The Voice and the Larynx in Older Adults: What's Normal, and Who Decides?

Identifying bias in models that detect vocal fold paralysis from audio recordings using explainable machine learning and clinician ratings

Detecting voice disorders from voice recordings could allow for frequent, remote, and low-cost screening before costly clinical visits and a more invasive laryngoscopy examination. Our goals were to detect unilateral vocal fold paralysis (UVFP) from voice recordings using machine learning, to identify which acoustic variables were important for prediction to increase trust, and to determine model performance relative to clinician performance. Patients with confirmed UVFP through endoscopic examination (N = 77) and controls with normal voices matched for age and sex (N = 77) were included. Voice samples were elicited by reading the Rainbow Passage and sustaining phonation of the vowel "a". Four machine learning models of differing complexity were used. SHapley Additive exPlanations (SHAP) was used to identify important features. The highest median bootstrapped ROC AUC score was 0.87 and beat clinician's performance (range: 0.74-0.81) based on the recordings. Recording durations were different between UVFP recordings and controls due to how that data was originally processed when storing, which we can show can classify both groups. And counterintuitively, many UVFP recordings had higher intensity than controls, when UVFP patients tend to have weaker voices, revealing a dataset-specific bias which we mitigate in an additional analysis. We demonstrate that recording biases in audio duration and intensity created dataset-specific differences between patients and controls, which models used to improve classification. Furthermore, clinician's ratings provide further evidence that patients were over-projecting their voices and being recorded at a higher amplitude signal than controls. Interestingly, after matching audio duration and removing variables associated with intensity in order to mitigate the biases, the models were able to achieve a similar high performance. We provide a set of recommendations to avoid bias when building and evaluating machine learning models for screening in laryngology.

Stratifying Presbylarynx: Characterization of its Three Types

OBJECTIVES

To describe the clinical characteristics of each type of glottal configuration proposed in a classification that allows stratifying presbylarynx into three types: Type 0: there is no significant morphological alteration in glottal configuration. Type 1: two or more of the following endoscopic signs are present: atrophic vocal folds, vocal fold bowing, prominence of vocal processes, supraglottal hyperactivity, decreased amplitude (vocal fold edge stiffness without evidence of vocal post-traumatic scar), and tremor of laryngeal structures. Type 2: elderly patients with type 1 plus glottal insufficiency.

METHODS

Prospective, observational, and cross-sectional study. Data regarding age, gender, type of glottal configuration, self-assessment regarding vocal complaints, as well as acoustic and audio-perceptive examination were collected.

RESULTS

The study population included 306 subjects (180 females), and the mean age was 73.85. Type 0 was the most frequent (61%). A statistically significant association was found between type 0 and female subjects, as well as type 2 and male subjects. The mean age of the patients with type 0 was statistically lower than the mean age of the patients with type 1 or type 2. The mean score of self-assessment questionnaires of patients with type 2 was statistically higher than the mean score registered for patients with type 0 and type 1. The mean score of Jitter, Shimmer, and HNR in type 2 were statistically higher than in type 0 or type 1 (P < 0.05). The mean score of Maximum Phonation Time in type 2 was statistically lower than in type 0 or 1 (P < 0.001). Based on vocal complaints, 115 subjects (38%) had presbyphonia (46 males; 69 females), and the mean age was 75.97 years (SD = 6.98; range 65 - 97). Among subjects without presbyphonia (n = 191), type 0 was the most frequent (84%), followed by type 1 (15%) and type 2 (1%). Among subjects with presbyphonia (n = 115), type 1 was the most frequent (58%), followed by type 0 (23%) and type 2 (19%).

CONCLUSIONS

This study presented the clinical characterization of a classification that allows stratifying endoscopic laryngeal signs in elderly patients with presbyphonia. We believe this classification will enrich the diagnostic protocol of the aging voice and will improve the documentation of treatment outcomes.

Laryngoscopic Findings of Age-related Vocal Fold Atrophy are Reliable but not Specific

OBJECTIVES

We aimed to evaluate the reliability of laryngoscopic features of vocal fold atrophy as assessed by novice otolaryngology trainees and expert laryngologists.

DESIGN

Two expert fellowship-trained laryngologists and three non-expert otolaryngology resident trainees were recruited to view 50 anonymized laryngo-stroboscopic examinations of patients presenting with dysphonia and non-voice, laryngeal complaints. Reviewers were asked to stratify the patient's age, provide an opinion about the presence of age-related vocal fold atrophy, and specify which laryngoscopy features were present to make the diagnosis.

SETTING

Tertiary care laryngology practice.

PARTICIPANTS

Two fellowship-trained laryngologists and three trainee otolaryngologists.

MAIN OUTCOME MEASURES

Accuracy of age categorization was determined and Kappa analysis was performed to assess inter-rater agreement.

RESULTS

The mean age of patients was 54.9 years old with near equal male to female distribution. The overall accuracy of age category determination by raters was only 30.8%. Kappa analysis demonstrated fair agreement regarding the presence of vocal fold atrophy in non-expert reviewers, and moderate agreement amongst expert reviewers. Features of glottic gap, muscular atrophy of vocal folds, and prominent vocal processes were all identified with high agreement (>80.0%).

CONCLUSION

Our study illustrates that while raters can agree on the presence of age-related vocal fold atrophy, the findings may be non-specific and do not necessarily correlate with age.

Sectorial Analysis of the Fibrous Matrix of Vocal Folds in the Elderly

OBJECTIVE

to determine the sectorial concentration of fibrous matrix along the vocal folds of the elderly.

METHODS

Thirteen fresh vocal folds were removed from 13 cadavers. The subjects were divided into two groups: control group (CG) - age range 25-40 years (n = 5), and elderly group (EG), aged ≥70 years (n-8). The vocal folds were separated according to regions in: anterior macula flava (AMF), posterior macula flava (PMF), medial region (M) and lateral region of the membranous vocal fold (L). The material was examined using a scanning electron microscope and 10 pictures per region were obtained at 1200x magnification. The area of fibrous matrix was quantified in both groups using the AVSOFT Biovew Program.

RESULTS

Percentage (%) of fibrous matrix per region (%): AMF (EG = 86.78%; CG = 81.63%); PMF (EG = 88.19%; CG = 81.52%); M (EG = 84.92%; CG = 77.96%); L (EG = 83.93%; CG = 78.91%). Irregular distribution of a dense fibrous matrix was also more evident in the elderly's larynges in a qualitative assessments.

CONCLUSION

A higher concentration of fibrous matrix was observed in the vocal folds of the elderly when compared to the control at all the studied regions. The greatest increase in the concentration of fibers in the medial portion of the body of the vocal folds is probably due to the phonatory stimulation. The macula flava remains functional even in the senile larynx.