Clinical Development and Evaluation of a Deep Learning Approach to Improve Diagnostic Accuracy

PROJECT SUMMARY Introduction: PhotoniCare, Inc. is a medical device company developing the TOMi Scope, a handheld, optical imaging device for improved diagnosis of middle ear health. The purpose of this proposal is to establish and evaluate a machine learning approach to facilitate both: (1) ease and reliability of quality data capture in a pediatric population from users with a range of otscopy expertise, and; (2) assist interpretation of the TOMi Scope’s correlated otoscopy and depth-resolved images in order to enable improved diagnostic accuracy and, ultimately, effective management. Significance: Ear infections affect 93% of all children, yet they are one of the most poorly diagnosed (~50% accuracy) and managed diseases in all of medicine, resulting in high antimicrobial over-prescription and resistance development. Correctly identifying the absence or presence/type of middle ear effusion (MEE; fluid) through the non-transparent eardrum is critical to accurate diagnosis, and the limited current diagnostic tools suffer poor diagnostic adoption (7-38% reported use) and accuracy (50-70%) due to inherent subjectivity and dependence on user expertise. Therefore, there is a clear and unmet need for superior, objective screening, starting with a definitive yet easily and reliably usable diagnostic tool for this extremely prevalent yet poorly managed disease. Hypothesis: Applying a machine learning approach to TOMi Scope imaging guidance and diagnostic classification will facilitate both: 1) ease-of-use and reliable quality data collection improvement, and 2) accurate detection of the presence or absence of MEE, as well as classification of the type of infection, regardless of user experience. Specific Aims: (1) Collect labeled TOMi Scope data (otoscopy and depth-scan images) from 268 patients at pediatric offices affiliated with UPMC Children’s Hospital of Pittsburgh, (2) Achieve reliable usability of the TOMi Scope by guiding image capture using TOMi-net, a deep learning model, (3) Develop a multimodal deep learning model to provide diagnostic assistance using TOMi Scope otoscopy and depth-scan data. Commercial Opportunity: The TOMi Scope will provide physicians with a superior user experience and new, objective information, enabling better decision-making for antibiotic prescription and surgical intervention. This has the potential to impact the standard of care for ~1B children worldwide that experience ear infections, representing a multi-billion-dollar commercial opportunity.PROJECT NARRATIVE Ear infections (otitis media) are highly prevalent in the pediatric population and represent a significant clinical challenge due to the limitations of the gold-standard diagnostic tools, resulting in high antimicrobial prescription and consequent resistance development. Accurate detection and classification of effusion (fluid) in the middle ear is a critical element for this diagnosis, and for making informed medical treatment decisions, particularly regarding antibiotic stewardship. The long-term goal of this work is to reduce antibiotic resistance and healthcare costs through improving patient outcomes by addressing the low diagnostic accuracy and user experience issues of current subjective methods, with a novel, non-invasive imaging tool capable of quantitative depth-resolved measurements to not only visualize the underlying infection behind the eardrum, but also, with automated machine learning image analysis algorithms, minimize user experience dependence and variability.