AUTHOR=Drai-Zerbib Véronique , Ansart Manon , Grenot Clément , Poulin-Charronnat Bénédicte , Perra Joris , Baccino Thierry TITLE=Classifying musical reading expertise by eye-movement analysis using machine learning JOURNAL=Frontiers in Cognition VOLUME=3 YEAR=2024 URL=https://www.frontiersin.org/journals/cognition/articles/10.3389/fcogn.2024.1417011 DOI=10.3389/fcogn.2024.1417011 ISSN=2813-4532 ABSTRACT=

Music reading is the key to literacy for musicians in the Western music tradition. This high-level activity requires an efficient extraction of the visual information from the score to the current needs of the execution. Differences in eye movements between expert and non-expert musicians during music reading have been shown. The present study goes further, using a machine learning approach to classify musicians according to their level of expertise in analyzing their eye movements and performance during sight-reading. We used a support vector machine (SVM) technique to (a) investigate whether the underlying expertise in musical reading could be reliably inferred from eye movements, performance, and subjective measures collected across five levels of expertise and (b) determine the best predictors for classifying expertise from 24 visual measures (e.g., the number of progressive fixations, the number of regressive fixations, pupil size, first-pass fixations, and second-pass fixations), 10 performance measures (e.g., eye–hand span, velocity, latency, play duration, tempo, and false notes), and 4 subjective measures (perceived complexity and cognitive skills). Eye movements from 68 pianists at five different levels of music expertise (according to their level in the conservatory of music—from first cycle to professional) were co-registered with their piano performance via a Musical Instrument Digital Interface, while they sight-read classical and contemporary music scores. Results revealed relevant classifications based on the SVM analysis. The model optimally classified the lower levels of expertise (1 and 2) compared to the higher levels (3, 4, and 5) and the medium level (3) compared to higher levels (4 and 5). Furthermore, across a total of 38 measures, the model identified the four best predictors of the level of expertise: the sum of fixations by note, the number of blinks, the number of fixations, and the average fixation duration. Thus, efficiently classifying musical reading expertise from musicians' eye movements and performance using SVM is possible. The results have important theoretical and practical implications for music cognition and pedagogy, enhancing the specialized eye and performance behaviors required for an expert music reading.