[Paper Review] Automatic pain recognition from Blood Volume Pulse (BVP) signal using machine learning techniques

Fatemeh Pouromran, Yingzi Lin+1 more|arXiv (Cornell University)|Mar 19, 2023

Pain Mechanisms and TreatmentsMedicine3 citations

TL;DR

This study proposes an automated pain recognition system using Blood Volume Pulse (BVP) signals and machine learning, extracting 44 features (time-domain, frequency-domain, and nonlinear dynamics from BVP and IBI) to classify pain presence and intensity. XGBoost achieved the highest performance, with AUCs of 80.06%, 85.81%, and 90.05% for low, medium, and high pain vs. no pain, respectively, and 91% AUC for distinguishing medium from high pain.

ABSTRACT

Physiological responses to pain have received increasing attention among researchers for developing an automated pain recognition sensing system. Though less explored, Blood Volume Pulse (BVP) is one of the candidate physiological measures that could help objective pain assessment. In this study, we applied machine learning techniques on BVP signals to device a non-invasive modality for pain sensing. Thirty-two healthy subjects participated in this study. First, we investigated a novel set of time-domain, frequency-domain and nonlinear dynamics features that could potentially be sensitive to pain. These include 24 features from BVP signals and 20 additional features from Inter-beat Intervals (IBIs) derived from the same BVP signals. Utilizing these features, we built machine learning models for detecting the presence of pain and its intensity. We explored different machine learning models, including Logistic Regression, Random Forest, Support Vector Machines, Adaptive Boosting (AdaBoost) and Extreme Gradient Boosting (XGBoost). Among them, we found that the XGBoost offered the best model performance for both pain classification and pain intensity estimation tasks. The ROC-AUC of the XGBoost model to detect low pain, medium pain and high pain with no pain as the baseline were 80.06 %, 85.81 %, and 90.05 % respectively. Moreover, the XGboost classifier distinguished medium pain from high pain with ROC-AUC of 91%. For the multi-class classification among three pain levels, the XGBoost offered the best performance with an average F1-score of 80.03%. Our results suggest that BVP signal together with machine learning algorithms is a promising physiological measurement for automated pain assessment. This work will have a national impact on accurate pain assessment, effective pain management, reducing drug-seeking behavior among patients, and addressing national opioid crisis.

Motivation & Objective

• To develop a non-invasive, automated system for objective pain assessment using BVP signals.
• To identify a comprehensive set of physiological features from BVP and Inter-beat Intervals (IBIs) sensitive to pain.
• To evaluate multiple machine learning models for classifying pain presence and intensity levels.
• To determine the optimal model for pain recognition to support clinical pain management and reduce opioid misuse.

Proposed method

• Collected BVP signals from 32 healthy subjects during controlled pain induction.
• Extracted 24 BVP-specific features (time-domain, frequency-domain, and nonlinear dynamics) and 20 IBI-derived features.
• Trained and compared five machine learning models: Logistic Regression, Random Forest, SVM, AdaBoost, and XGBoost.
• Used 10-fold cross-validation to evaluate model performance on pain classification and intensity estimation.
• Optimized hyperparameters for each model using grid search and Bayesian optimization.
• Evaluated performance using ROC-AUC, F1-score, and confusion matrices for multi-class and binary classification tasks.

Experimental results

Research questions

• RQ1 Can BVP signals and derived IBI features effectively capture physiological changes associated with pain?
• RQ2 Which machine learning model performs best in classifying pain presence and intensity from BVP-derived features?
• RQ3 How accurately can the system distinguish between low, medium, and high pain levels compared to no pain?
• RQ4 Can the model differentiate between medium and high pain with high sensitivity and specificity?
• RQ5 What is the contribution of nonlinear dynamics and frequency-domain features to pain recognition performance?

Key findings

• XGBoost achieved the highest ROC-AUC of 90.05% for distinguishing high pain from no pain, outperforming all other models.
• The XGBoost model achieved an AUC of 85.81% for medium pain and 80.06% for low pain versus no pain.
• For binary classification between medium and high pain, XGBoost achieved an AUC of 91%, indicating strong discriminative capability.
• The XGBoost model achieved an average F1-score of 80.03% in multi-class classification across three pain levels (low, medium, high).
• The inclusion of nonlinear dynamics and IBI-derived features significantly enhanced model performance compared to BVP-only features.
• XGBoost demonstrated robustness and generalization across all pain intensity levels, making it the most suitable model for automated pain recognition.