Development & Validation of a Machine Learning Model That Uses Voice to Predict Aspiration Risk.

Background: Aspiration causes or aggravates a variety of respiratory diseases. Subjective bedside evaluations of aspiration are limited by poor inter- and intra-rater reliability, while gold standard diagnostic tests for aspiration, such as video fluoroscopic swallow study (VFSS) and fiberoptic endoscopic evaluation of swallowing (FEES), are cumbersome or invasive and healthcare resource intensive. Research Aim: To develop and validate a novel machine learning algorithm that can analyze voice features, as an objective screening test to predict aspiration risk. Methods: Recorded [i] phonations during routine nasal endoscopy from 163 unique patients were retrospectively analyzed for acoustic features including pitch, jitter, shimmer, harmonic to noise ratio (HNR), and others. Supervised machine learning (ML) was performed on the vowel phonations of those at high-risk for aspiration versus those at low-risk for aspiration. Ground truth of aspiration risk classification for model development was established using VFSS. The performance of the ML model was tested on an independent, external cohort of patient voice samples. The performance of trained Speech Language Pathologists (SLPs) to categorize high versus low-risk aspirators by listening to phonations was compared against the ML model. Results: Mean ML risk score for those with the ground truth of high versus low aspiration risk was 0.530± 0.310 vs 0.243±0.249, which was a significant difference (0.287, 95% CI: 0.192-0.381) p<0.001. In the development cohort, the model showed an area under the curve (AUC) for the Receiver Operator Characteristic (ROC) of 0.76 (0.67-0.84) with specificity of 0.76 and F1 score of 0.63. The performance of the model in an external testing cohort was comparable, with AUC of 0.70 (0.52-0.89) with a specificity of 0.81, and F1 score of 0.67. The ML model had better accuracy, sensitivity, specificity, negative and positive predictive values compared to trained SLPs in classifying aspiration risk by evaluating vowel phonations. Conclusion: High risk aspirators have quantifiable voice characteristics that significantly differ from those who are at a low risk for aspiration, as detected by a ML model trained to analyze sustained phonation and tested on an independent cohort. The ML model's performance was superior to human experts in predicting aspiration risk by evaluating vowel phonations.