alongside plateaus and increases.
• Data Collection: Multimodal physiological signals were recorded simultaneously:
  • Electrodermal Activity (EDA): Skin conductance (Shimmer3 device).
  • Heart Rate (HR): Photoplethysmography (PPG) via earlobe.
  • Pupil Diameter: Eye tracking (Smart Eye AI-X).
  • EEG: 8-channel scalp EEG (Neuroelectrics Enobio 8).
  • Facial Expressions: Video recording analyzed via AFFDEX software.
  • Ground Truth: Continuous subjective pain ratings via mouse movement on a digital VAS.

Data Preprocessing

• Causal Constraint: To ensure real-time applicability, all preprocessing was strictly causal (using only current and past data). This excluded non-causal techniques like deconvolution or symmetric filtering that would introduce data leakage.
• Signal Processing:
  • EDA, HR, and Pupil data were downsampled to 10 Hz.
  • EEG data was processed at 250 Hz with causal filtering (high-pass, notch).
  • Blink artifacts in pupil data were filled via forward propagation.
• Sample Creation:
  • Binary Classification Task: "Pain Decrease" vs. "Non-Decrease" (Increases/Plateaus).
  • Window Size: 7-second temporal windows.
  • Labeling: Decrease samples were shifted 1 second from the start of the temperature drop to account for physiological latency.
  • Dataset: 2,826 total samples (1,413 per class), balanced via downsampling.

Deep Learning Models

The study evaluated 11 feature sets (ranging from single modalities to full multimodal ensembles) across various architectures:

• Non-EEG Models: MLP, LightTS, TimesNet, PatchTST (Transformer with patch-based tokenization), and iTransformer.
• EEG Models: EEGNet, iTransformer.
• Ensemble Models: Feature-level fusion combining EEGNet (for EEG) and PatchTST (for peripheral signals).
• Training Strategy: Group-based splitting (60% train, 20% validation, 20% test) to prevent participant data leakage. Hyperparameter optimization was performed using Optuna.

3. Key Results

Model Performance

• Optimal Configuration: The combination of EDA, Heart Rate, and Pupil Diameter using the PatchTST architecture achieved the best performance.
  • AUROC: 0.854
  • Accuracy: 76.8%
  • Precision/Recall: 0.762 / 0.788
• Single Modality Performance:
  • EDA: The strongest single predictor (AUROC = 0.805).
  • Pupil Diameter: Second best (AUROC = 0.738).
  • Heart Rate: Poor discriminative power alone (AUROC = 0.557).
  • EEG & Facial Expressions: Demonstrated very low group-level performance (AUROC ~0.53–0.58), often performing near or below chance level.
• Multimodal Synergy: Adding EEG or facial expressions to the optimal EDA/HR/Pupil set did not improve performance and, in some cases, slightly degraded it, suggesting noise accumulation rather than information gain.

Continuous Stream Analysis (Real-Time Simulation)

• Latency: The model detected pain decreases with a median latency of 5.75 seconds (MAD = 1.50s) after the onset of the temperature drop.
• Sensitivity: At a conservative threshold (0.90), sensitivity was 70.4%.
• Trade-off: Lowering the threshold to 0.50 reduced latency to 4.25 seconds but increased false positives.
• Generalization: The model successfully detected decreases that were shorter and smaller in magnitude than those used during training.

Interindividual Variability

• Significant variability was observed across participants. While EDA and Pupil responses were relatively stable, EEG and facial expressions showed high variance, with some individuals performing below chance. This suggests that universal models for EEG/facial cues are ineffective, whereas peripheral autonomic signals offer more robust group-level detection.

4. Key Contributions

1. Shift from Static to Dynamic Assessment: The study moves beyond estimating absolute pain intensity to detecting pain dynamics (specifically decreases) in a fluctuating, tonic pain environment, which better mimics chronic pain conditions.
2. Real-Time Feasibility: Demonstrated that deep learning models trained on causally preprocessed data can achieve high accuracy suitable for near real-time deployment, overcoming the latency constraints of traditional signal processing.
3. Signal Hierarchy: Established a clear hierarchy of signal utility for this specific task: EDA > Pupil > HR >> EEG/Facial Expressions. It challenges the assumption that EEG is the gold standard for pain monitoring, showing it is highly variable and difficult to generalize without personalized fine-tuning.
4. Architectural Validation: Validated PatchTST (a transformer-based architecture using patch-based tokenization) as superior to CNNs and standard MLPs for processing raw, unengineered physiological time series in this context.

5. Significance and Future Implications

• Therapeutic Application: The findings provide a proof-of-concept for closed-loop pain interventions. By detecting spontaneous pain decreases, a device could trigger a treatment (e.g., TENS) at the precise moment pain is naturally subsiding. This creates an "illusion of control," potentially reinforcing self-efficacy and breaking the cycle of learned helplessness in chronic pain patients.
• Clinical Translation: The reliance on EDA and Heart Rate (measurable via commercial wearables) makes the solution highly practical for everyday use, unlike EEG or facial expression analysis which require specialized, intrusive equipment.
• Precision Medicine: The high interindividual variability in EEG and facial responses underscores the need for personalized models for those modalities, while peripheral autonomic signals may support more generalized "off-the-shelf" solutions.
• Future Work: The authors note that clinical validation in chronic pain populations is the necessary next step, as physiological signatures may differ significantly from experimentally induced pain in healthy subjects.

In summary, this paper establishes that noninvasive, objective, and near real-time detection of pain decreases is feasible using standard wearable sensors (EDA, HR, Pupil) and transformer-based deep learning, paving the way for intelligent, responsive pain management systems.