DETECTING STRESS AND DEPRESSION IN ADULTS WITH APHASIA THROUGH SPEECH ANALYSIS Stephanie Gillespie1, Elliot Moore1, Jacqueline Laures-Gore2, Matthew Farina2 , Scott Russell3, Yash-Yee Logan1 1 Georgia Institute of Technology, 2 Georgia State University, 3 Grady Memorial Hospital {sgillespie6, em80, ylogan3}@gatech.edu; jlaures@gsu.edu; mfarina2@student.gsu.edu; smrussell@gmh.edu Background Results Aphasia is an acquired communication disorder resulting from brain damage and impairs an individual’s ability to use, produce, and comprehend language. Loss of communication skills can be stressful and may result in depression, yet most stress and depression diagnostic tools are designed for adults without aphasia. This project is a research effort to predict stress and depression from acoustic profiles of adults with aphasia using linear support-vector regression. The labels were obtained through caregiver surveys (SADQ-10) or surveys not designed for adults with aphasia (PSS). SADQ-10 Depression Prediction Results PSS Stress Prediction Results Feature Type MAE (SADQ-σ) R2 P1SD (%) (PSS-σ) All 1.24 0.04 46.0% 1.51 0.12 36.8% Pitch + Jitter 1.08 0.34 48.4% 1.05 0.18 54.4% RMS-Energy 0.05 52.3% 0.94 61.4% LSF + Δ 1.11 0.07 53.0% 1.17 0.11 43.9% MFCC + Δ 1.25 47.7% 1.33 HNR 0.97 0.15 57.5% 1.02 55.8% CPP 1.03 0.13 54.7% TEO-All 1.15 0.03 1.01 0.00 55.3% TEO-AM 1.04 56.8% 0.29 61.1% TEO-FM 0.91 62.1% 1.06 0.25 53.2% TEO-CBarea 53.7% 55.0% TEO-RMS Energy 54.0% 0.40 56.1% TEO-log Energy 0.14 50.2% 0.98 0.09 59.4% Glottal-All 1.16 49.5% 0.02 52.9% H1-H2 PSP 0.44 51.9% 0.24 HRF 1.07 0.17 0.89 59.9% GLTP 1.21 44.9% 1.00 0.01 53.5% Data Collection Full Recording Material: Speech components of the Western Aphasia Battery-Revised (WAB-R) Protocol 2 additional Picture Descriptions Participants Included: 19 Participants selected for depression analysis based on Stroke Aphasia Depression Questionaire-10 (SADQ-10) 18 participants selected for stress analysis based on Perceived Stress Scale (PSS) Speech Analyzed: Used only phrase or sentence responses- no single words Spontaneous responses only- no repeat-after-me or fill in the blank responses. 18 sentences per person Pictures used to elicit spontaneous speech (from left to right): Picnic Scene included in the WAB, Cat in Tree photo, and Cookie Theft photo from the BDAE. Sample PSS questions (participant): Sample SADQ-10 questions (caregiver): -In the last month, how often have you been able to control irritations in your life? -Does he/she avoid eye contact when you talk to him/her? -In the last month, how often have you felt nervous and “stressed”? -Does he/she sit without doing anything? Participant characteristics that may influence speech acoustics Participant characteristics included in this study Age Gender Aphasia Type (WAB-R) Aphasia Quotient (WAB-R) Dysarthria (FDA-2) Apraxia (ABA-2) Mood # Males 12* # Females 7 Age Range 31-70 AQ 31.9-99.4 # with AQ>93.8 2 SADQ-10 score 6-25 PSS score 14-40 SADQ-10 and PSS regression results by feature subtype after feature selection. MAE= Mean Absolute Error, units are with respect to score-σ R2=R-Squared Score P1SD= Percentage of predictions within one score-σ from the actual value Depression (SADQ-10) Stress (PSS) Discussion Feature Extraction and Regression Correlation analysis between prediction median, mean, standard deviation, IQR, and accuracy and clinical/demographic information found no statistically-significant correlations Models do not appear to perform better or worse based on any specific clinical or demographic characteristics of the individuals Linear-SVR may not be able to handle complexities of the data, or possibly the database is too limited for generalizations of stress/depression to be drawn Future work will investigate: Emotional state/mood of the participants Impact of motor disorders on speech from adults with aphasia as it relates to affect Multi-session interview and recording data collection Pre-Processing Segment recordings into individual responses Voiced Speech Detection Results of prior work indicate poor classification performance on participants with scores just above/below threshold for binary depression label. Regression avoids the boundary errors by predicting the depression score instead of classifying. Feature Extraction Prosodic Spectral Teager Energy Operator Glottal Experiment Setup in Matlab Feature Selection Remove correlations >0.75, then 10-fold cross validation SFS Leave-one-participant-out train/test sets Linear Support-Vector regression model built and predictions observed Acknowledgements Supported by the Emory-Georgia Institute of Technology Healthcare Innovation Program and the National Center for Advancing Translational Sciences of the National Institutes of Health under Award Number UL1TR000454. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. Work supported by the National Science Foundation Graduate Research Fellowship, Grant No. DGE-1148903. SADQ-10 range: 0-30 PSS range: 0-56