Classifying social and physical pain from multimodal physiological signals using machine learning

Abstract

Accurate pain assessment is essential for effective management; however, most studies have focused on differentiating pain from non-pain or estimating pain intensity rather than distinguishing between distinct pain types. We present a machine learning method for classifying physical and social pain using physiological signals. Seventy-three healthy adults participated in experiments involving baseline, neutral, and pain-inducing stimuli related to both types of pain. Physical pain was elicited by pressure cuff inflation, whereas social pain was induced by watching a video depicting a loved one's death. The electrocardiogram, electrodermal activity, photoplethysmogram, respiration, and finger temperature were recorded, and 12 physiological features were extracted. Three machine learning algorithms-logistic regression, support vector machine, and random forest-were employed to classify the input data into baseline versus painful states and physical versus social pain. Our findings demonstrated high accuracy in identifying social pain (0.82) and physical pain (0.90) compared to the baseline. Classification accuracy between physical and social pain was moderate (0.63) when using painful state data alone but improved to 0.77 when incorporating reactivity from neutral to painful states. This study highlights the potential of multimodal physiological signals for differentiating pain types and enhancing personalized pain management strategies.

Keywords: Autonomic nervous system; Machine learning; Multimodal; Physical pain; Physiological signals; Social pain.

Fig. 1

Experimental methods (a) overall procedure (b) emotion ratings (c) physiological signals observed during the experiment. Pain-inducing sessions were presented in random order.

Fig. 2

Bar graphs showing performance measures for classifying baseline and painful states of the social and physical pain stimuli. Brackets indicate significant differences in paired comparisons.

Fig. 3

Bar graphs showing performance measures for classifying physical and social pain types based on painful state and reactivity data. Brackets indicate significant differences in paired comparisons.

Fig. 4

Learning curves for pain-type classification using reactivity data: (a) LoR, (b) SVM, and (c) RF. Each subplot shows training and test accuracy across varying fractions of the dataset, where a fraction of 1 represents the entire dataset. Solid lines indicate mean values, and dashed lines represent standard deviations.

Fig. 5

SHAP values for classifying physical and social pain types using (a) painful state data and (b) pain reactivity data, with no feature engineering method applied.