StrepApp: A Deep Learning-Based Tool for Diagnosis of GAS Pharyngitis in Pediatric Patients

Corresponding Author Rana F. Hamdy, MD, MPH, MSCE; Children’s National Hospital; rhamdy2@childrensnational.org Funding This work is supported by the Advanced Research Projects Agency for Health (ARPA-H). Conflict of Interest T. Nguyen, A, Khazraei, R. Shekhar are employees of AusculTech Dx. R. Hamdy, J. Dome, R. Shekhar, T. Nguyen, Y. Arjoune are co-inventors of a pending patent application for StrepApp. Background Pharyngitis is a leading cause for pediatric ambulatory visits in the U.S., including in telehealth settings. Approximately one-third of cases are diagnosed as Group A Streptococcus (GAS) pharyngitis, for which antibiotic treatment is the standard of care. Because clinical signs and symptoms alone cannot reliably identify GAS, accurate diagnosis requires a throat swab for microbiologic testing. In telemedicine and other settings where microbiologic testing is not readily available, antibiotics are frequently overprescribed, contributing to unnecessary use and antimicrobial resistance. Objective This study aims to develop a real-time, deep learning-based mobile app that integrates clinical features with photographic pharyngeal images to rule out GAS pharyngitis. Methods Over a 14-month data collection period, patients aged 3-17 years were prospectively enrolled from eight pediatric urgent care or emergency department sites across the U.S. if they underwent evaluation for GAS pharyngitis with a throat swab test. Patients were excluded if GAS test results were unavailable or if they had received antibiotics within the preceding 48 hours. We used a custom mobile app to capture throat images along with clinical history (age, recent known contact with GAS, history of tonsillectomy), presenting symptoms (presence of sore throat, fever, dysphagia, abdominal pain, cough, congestion or rhinorrhea, scarlatiniform rash), clinical signs (lymph node enlargement or tenderness, tonsillar exudate), and GAS test results. We determined the association between each clinical feature and GAS test positivity using a mixed-effects multivariable logistic regression model, accounting for clustering by practice site. We then developed a multimodal deep learning method to distinguish GAS from non-GAS cases by integrating symptoms with cropped throat images, automatically processed using a novel segmentation algorithm. Our segmentation model was trained and validated to identify key anatomical regions—palate, tonsils, pharynx, and uvula – using a database of manually segmented throat images. Results The database included 2995 unique pediatric patients, with 42% testing positive for GAS. The GAS positivity rate varied by age (Fig. 1). In a multivariable logistic regression model, the presence of fever (aOR 1.48, 95% CI: 1.26-1.74), enlarged or tender cervical lymph nodes (aOR 1.67, 95% CI: 1.32-2.10), and recent known GAS contact (aOR 1.64, 95% CI: 1.35-2.00) were significantly associated with higher odds of GAS pharyngitis. Conversely, the presence of cough (aOR 0.67, 95% CI: 0.56-0.79), congestion or rhinorrhea (aOR 0.82, 95% CI: 0.68-0.97), and age > 14 years (aOR 0.27, 95% CI: 0.19-0.39) were associated with lower odds of GAS pharyngitis. Analysis of images was performed on 16 589 images deemed to be high quality images (37.7%). The segmentation model demonstrated high accuracy across anatomical structures of interest, with 95.6% accuracy for uvula detection, 89.6% for the tonsils, 86.3% for the pharynx, and 99.2% for the palate. The GAS diagnostic classifier achieved an accuracy of 89.3%, with a specificity of 96.6% and a negative predictive value of 93.6%. Discussion StrepApp has the potential to transform pediatric healthcare by enabling real-time, accurate GAS pharyngitis diagnosis using an AI-powered mobile app, reducing reliance on microbiologic tests that may be inaccessible in telemedicine settings.