Computer-assisted diagnosis of ear pathologies by combining digital otoscopy with complementary data using machine learning

摘要 耳部疾病，特别是急性中耳炎（AOM）和中耳积液，是最常见的 治疗儿童疾病据估计，耳部疾病的经济负担超过32亿美元， 年由于耳部疾病很常见，一个重要的问题是过度诊断和过度治疗， 因素首先，诊断耳部疾病的主观性-基于对鼓膜的简短一瞥， 耳镜-即使对于有经验的初级保健，急诊医学或 耳鼻喉科（ENT）医师。其次，随着美国初级保健医生的日益短缺， 更多的高级实践提供者（执业护士和医师助理）作为一线临床医生， 初级保健和紧急情况但缺乏耳镜检查方面的广泛培训（即，临床检查 鼓膜）。因此，临床医生往往错误的一边作出诊断的耳朵感染和处方 口服抗生素每年有超过800万种不必要的抗生素被处方，导致了 抗生素耐药细菌，并创造了最大数量的儿科药物相关的不良事件。 耳部诊断不准确的儿童经常被转诊到耳鼻喉科进行耳管手术， 复发性感染，其中高达70%的病例没有指征。诊断耳部病变仍然取决于 临床医生的主观性，基于对鼓膜的短暂一瞥。这种诊断的主观性创造了一个关键的 这是降低医疗保健成本和减少耳病过度诊断和过度治疗的障碍。设备 以帮助更准确、一致和客观地诊断耳部病理。我们以前的工作 为开发机器学习方法提供客观的耳部诊断方法奠定了基础 使用数字耳镜计算机辅助图像分析。这个项目将大大扩大我们以前的 工作的首要目标是开发新的机器学习应用程序来分析鼓膜视频 使用数字耳镜收集，将结合鼓室测量、人口统计学和临床数据， 实现诊断的客观性。长期目标是提高临床医生对耳部疾病的诊断准确性， 使用新颖的计算机辅助方法。为了实现这些目标，我们提出了三个具体目标： 具体目标1将完善客观的计算机辅助图像分析（CAIA）软件，以区分 多处耳膜异常具体目标2将开发耳镜检查临床决策支持 通过将CAIA与其他数据源相结合，包括鼓室导抗法，人口统计学， 临床信息。具体目标3将确定耳镜检查临床决策支持系统 提高临床医生的诊断能力。

ABSTRACT Diseases of the ear, particularly acute otitis media (AOM) and middle ear effusions, are the most commonly treated childhood pathologies. The financial burden of ear disease is estimated at more than $3.2 billion per year. Because ear diseases are common, a significant problem is over-diagnosis and over-treatment, due to two factors. First, the subjective nature of diagnosing ear disease - based on a brief glimpse of the eardrum with an otoscope - makes an accurate diagnosis difficult, even for experienced primary care, emergency medicine, or ear, nose, and throat (ENT) physicians. Second, with a growing shortage of primary care physicians in the US, more Advanced Practice Providers (Nurse Practitioners and Physician Assistants) serve as first-line clinicians in primary care and emergency settings but lack extensive training in otoscopy (i.e., clinical examination of the eardrum). Consequently, clinicians often err on the side of making a diagnosis of ear infection and prescribing oral antibiotics. Over 8 million unnecessary antibiotics are prescribed annually, contributing to the rise of antibiotic-resistant bacteria and creating the largest number of pediatric medication-related adverse events. Children with inaccurate ear diagnoses are often referred to ENTs for surgical placement of ear tubes for recurrent infections, and up to 70% of these cases are not indicated. Diagnosing ear pathologies still depends on clinician subjectivity, based on a brief glimpse of the eardrum. This diagnostic subjectivity creates a critical barrier to decreasing healthcare costs and reducing over-diagnosis and over-treatment of ear disease. Devices are needed to assist in a more accurate, consistent, and objective diagnosis of ear pathology. Our previous work laid the foundation to develop machine-learning approaches to provide an objective approach to ear diagnosis using digital otoscopy computer-assisted image analysis. This project will dramatically expand on our previous work with the overarching goal of developing new machine learning applications to analyze eardrum videos collected with a digital otoscope, which will be combined with tympanometry, demographic, and clinical data, to achieve diagnostic objectivity. The long-term goal is to improve clinicians’ diagnostic accuracy for ear diseases, using novel computer-assisted approaches. To accomplish these goals, we propose three Specific Aims: Specific Aim 1 will refine an objective computer-assisted image analysis (CAIA) software to differentiate multiple eardrum abnormalities. Specific Aim 2 will develop an otoscopy clinical decision support system by combining CAIA with additional data sources, including tympanometry, demographic, and clinical information. Specific Aim 3 will determine how the otoscopy clinical decision support system improves clinicians’ diagnostic performance.