Development and evaluation of facial recognition for unique patient matching in a resource-limited setting

Abstract

Proper patient identification is pivotal for day-to-day operations of health care organizations. In low-and-middle-income countries (LMICs), the lack of unique patient identifiers remains challenging. Patients often have multiple IDs, with national IDs not readily available for younger populations or foreigners. Further, the identification system for patients varies between health- care facilities. Even when patients have identification numbers, there are numerous cases where individuals present to care without identifying documents. Patient misidentification often results in misdiagnosis and sometimes sentinel events. Historical probabilistic and deterministic match- ing approaches based on patient demographics have also proven suboptimal in LMIC settings. Biometric solutions offer a potential approach for unique patient identification, but these have not been rigorously evaluated within health care settings in LMICs. The study objective was to evaluate facial recognition biometric approach for unique patient identification and matching in a resource-limited setting. The specific objectives were to develop, implement, and evaluate the performance of a facial recognition solution integrated to an Elec- tronic Health Record system within an HIV care clinic in the Academic Model Providing Access to Healthcare (AMPATH) program in Western Kenya. A facial recognition module, employing deep neural networks to match facial images stored in a patient demographic database, was developed within the AMPATH medical record system (AMRS). The system was programmed to take between 10 and 14 training facial images at the time of registration. The performance of the facial recognition system to identify a patient and retrieve their medical record, was evaluated using a convenient sample of adult consenting patients presenting for routine care at the AMPATH outpatient clinic. At registration, patient facial images were captured and stored in a database. At a different station within the clinic (akin to a next visit or presentation at a distinct part of the care institution), facial images for the patients were then matched against those in the database. Accuracy of facial recognition was evaluated using standard measures, namely: Sensitivity; False Acceptance Rate (FAR); False Rejection Rate (FRR); Failure to Capture Rate (FTC) and Failure to enroll rate (FTE). A total of 103 patients (mean age 37.8; SD 13.6; 49.5% female; 7% with spectacles) were enrolled. On average 13.0 training images (SD 1.1) per participant were captured. For all participants, the system had a sensitivity of 99.0% at accurately identifying a patient. FAR for the system was <1% (0.0097), FRR was 0.00, FTC was 0.00 and FTE was 0.00. Wearing spectacles did not affect performance. The facial recognition system correctly and accurately identified almost all patients during the first match. Care systems needing to match patients accurately should strongly consider facial recognition as a potential approach for adults, in settings without unique patient identifiers.