Medizinische Universität Graz Austria/Österreich - Forschungsportal - Medical University of Graz
Gewählte Publikation:
SHR Neuro Krebs Kardio Lipid Stoffw Microb
Kulvicius, T; Lang, S; Widmann, CAA; Hansmann, N; Holzinger, D; Poustka, L; Zhang, DJ; Marschik, P.
Facilitating Deep Acoustic Phenotyping: A Basic Coding Scheme of Infant Vocalisations Preluding Computational Analysis, Machine Learning and Clinical Reasoning
KINDH ENTWICKL. 2023; 32(3): 142-150.
Doi: 10.1026/0942-5403/a000418
Web of Science
FullText
FullText_MUG
- Führende Autor*innen der Med Uni Graz
- Marschik Peter
- Co-Autor*innen der Med Uni Graz
- Marschik Dajie
- Altmetrics:
- Dimensions Citations:
- Plum Analytics:
- Scite (citation analytics):
- Abstract:
- Theoretical Background: Very early infant development is still underresearched and emerging speech-language functions are underexplored. A reliable easy-to-apply annotation tool for very early infant vocalizations is lacking. Objective: To provide a tool - a ground-annotation scheme/coding scheme - to classify infant vocal data as a basis for the in-depth analysis of emerging verbal functions. Method: We analyzed longitudinal data from a prospectively recruited neurotypical infant cohort (N = 45). We filmed spontaneous movements and vocalizations fortnightly in seven consecutive lab sessions from the 4th-6th week postterm. We then semiautomatically segmented the audio sequences. We exported a total of 9,305 audio segments that included infant vocalizations and other sounds for human annotation. We defined a coding scheme with 5 classes: (1) voiced vocalization, (2) unvoiced vocalization, (3) fixed signal, (4) non-target (sound not vocalized by the infant), and (5) infant vocalization that cannot be assigned with certainty to classes (1)-(3). Two coders were involved, one with a rich experience in coding vocalizations, the other with no experience. The coders were instructed and familiarized with the scheme. No intensive training was needed for the coders to achieve high agreement on the pilot datasets. The coders proceeded to annotate the 9,305 segments. We then analyzed the consensually annotated vocalizations computationally to exemplify the use of the dataset generated with the proposed coding scheme. Results: For the 9,305 segments annotated by both coders, the 5-class Cohen's Kappa was .70 and .92 for classes (1)-(4). The test-retest reliability of the experienced coder was .82 and .68 of the novice coder. Of the 9,305 segments, 27.6 % were consensually classified as voiced vocalizations, 0.1 % as unvoiced vocalizations, and 1.5 % as fixed signals. The computational analysis replicated findings that the fundamental frequency and duration of vocalizations are comparable between baby boys and girls. Discussion and conclusion: The proposed coding scheme proved applicable for both the experienced and the novice coder. In contrast to other schemes, one can apply the current scheme without intensive training to achieve satisfactory annotation results. Experience may improve the rating, as suggested by the higher reliability and lower rating uncertainty of the experienced coder. This easy-to-apply scheme can be broadly utilized in research and clinical settings for the efficient initial annotation of infant vocalizations. The generated datasets suit further in-depth manual and computational analyses and pave the way to examine emerging verbal functions - deep acoustic phenotyping - of infants of different developmental and clinical phenotypes.
- Find related publications in this database (Keywords)
- acoustics
- infant
- infancy
- speech-language
- development
- vocalization
- coding-scheme