Extraction and Analysis of Speech Features to Understand Audio Phenotypes of Pain

INTRODUCTION: Pain evaluation of spine patients remains largely a subjective assessment in practice, but artificial intelligence provides the potential for objective pain assessment tools. METHODS: This study enrolled spine patients from a general neurosurgical clinic. Patients were asked to download the Beiwe smartphone application and prompted to complete pain surveys at home once per day and speech recordings once per week during the study. The pain surveys asked patients to rate their pain on a scale from 0 to 10. For the speech recordings, patients were asked to read a passage from a well-known novel. Patients were included in the study if they completed at least one speech recording and pain survey within 24 hours of each other. Audio features were extracted from the speech recordings using the Parselmouth library, and 13 Mel-frequency cepstral coefficients (MFCCs) were extracted using the Librosa package. The mean MFCC values were calculated for each speech recording. These speech features were then analyzed using a linear mixed model with pain as the model output and a random intercept for each patient. RESULTS: In total, there were 384 speech recordings from 60 patients. The linear mixed model was significantly associated with two MFCCs – MFCC 2 (Coef: 0.552; 95% CI: [0.254, 0.850]; p < 0.001) and MFCC 12 (Coef: -0.286; 95% CI: [-0.538, -0.034]; p = 0.026). However, we observed that none of the other available audio features were significantly associated with pain scores. CONCLUSIONS: This preliminary study shows that MFCCs 2 and 12 are of potential interest for phenotyping speech and objective pain assessment.