Differential diagnosis of RRP and VFN in the paediatric voice clinic using audio data: outcomes of a feasibility study exploring the use of Deep Learning as an adjunct in diagnosis and monitoring

Introduction and Aims -- Current incidence of paediatric Recurrent Respiratory Papillomatosis (RRP) in the UK is 1.42 per 100,000. RRP presents as hoarseness with or without stridor. There is a broad spectrum of voice pathologies in children that can present as hoarseness, including vocal fold nodules (VFN). Differential diagnosis of RRP requires visualisation of the larynx although the voice quality of RRP is distinct from VFN. Our aim was to explore the potential for developing a software diagnostic tool to improve the efficiency of diagnosing RRP and to measure if experienced otorhinolaryngologists can differentiate RRP from VFN from an audio recording alone. Methods -- Sixty-seven audio recordings, consisting of sustained vowels and standard sentences from the CAPE-V protocol, were analysed. These recordings were subject to various classification methods extracting voice pathology features alongside deep learning methods such as Convolutional Neural Networks (CNN). We presented a subset of full audio recordings to four experienced paediatric otorhinolaryngologists who rated them as ‘RRP’ or ‘not RRP’. Results -- We tested this software on a subset of 24 RRP and 24 VFN recordings, identifying presence of RRP 75% of the time. A moderate degree of reliability was found between the listener ratings. The average measure ICC was .670 with a 95% confidence interval from .406 to .834 (F(26,78)=3.029, p<0.001). Conclusion -- Clinical diagnosis of audio recording without visualisation of the larynx is insufficient. However there may be utility in analysing voice recordings using this software as a diagnostic or monitoring adjunct for RRP. A much larger dataset is required and further study is planned.