A primary care feasibility study using artificial intelligence in diagnosing paediatric middle ear problems

The study intervention is video-otoscopy image capture of the tympanic membrane and AI-driven computer vision model output generated using the drumbeat.ai algorithm (MRdrumbeat.ai). This will be administered by a primary care nurse or healthcare assistant within the normal clinical pathway for children aged 1- 10 years old attending a GP consultation at the study setting, Otara Local Doctors, over a 4-8 week pilot period. AI algorithm: DrumBeat.ai, has been developed at the University of Sydney, with early validation in both Australian Indigenous communities and New Zealand. It uses deep learning (Densenet-169 convolutional neural network) to classify otoscopic images across eight diagnostic stages, with reported stepwise accuracy of 86%. Study design: This study is a mixed-methods prospective feasibility study. The study includes both inductive and deductive qualitative analysis of provider and whanau perspectives, and prospective cross-sectional measurement of provider-reported useability, acceptability and participant clinical outcomes. The study setting is Otara Local Doctors, a large General Practice facility in South Auckland with an enrolled patient population of 21,000. The study will proceed in three phases. i. The first phase will collect baseline data by exploring the primary-care providers’ perspectives regarding the acceptability of MRdrumbeat.ai, implementation and adaptability factors. ii. The second phase will implement the study intervention: MRdrumbeat.ai tool, into the real-world general practice. Participants will be recruited opportunistically as children aged 1-10 years presenting within normal business hours for GP consultation. Whanau acceptability will be measured by anonymous survey responses after each appointment (up until 40 surveys have been returned). Feasibility and usage data will be captured in real time. Each eligible child participant will be offered MRdrumbeat.ai during the nurse triage consultation prior to GP consultation. The nurse will acquire video-otoscopic images of each ear sequentially (2 minutes). They will select the clearest right and left ear image and input directly into drumbeat.ai portal. Drumbeat.ai output will be generated immediately (within 2 minutes) and provided to the family and GP. The participant will subsequently undergo normal GP consultation. Image capture success, analysable image rate, and errors/malfunctions will be recorded by the triage nurse. iii. The third phase will explore whanau, administration/manager personnel, and primary-care provider perspectives using semi-structured interviews. Participants will be recruited through purposive sampling up to 15 participants. Clinical outcomes and demographic data will be collected for child participants from their medical records (including age, ethnicity, presenting symptom, drumbeat.ai prediction, GP diagnosis and management plan). Useability of the MRdrumbeat.ai tool will be assessed by participants using the System Useability Scale.