Poster Number-41 - Development of Machine Learning Algorithms Using EEG Features to Detect Presence of Chronic Pain

Development of Machine Learning Algorithms Using EEG Features to Detect Presence of Chronic Pain

Purpose:

The current standard of care for assessing pain intensity is a subjective patient self-report using the Numeric Rating Scale (NRS) or the Visual Analog Scale (VAS). Subjective pain measurement may lead to a non-replicable, unreliable assessment of a patient’s pain state, leading to inappropriate treatment and ultimately poorly managed pain.

 

The purpose of PainQx’s research was to develop a technology to assess an individual’s pain state objectively and empirically. With support from a National Institutes on Drug Abuse SBIR grant, PainQx conducted Phase I of a multi-site research study and developed algorithms using quantitative electroencephalography (qEEG) derived neural activity to objectively classify chronic pain patients as “in pain” or “not in pain.” Through this research effort, PainQx developed a large database of chronic pain patients and controls, EEG processing modules for quality control, artifact removal, and epoch selection, and a machine learning (ML) pipeline for EEG based chronic pain classification algorithm development.

 

Development of a chronic pain biomarker that is linked to pain intensity has long been recognized as a valuable goal in the clinical community and has received increasing attention among qEEG researchers in the last two decades. PainQx sought to ensure our findings regarding neurological differences between pain patients and healthy controls were congruent with domain knowledge and the published literature on qEEG and pain.

 

Ultimately, the use of EEG to detect a chronic pain state will provide clinicians and patients an empirical, additional data point in their determination of appropriate treatment paths. Additionally, an EEG-based biomarker for chronic pain would have significant value for future research on pain mechanisms and therapeutic development. It is PainQx’s goal to further pain research and improve outcomes for the millions of patients currently suffering from chronic pain.

Methods:

In the first phase of the project, a multi-site clinical study was conducted to collect EEG and clinical data from a healthy control cohort and a pain patient cohort. Male and female participants ages 18-80 were included in the study, and pain subjects were required to meet the IASP definition of chronic pain. Subjects with neurological disorders or other conditions affecting neurological patterns were excluded from the study. Additionally, patients who may have had a reason to misrepresent their pain (e.g., patients on workers compensation) were also excluded.  Participant data collected during the study included fifteen minutes of eyes closed, resting EEG data and clinical information relating to the subject’s pain history, functional impairment, and mental state.

 

During and following the clinical study, EEG data was sent to the PainQx cloud computing platform, where it was processed by a set of core modules. The first module assessed the quality of the EEG files to insure they met several standards. If an EEG file met these quality criteria, it was processed by automatic artifactor software that identified and removed any distortions in the EEG data caused by signals not originating from the brain. The output was a collection of EEG segments called epochs. Epoch selection algorithms were used to assemble an optimal set of EEG segments that were then passed to the feature extractor module. Over nine thousand unique EEG features were extracted and calculated from each EEG recording, such as signal power within a frequency range on an electrode and signal coherence within a frequency range between pairs of electrodes. Complex features such as Granger Causality and Cross Frequency Coupling were also computed.

 

PainQx took the resulting qEEG features and participant reported NRS scores as inputs to develop an algorithm to predict chronic pain. The algorithm is a binary classifier generated utilizing Machine Learning (ML) best practices. Several methodologies, including Elastic Net (ENET) and Support-Vector Machine were used during development. Algorithms were developed solely on a train/test set (60% of the data), with performance evaluated on the remaining data to insure against overtraining of the algorithm.

Results:

There was a total of 386 participants in the study, 93 healthy controls and 293 chronic pain patients. The data from 235 subjects were used for algorithm development and is referred to as Train/Test (TT) data set. The remaining 151 participants is referred to as Hold Out (HO) and used for performance evaluation. Participants were assigned to either the TT or HO dataset randomly, but stratified for gender, age, and pain intensity.

 

Using the TT data, multiple Pain vs No Pain binary classifiers were created. After utilizing several machine learning methodologies in algorithm development, ENET was the most successful in correct classification of Pain vs. No Pain.

 

Performance was evaluated for both the TT and HO datasets. The TT performance was derived from 20 repetitions of 10-fold cross-validation, in which the TT set is partitioned at random into a Train interval and Test interval and results are aggregated, whereas the HO performance was a single application of the TT-developed classifier applied to the HO dataset.

 

The performance of the classifier was assessed using multiple metrics. Area Under the Curve (AUC) is the preferred metric as it is independent of a specific operating point, which can be set to optimize sensitivity and specificity. The following results utilize an operating point that provides an equal emphasis on both sensitivity and specificity.

 

The AUC was 0.941 for TT and 0.903 for HO. Sensitivity/specificity was 0.886/0.874 for TT and 0.841/0.878 for HO. Accuracy was 0.883 for TT and 0.841 for HO. Using a pain prevalence model developed by PainQx, the Negative Predictive Value/Positive Predictive Value was 0.291/0.993 for TT and 0.238/0.984 for HO.

 

Using ENET, each qEEG feature received a weight, with heavily weighted features indicating that feature’s relevance to the classification of Pain vs No Pain. There were several features that were heavily weighted in PainQx’s algorithm that are consistent with previously reported results. For example, an increase in theta-band activity in chronic pain patients has been reported in multiple publications [Mussigmann et al. 2022].  Additionally, a recent paper found five measures of EEG connectivity were the most predictive for identifying severe chronic pain vs pain-free controls [Ploner et al. 2019], which is also consistent with our results. The consistency of these features across multiple studies supports that the ML analysis has captured fundamental underlying physiology related to chronic pain, and not spurious anomalies present by chance in our training data.

Conclusion:

The research effort successfully demonstrated that machine learning based algorithms utilizing EEG data can be used to classify healthy controls in no pain versus chronic pain patients in pain. Several EEG features identified with the machine learning pipeline are congruent with domain knowledge identified by previous chronic pain and EEG studies. The ability to empirically differentiate patients in pain from those not in pain has several clinical use cases, including determination of proper opioid dispensing practices or validation of chronic pain state in legal cases.

 

Even with the success of this research, limitations in the performance achieved indicates that additional research is needed to refine the technology.  Areas of future research opportunities include conducting additional studies to increase the training database or combining EEG with other physiological data sources (such as heart rate or blood-based biomarkers) to further improve the technology and its pain intensity prediction capabilities.

References: Mussigmann, Thibaut, et al. “Resting-State Electroencephalography (EEG) Biomarkers of Chronic Neuropathic Pain. A Systematic Review.” NeuroImage, vol. 258, 2022, p. 119351., https://doi.org/10.1016/j.neuroimage.2022.119351.

Ta Dinh S, Nickel MM, Tiemann L, May ES, Heitmann H, Hohn VD, Edenharter G, Utpadel-Fischler D, Tölle TR, Sauseng P, Gross J, Ploner M. Brain dysfunction in chronic pain patients assessed by resting-state electroencephalography. Pain. 2019 Dec;160(12):2751-2765. doi: 10.1097/j.pain.0000000000001666. Erratum in: Pain. 2020 Jul 1;161(7):1684. PMID: 31356455; PMCID: PMC7195856.