34. Development of a Machine Learning Algorithm For Hypernasality Diagnosis Using Caps-a-am Rated Speech Samples In Pediatric Vpd Patients

PURPOSE: Velopharyngeal dysfunction (VPD), commonly associated with cleft palate, results in hypernasal speech that affects intelligibility and can lead to social stigmatization. Accurate diagnosis of VPD requires the expertise of trained speech-language pathologists (SLPs) using the CAPS-A-AM, the gold standard for assessing hypernasality. However, access to these specialists is limited in many centers. Recent efforts to incorporate machine learning (ML) into VPD diagnosis have relied on samples limited to isolated phonemes, internet speech samples, and crowd-sourced data. These models have not been trained on prospectively accrued and reproducibly rated speech samples using CAPS-A-AM. Our project aims to develop a clinically viable ML algorithm for VPD diagnosis based on CAPS-A-AM-rated, standardized speech samples, advancing beyond the approaches previously attempted. METHODS: Speech samples were collected from English-fluent pediatric patients (ages 2-17) with and without VPD. Each participant provided a standardized speech sample, recorded by an SLP following CAPS-A-AM protocols. Demographics, clinical/surgical history, and nasalance scores were also collected. These speech samples were independently rated by both unblinded and blinded SLPs on the CAPS-A-AM scale to ensure inter- and intra-rater reliability. The speech samples were pre-processed to remove the SLP’s voice and converted into spectrograms using Praat software. A large language model (LLM) was then used to identify acoustic differences in the spectrograms between hypernasal and control samples. Currently, 28 recordings have been completed, with additional samples being collected. Next, these identified acoustic features will be used to train and validate the ML model to detect hypernasality and assign severity ratings based on the CAPS-A-AM scale. RESULTS: Currently, 28 speech samples have been collected and processed. (VPD: n=18, control: n=10). The average age of participants was 8.3 years (range: 5.1-12.3 years). For VPD cases, the average CAPS-A-AM hypernasality score was 2.92 (range: 1-4) and overall speech acceptability was 2.58 (range: 2-3). Preliminary results show that the LLM was able to correctly identify hypernasal cases on a binary scale in 100% of the initial tests. Further training and validation of the algorithm to assign ratings on the CAPS-A-AM scale is ongoing, with additional data being processed to improve model accuracy and performance. CONCLUSION: Our preliminary work demonstrates that a standardized, CAPS-A-AM-based ML model holds significant potential for improving the accuracy of VPD diagnosis, especially in areas with limited access to specialized SLPs. With continued refinement and expansion of the dataset, this tool could aid in clinical decision-making, providing an objective, scalable approach to the diagnosis and severity grading of hypernasality in patients with VPD.