Development of an objective pain assessment using functional near-infrared spectroscopy and machine learning

Abstract

Pain is a subjective experience, and no objective clinically available diagnosis test exists to assess it. In clinical practice, the most accepted and valid method is self-reports, this method relies on a patient’s ability to communicate a self-assessment of pain. However, the absence of verbal (or writing) communication in some patients (also referred as non-verbal) is an obstacle to the evaluation of pain, which may derived in risk to death and under- or over-treatment. Patients with impaired communication, unconscious patients, infants, stroke survivors, the critically ill, and persons suffering from advance dementia are examples of vulnerable individuals who are unable to communicate.
Therefore, the need for a reliable and objective pain assessment to assist medical
practitioners in the diagnosis and management of pain.
Attempts to use neuroimaging and computational methods to identify pain in healthy humans have shown potential to aid the assessment of pain. Previous studies using functional Magnetic Resonance Imaging (fMRI), Electroencephalography (EEG) and Near-Infrared Spectroscopy (NIRS), to predict and identify patterns of noxious signals (e.g., heat, cold) from non-painful stimuli have proved the use of brain signals to retrieve pain information. These results show that pain recognition and classification is plausible using neuroimaging methods, also the results from these studies advocate for the use of machine learning techniques to predict human pain.
The objective of this PhD research is to investigate the use of functional NIRS (fNIRS) and machine learning techniques to objectively estimate the pain status of non-verbal patients. This PhD r search also aims to expand previous studies by exploring the classification of fNIRS signals according to different types of thermal (cold and hot) pain and corresponding pain intensity (low and high) using machine learning models.
In addition, another purpose of this PhD research is the identification of a potential biomarker of pain based on fNIRS.
Based on the literature reviewed, the methodology created in this PhD thesis consisted of data collection, data analysis and interpretation. This included designing a noxious stimulation procedure using the thermal test following the Quantitative Sensory Testing (QST) method to obtain the pain threshold and pain tolerance in eighteen (n = 18) volunteers while recording their haemodynamic activity using an fNIRS system.
This database was labelled according to their corresponding type of stimuli:
low-cold, low-heat, high-cold, and high-heat. Then, from the fNIRS signals a total
of 69 features were established, 9 features in time, 23 features in frequency, and 37 features in wavelet domain; these features represent different distinctive characteristics in each domain. The features were evaluated by five feature selection methods and ranked based on relevance to the task and redundancy according to: Information Gain (IG), Joint Mutual Information (JMI), Student’s t-test (t-test), Chi-squared (X2), and Fast Correlation Based Filter (FCBF). The feature rankings were used to train and test three classifiers separately, the Linear Discriminant Analysis (LDA), K-Nearest Neighbour (K-NN) (with K = 1, 3, 5, 7, 9), and Support Vector Machines (SVM) (linear, Gaussian, and polynomial kernels).
The results of the this PhD research proved the adequacy of the designed methodology.
The stimulation paradigm showed a haemodynamic response in the primary
somatosensory cortex (S1) and opposite to the hand of application, which was
consistent with similar published literature. Most of the obtained features showed a normal distribution, while other presented a slightly positive skewed distribution; some features also showed high correlation among them, suggesting redundancy.
Feature selection methods ranked the obtained features, showing that timemax (time to maximum amplitude) and F7 (Fourier coefficient) were among the top ranked features.
The classification models tested the significance of each ranking to classify the data in four types of noxious stimuli, these results showed that the best classifiers were 1-NN, the Gaussian SVM and the polynomial SVM. The highest accuracy (94.16%) was exhibited by the Gaussian SVM using the top 25 features ranked by (JMI), however, using only 13 features this classifier showed an accuracy of 89.44%. Therefore, this subset of 13 features was identified as probable biomarker of pain using fNIRS.
The major contributions of this PhD research can be summarized in five items: a)
presenting the classification of four types of noxious stimuli, b) showing that pain
threshold and pain tolerance from the QST could be used to obtain pain information from non-verbal patients, c) increasing generalization by taking measurements from different domain representations, d) presenting a comparison between feature selection methods and learning models, to obtain the best representation of the pain database, e) finding a subset of features to be defined as potential biomarker of pain using fNIRS This PhD research demonstrates the application of fNIRS in the development of a
physiologically-based diagnosis of human pain that would benefit vulnerable patients who cannot self-report pain, and presents a set of features as possible biomarker of pain as measured by fNIRS. However, there is still further research need to develop a bedside monitor for the diagnosis of pain in clinical settings.

Date of Award	2018
Original language	English
Supervisor	Xu HUANG (Supervisor) & Dat Tran (Supervisor)