OR13-6 Deep learning-based voice screening technique for cystic fibrosis related diabetes

Abstract Background Cystic fibrosis-related diabetes (CFRD) is a unique type of diabetes that is associated with significantly increased morbidity and mortality in both children and adults with cystic fibrosis (CF). The prevalence of CFRD progressively increases with age such that more than half of adults with CF develop CFRD. Early diagnosis and treatment are associated with improvements in body weight and pulmonary function, reduction in the frequency of pulmonary exacerbations, and improved overall survival. Since patients with CFRD may present with no symptoms, screening is recommended starting from the age of 10 years with an annual oral glucose tolerance test (OGTT). Especially during the COVID-19 pandemic, requiring an in-person clinic visit can be challenging, which may lead to a delayed diagnosis of CFRD. Objectives The purpose of this project was to develop a state-of-the-art technique to detect changes in glucose levels of patients with CF by developing a deep learning-based audio classification tool. Preliminary work by our group suggested that voice characteristics could distinguish between patients with CFRD patients and patients with CF but without CFRD. We hypothesize that high blood glucose levels may cause laryngeal soft tissue swelling leading to changes in voice characteristics. Methods We performed a prospective cross-sectional study in adult patients with CF recruited from Emory CF Clinic from March to December 2021. We recorded 5-second voice samples of a sustained /a/ vowel via a portable digital microphone. The spectrogram was extracted via the Mel frequency cepstral coefficient. The training to test the dataset ratio was 80: 20. 20% of the training dataset were randomly selected to serve as a validation dataset. We designed a convolutional neural network (CNN) architecture for CFRD patients’ voice classification. Results There were a total of 100 subjects consisting of 43 patients with CFRD and 57 patients with CF without diabetes. The male to female ratio was approximately 60: 40 in both groups. Patients with CFRD had similar mean age and mean BMI to patients without CFRD. There was a significantly higher point of care glucose level in CFRD patients. The mean duration of a CFRD diagnosis was 9 years and the mean HbA1c level was 7.26 in the CFRD group. The performance of the VGG model CNN classifier achieved 98.7% and 94.92% accuracy on training and validation datasets, respectively. On the test dataset, the model achieved 73.53% sensitivity 69.77% specificity and 71.43% accuracy. Conclusions We found a deep learning-based audio screening tool for CFRD could be potentially used as an alternative tool for screening in the CF community. A convolutional neural network algorithm demonstrated high sensitivity and specificity to adequately differentiate between patients with and without CFRD. Larger prospective studies are required to test this technology in patients with every form of diabetes. Presentation: Sunday, June 12, 2022 12:15 p.m. - 12:30 p.m.