Buller, Kevin

Pietrowicz et al. (2022) demonstrated feasibility of detecting potential speech-based biomarkers to identify gastroesophageal reflux disease (GERD) and Barrett’s esophagus (BE); however, speech samples were obtained using research-grade recording equipment (TASCAM DR-40X with AKG C555L microphone). The use of more readily available equipment such as the Apple iPad with a native microphone could simplify data collection and allow access to a broader range of participants. Therefore, this study compared the research-grade recordings to iPad recordings for feasibility of detecting GERD and BE. One-hundred-fifty adults, clinically classified as GERD negative or GERD positive with or without BE, recorded a paragraph (The Rainbow Passage) simultaneously on both a TASCAM and iPad. Both data sets were analyzed manually for Alpha Ratio and Cepstral Peak Prominence (CPP) with Praat. Initial results show significantly higher alpha ratio and lower CPP in iPad recordings, indicating greater noise energy in the high-frequency range. Additionally, the capability of machine and deep learning model types to discern across GERD negative, GERD positive, BE, and normal voices were compared. Detection across these conditions and comparison of acoustic findings were explored to determine usefulness of the less controlled option of recording using an iPad with its internal microphone.

This study explores the feasibility of detecting gastroesophageal reflux disease (GERD) and Barrett’s esophagus (BE) using acoustic features extracted from speech recordings. GERD is a common condition affecting up to 27.8% of adults in the US. Chronic GERD is associated with BE, the precursor to esophageal adenocarcinoma. An automatic screening tool for GERD would improve clinical outcomes by improving early detection. Biomarkers based on acoustic features of speech may hold promise as hoarseness often co-occurs with GERD. The study cohort consisted of 49 adults with and 63 without GERD based on esophagogastroduodenoscopy or ambulatory pH studies. Speech, including sustained vowels, sentences, and reading, was systematically recorded. Voice quality was validated through auditory-perceptualratings with three speech-language pathologists. Automated, cross-validated, search across multiple machine and deep learning model types was performed using the recorded speech. Detection of condition presence and severity are explored, and model performance is reported, along with an assessment of the ability of acoustic features to discern across conditions.