Immediate and partial neural dysfunction after thyroid and parathyroid surgery: Need for recognition, laryngeal exam, and early treatment

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

Introduction

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.

Methods

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.

Results

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.

Conclusion

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.

Intraoperative Neuromonitoring: Evaluating the Role of Continuous IONM and IONM Techniques for Emerging Surgical and Percutaneous Procedures

Continuous and intermittent intraoperative nerve monitoring (IONM) has become an important asset for endocrine surgeons over the past few decades. The ability to determine neurophysiologic integrity of the recurrent laryngeal nerve (RLN) and external branch of the superior laryngeal nerve (EBSLN) on top of identification and meticulous dissection of the nerve in the surgical field, has allowed for advances in technology and improved outcomes when it comes to prevention of vocal fold immobility. This article aims to compare in review continuous and intermittent nerve monitoring (CIONM, IIONM), as well as review the current paradigms of their use. This article will also discuss the future of intraoperative nerve monitoring technologies in scarless thyroid surgery and percutaneous approaches to thyroid pathology in form of radiofrequency ablation (RFA).

Comparison of a Handheld Device vs Endotracheal Tube-Based Neuromonitoring for Recurrent Laryngeal Nerve Stimulation

OBJECTIVE

To measure the effect of thyroidectomy difficulty on intraoperative neuromonitoring false loss of signal (LOS) and to compare intraoperative endotracheal tube-based neuromonitoring (ETNM) and Checkpoint palpation-based direct stimulation (pDS) signals with postoperative laryngoscopy. We hypothesized that pDS has higher a positive predictive value for postdissection confirmation of recurrent laryngeal nerve function than ETNM and that this difference is accentuated with increasing thyroidectomy difficulty.

STUDY DESIGN

Prospective single-arm cross-sectional study comparing ETNM and pDS for patients undergoing hemi-, total, or completion thyroidectomy from July 2018 to March 2020.

SETTING

Single-surgeon series at a tertiary care hospital.

METHODS

Percentage concordance and positive and negative predictive values were measured. Each thyroidectomy was assigned a validated thyroidectomy difficulty score, and recorded recurrent laryngeal nerve signals were compared with postoperative vocal fold mobility.

RESULTS

Percentage concordance was 90.09%. Positive and negative predictive values were 0.19 (95% CI, 0.09-0.31) and 1.0 for ETNM and 0.59 (95% CI, 0.35-0.82) and 1.0 for pDS. The difference in positive predictive value was significant (0.40 [95% CI, 0.33-0.47], P < .001). False LOS rates for ETNM and pDS were 13.19% versus 3.30% (9.89% [95% CI, 1.80%-18.62%], P = .0155), 44.11% versus 0% (44.11% [95% CI, 25.80%-60.54%], P < .001), and 73.33% versus 13.33% (60% [95% CI, 24.76%-78.46%], P = .001) for the second through fourth thyroidectomy difficulty score quartiles, respectively. False LOS with ETNM was linearly correlated with increasing difficulty (R² = 0.97).

CONCLUSION

ETNM was subject to high rates of postdissection false LOS that increased with thyroidectomy difficulty score. pDS is a reliable alternative that has higher positive predictive value than ETNM, particularly in more challenging cases such as those with posteriorly fixed thyroid cancers and fibrotic glands.

EVIDENCE LEVEL

2.