Preliminary study: quantification of chronic pain from physiological data

Zhuowei Cheng, Franklin S. Ly, Tyler Santander, E. Turki, Yun Zhao, Jamie Yoo, Kian Lonergan, Jordan Gray, Christopher H. Li, Henry T. Y. Yang, Michael B. Miller, Paul K. Hansma, Linda PetzoldPAIN 2022 · DOI

TL;DR

This study presents a novel approach to quantifying chronic pain by analyzing physiological data collected from wearable sensors. The researchers demonstrate that machine learning models can predict pain scores with moderate accuracy, offering a potential pathway towards objective, low-cost chronic pain assessment beyond subjective self-reports.

What Problem Does This Paper Address?

Chronic pain is a pervasive and costly public health issue, affecting millions and imposing significant personal and economic burdens. Despite its prevalence, quantifying chronic pain remains a significant challenge, largely relying on subjective self-reports from patients. This subjectivity can lead to diagnostic delays, undertreatment, or overtreatment. While advanced neuroimaging techniques like fMRI show promise in identifying physiological correlates of pain, their high cost and complexity limit their routine clinical use. This paper aims to bridge this gap by investigating whether inexpensive, readily available physiological sensors, combined with machine learning, can provide a more objective and accessible method for quantifying chronic pain.

Methods

The study employed a custom-built "Pain Meter" prototype, designed for home use, to collect physiological data. This device avoided invasive sensors, relying on non-contact methods. Key physiological signals measured included:

  • Pulse: Using photoplethysmogram (PPG) sensors placed on the head, neck, wrist, and finger.
  • Temperature: Using thermistors at various locations (temple, forehead, wrist, finger).
  • Motion: Using accelerometer and gyroscope sensors.
  • Force: Using force sensors on the carotid and basilar arteries.

Data was collected in 10-second windows. The core methodological steps involved:

  1. Preprocessing: Removing unstable segments of data (e.g., due to movement) and segmenting stable data into 10-second samples.
  2. Feature Extraction: Deriving features from pulse signals (e.g., pulse width parameters like LR, LF, PPIH, PPIL, and their standard deviations) and temperature readings (mean and standard deviation, as well as ratios and differences between different temperature locations). Features were z-score normalized.
  3. Model Training:
    • Individual-level models: For each subject, recursive feature elimination was used to select the top 5 features, followed by fitting a linear model to predict pain scores. This approach aimed to capture subject-specific physiological responses.
  4. Population-level model: A random forest regression model was employed to predict pain scores across all subjects. This ensemble method is known for its robustness and ability to handle high-dimensional data.
  5. Evaluation: Performance was assessed using leave-one-recording-out cross-validation. Metrics included Pearson correlation coefficient (r), intraclass correlation coefficient (ICC), and root mean square error (RMSE). Nonparametric permutation tests were used to assess model performance against chance. Bland-Altman analysis was conducted for the population-level model to assess agreement.

Datasets

The study utilized a newly collected dataset from 12 participants (9 women, 3 men) who self-reported their chronic pain levels on a visual analogue scale (0-10). A total of 183 10-minute recordings were collected, with participants recording their pain twice daily for two weeks. This dataset, collected in participants' homes, aimed to capture natural variations in pain levels. The paper does not explicitly state the use of any pre-existing benchmark datasets for training or evaluation.

Key Insights and Ideas

  • Physiological Correlates of Chronic Pain Exist: The study provides the first demonstration that physiological data, collected non-invasively, can be correlated with subjective reports of chronic pain. This is a significant step towards objective pain measurement.
  • Individualized vs. Population Models: The research highlights the trade-offs between personalized models (which can capture unique physiological responses) and generalized models (which offer broader applicability). While individual models showed promise for some subjects, a population-level model demonstrated more consistent, albeit moderate, predictive power.
  • Importance of Multivariate Analysis: Consistent with prior research, the findings suggest that combining multiple physiological signals and features is more effective than relying on single parameters for pain quantification.
  • Potential for Low-Cost, Accessible Pain Assessment: The use of inexpensive sensors and a home-based data collection paradigm points towards the feasibility of developing a "chronic pain meter" that could be widely accessible to patients and clinicians.
  • Limitations of Subjective Reporting: The study implicitly underscores the limitations of current subjective pain reporting by demonstrating that objective physiological measures can offer complementary information.

Key Results

  • Individual-level Models: For 5 out of 12 subjects, moderately strong predictive accuracies were achieved, with intraclass correlation coefficients (ICC) ranging from 0.46 to 0.75.
  • Population-level Model: The random forest regression model achieved an ICC of 0.58 and a Pearson correlation coefficient (r) of 0.60 (P < 0.001) between reported and predicted pain scores.
  • Model Tendencies: The population-level model tended to overestimate lower pain scores and underestimate higher pain scores, as indicated by Bland-Altman analysis.
  • Sensor Performance: Models using only the pointer finger sensor or only the temple pulse sensor showed the strongest correspondences (r = 0.60, RMSE = 1.52) in the population-level analysis.

Limitations and Future Work

The authors acknowledge several limitations:

  • Limited Data: The study was conducted with a relatively small sample size (12 participants) and a limited duration of data collection, partly due to the COVID-19 pandemic.
  • Imbalanced Data: Extreme pain scores (0, 8, and 9) had fewer samples, potentially impacting model performance at these ends of the spectrum.
  • Approximation in Data Assignment: Assigning pain scores to the first/second half of recordings is an approximation, as pain levels can fluctuate within a session.
  • Sensor Optimization: The current sensors have limitations (e.g., filtering out low-frequency pulse variations).

Future work should focus on:

  • Larger and More Diverse Datasets: Collecting data from a broader range of chronic pain conditions, demographics, and durations.
  • Sensor Optimization and Expansion: Exploring more advanced sensors, including ECG, electromyography, and skin impedance sensors, while carefully considering safety for unsupervised home use.
  • Advanced Machine Learning Models: Investigating more sophisticated algorithms for feature selection and prediction.
  • Clinical Validation: Moving towards validating the approach in clinical settings to assess its utility for diagnosis and treatment monitoring. The ultimate goal is to develop a low-cost, easy-to-use system for objective chronic pain measurement, which could significantly improve pain management and reduce issues related to medication.

Summary updated

Abstract

This is the first demonstration that physiological data can be correlated with chronic pain, both for individuals and populations. Further research and more extensive data will be required to assess whether this approach could be used as a "chronic pain meter" to assess the level of chronic pain in patients.
BibTeX
@inproceedings{cheng2022preliminary,
  author = {Zhuowei Cheng and Franklin S. Ly and Tyler Santander and E. Turki and Yun Zhao and Jamie Yoo and Kian Lonergan and Jordan Gray and Christopher H. Li and Henry T. Y. Yang and Michael B. Miller and Paul K. Hansma and Linda Petzold},
  title = {{Preliminary study: quantification of chronic pain from physiological data}},
  booktitle = {PAIN},
  year = {2022},
  doi = {10.1097/pr9.0000000000001039},
}
Figure