Mobile Digital Pupillometry for Rapid Triage of Patients With Severe Traumatic Brain Injury. (1st September 2019)

• Record Type: Journal Article
• Authors: McGrath, Lynn B; Eaton, Jessica C; Law, Anthony; Mariakakis, Alex; Patel, Shwetak; Levitt, Michael R
• Is Part Of: Neurosurgery. Volume 66(2010)Supplement 1
• Publication Date: 2019-09-01
• Subjects: Nervous system -- Surgery -- Periodicals; 617.48005
• DOI: 10.1093/neuros/nyz310_844
• Languages: English
• ISSNs: 0148-396X

Abstract

Abstract: INTRODUCTION: Traumatic brain injury (TBI) is the leading cause of mortality in people under age 45 and accounts for 2.5 million ED visits and $75 billion in healthcare costs each year in the United States. The key to ensuring the best possible clinical outcome for TBI patients is to facilitate their care at a designated trauma center. Unfortunately, up to 60% of severe TBI patients are undertriaged and admitted to non-trauma hospitals, a systemic problem which the National Study on the Costs and Outcomes of Trauma has demonstrated results in an excess mortality of 25%. Evaluation of the pupillary light reflex (PLR) is a crucial factor in triaging TBI patients, but penlight-based manual pupillometry is known to be inaccurate and digital infrared pupillometry impractical for field use. PupilScreen, a pupillometry technology developed for smartphones, integrates the convenience of manual pupillometry with the accuracy of a digital infrared pupillometer and may represent a practical way to improve the triage of severe TBI patients. METHODS: A smartphone-based pupillometer, PupilScreen, was developed utilizing a fully convolutional neural network (FCNN) and the flash and camera from a typical iPhone. The FCNN was trained for 175 000 iterations on a database consisting of 4410 pupil image samples from 42 healthy volunteers and 7 TBI patients. Once sufficient accuracy in generating PLR curves was achieved (0.3 mm mean error in diameter estimation), the trained network analyzed n = 48 new samples (n = 24 from healthy patients and n = 24 from patients with severe TBI [GCS < 8]). PLR curves were generated for each cohort and presented to 2 blinded clinicians to assess each pupil as either "responsive" or "non-responsive". RESULTS: When presented with PLR curves, one blinded clinician was able to correctly classify a patient as demonstrating "responsive" or "non-responsive" pupils with 100% accuracy, and the other with 99% accuracy (one "non-responsive" curve was misclassified as "responsive"). While limited in scope, these results demonstrate substantial improvement over manual pupillometry. CONCLUSION: A smartphone-based pupillometer can provide PLR assessments accurate and intuitive enough for clinicians to identify one of the clinical hallmarks of TBI, and may be useful for patient triage.