A Machine Learning Approach for the Identification of a Biomarker of Human Pain using fNIRS

Abstract

Pain is a highly unpleasant sensory and emotional experience, and no objective diagnosis test exists to assess it. In clinical practice there are two main methods for the estimation of pain, a patient's self-report and clinical judgement. However, these methods are highly subjective and the need of biomarkers to measure pain is important to improve pain management, reduce risk factors, and contribute to a more objective, valid, and reliable diagnosis. Therefore, in this study we propose the use of functional near-infrared spectroscopy (fNIRS) and machine learning for the identification of a possible biomarker of pain. We collected pain information from 18 volunteers using the thermal test of the quantitative sensory testing (QST) protocol, according to temperature level (cold and hot) and pain intensity (low and high). Feature extraction was completed in three different domains (time, frequency, and wavelet), and a total of 69 features were obtained. Feature selection was carried out according to three criteria, information gain (IG), joint mutual information (JMI), and Chi-squared (χ²). The significance of each feature ranking was evaluated using three learning models separately, linear discriminant analysis (LDA), the K-nearest neighbour (K-NN) and support vector machines (SVM) using the linear and Gaussian and polynomial kernels. The results showed that the Gaussian SVM presented the highest accuracy (94.17%) using only 25 features to identify the four types of pain in our database. In addition, we propose the use of the top 13 features according to the JMI criteria, which exhibited an accuracy of 89.44%, as promising biomarker of pain. This study contributes to the idea of developing an objective assessment of pain and proposes a potential biomarker of human pain using fNIRS.

Figure 1

Thermal threshold and tolerance levels perceived by the participants after cold (left panel) and heat (right panel) stimuli. Horizontal red lines are the median values across all participants for each test. Pain threshold (tests 1–3) and pain tolerance (tests 4–6).

Figure 2

Classification results by seven different learning models using the ranked features according to the information gain (IG) criterion.

Figure 3

Classification results by seven different learning models using the ranked features according to the joint mutual information (JMI) criterion.

Figure 4

Classification results by seven different learning models using the ranked features according to the Chi-squared (Chi-2) criterion.

Figure 5

Time-frequency analysis (bottom-right panel) of a raw HbO signal using the wavelet transform. Heartbeat signal can be seen in the frequency of ~1.25 Hz, it is exhibited as a large peak in the frequency spectrum (bottom-left panel) and affects the data during the whole experiment as observed in the wavelet domain. The effect of a moving artefact is also observed after the last stimulus (after time 200 sec) in the temporal graph (top panel), which affects several frequency bands (only observed in the wavelet domain).

Figure 6

Stimulation paradigm. In this example, pain threshold test was first measured followed by pain tolerance test. In each test, cold and hot stimulus were applied on the back of the hand of each subject. Each stimulus was applied in a random order.

Figure 7

Channel location and configuration. Channel probes were located around the C3 and C4 areas. Source-detector distance was 3 cm.