Comparing pain related EEG patterns in musculoskeletal pain, healthy participants, and experimental induced pain, using a machine learning algorithm
Author
Laursen, Anders Kruse
Term
4. term
Publication year
2025
Submitted on
2025-05-30
Pages
59
Abstract
Muskuloskeletal (MSK) smerte er udbredt og vurderes ofte med subjektive metoder, som i begrænset omfang afspejler smerters psykologiske dimensioner. Dette projekt havde til formål at udvikle en maskinlæringsmodel, der på baggrund af EEG-data kan klassificere målinger som kronisk smerte, akut (eksperimentelt fremkaldt) smerte eller rask, og samtidig visualisere hvilke tidspunkter i signalet der er mest betydningsfulde for klassifikationen. EEG-data blev opdelt i 5-sekunders segmenter, justeret til bevægelsesstart, og i alt 1.383 segmenter blev anvendt til at træne og teste en Long Short-Term Memory (LSTM) Convolutional Neural Network (CNN). Derudover blev der genereret heatmaps for at illustrere, hvor modellen fokuserede under analysen. Modellen opnåede en samlet prædiktionsnøjagtighed på 54,55 % og en F1-score på 0,543, og heatmaps viste opmærksomhedsvægte for både enkeltsegmenter, de enkelte klasser og modellen som helhed. Resultaterne viser et proof-of-concept med moderat ydeevne og forklarlige modeludfald, men indikerer også behov for forbedringer – særligt for klassificering af raske deltagere – for at øge den samlede nøjagtighed og den praktiske anvendelighed som supplement til traditionelle smertevurderinger.
Musculoskeletal (MSK) pain is widespread and typically assessed by subjective methods that only partly capture the psychological dimensions of pain. This project aimed to develop a machine learning model that uses EEG data to classify recordings as chronic pain, acute (experimentally induced) pain, or healthy, while also visualizing when in the signal the discriminative features occur. EEG recordings were segmented into 5‑second windows aligned to movement onset, and a total of 1,383 segments were used to train and test a Long Short‑Term Memory (LSTM) Convolutional Neural Network (CNN). Heatmaps were generated to illustrate where the model focused during analysis. The model achieved an overall prediction accuracy of 54.55% and an F1 score of 0.543, with heatmaps depicting attention weights for individual segments, each class, and the model overall. These findings demonstrate a proof of concept with moderate performance and interpretable outputs, while highlighting the need for further improvements—especially for the healthy class—to enhance overall accuracy and potential utility as a complement to traditional pain assessment.
[This abstract was generated with the help of AI]
Documents