Exploring Shapely Values for Blood Glucose Level Prediction from Speech
Konferenz: Speech Communication - 15th ITG Conference
20.09.2023-22.09.2023 in Aachen
doi:10.30420/456164015
Tagungsband: ITG-Fb. 312: Speech Communication
Seiten: 5Sprache: EnglischTyp: PDF
Autoren:
Pompe, Simone (EIHW – Chair of Embedded Intelligence for Health Care & Wellbeing, University of Augsburg, Germany)
Mallol-Ragolta, Adria (EIHW – Chair of Embedded Intelligence for Health Care & Wellbeing, University of Augsburg, Germany & Centre for Interdisciplinary Health Research, University of Augsburg, Germany)
Schauer, Nicolas (Vindex GmbH, Germany)
Schuller, Bjoern W. (EIHW – Chair of Embedded Intelligence for Health Care & Wellbeing, University of Augsburg, Germany & Centre for Interdisciplinary Health Research, University of Augsburg, Germany & GLAM – Group on Language, Audio, & Music, Imperial College London, UK)
Inhalt:
We explore a novel dataset for blood glucose level prediction from self-recorded speech. The dataset contains 10 h 30m 25 s of data from 63 German patients (44 f, 19 m). We model the paralinguistic information embedded in the voice, exploiting the Low-Level Descriptors (LLD) of the eGeMAPS feature set. We investigate the use of Shapely values to understand the contribution of each individual LLD on the inferences produced by a Support Vector Machine (SVM). We also compare the performance of subsets of the LLDs selected by the Shapely values, or transformed using Principal Component Analysis (PCA). We tackle the task as a 3-class classification problem with the Unweighted Average Recall (UAR) as the evaluation metric. The baseline SVM model scores a UAR of 51.8% on the test partition. The best model selecting a subset of the LLDs based on the Shapely values obtains a UAR of 56.8 %, while the top model transforming the LLDs with PCA reaches a UAR of 42.0 %, both on the test partition.