PC3001: A Novel Method to Identify Streptococcus Pharyngitis Using Smartphone Images and Artificial Intelligence

Background: Recent studies show artificial intelligence(AI) can identify streptococcal pharyngitis (strep throat) from a smartphone image, which has implications for telehealth. However, these studies lacked children (who more commonly experience strep throat), compared diseased vs. asymptomatic patients, or used hardware devices, which limits generalizability to telehealth.

Aim 1: To demonstrate feasibility of an AI algorithm to differentiate strep from non-bacterial pharyngitis using smartphone images of pediatric and adult throats. Aim 2: To compare the AI algorithm to Centor and McIsaac scores, and the treating clinicians’ prediction of strep throat.

Methods: This cross-sectional study was conducted at urban, tertiary care adult and pediatric emergency departments, May ‘21-June ‘22, with IRB approval. In this convenience sample patients were included if a rapid antigen strep test or throat culture was ordered. The ground truth was defined as: positive strep throat if the rapid antigen or culture were positive; negative strep throat if the culture was negative. Centor/McIsaac scores and throat image(s) were recorded on a secure smartphone. If the strep throat test had not yet resulted, the clinician was asked their prediction of the strep throat test outcome.

The throat images were inputs used to train a deep convolutional neural network; the output was a binary prediction of positive/negative for strep throat. 1-3 images per patient comprised the dataset. The dataset was randomly divided for each phase of algorithm development: ~50% (116 images) for training, ~25% (59 images) for validation, and ~25% (71 images) for testing. Results of the AI algorithm characteristics are reported on the testing image dataset.

Outcomes were sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) of the AI algorithm, Centor score, McIsaac score, and clinician prediction.

Results: In total, 193 patients were approached, 4 declined consent/assent, and 5 had missing test data; 185 patients were included in the analysis. Patient, clinical score, and clinician characteristics are noted in Table 1. 83 clinician predictions were recorded; for the remainder, the strep throat test resulted before the provider could offer a prediction. The overall prevalence of strep throat was 25.4%.

The AI algorithm has a specificity similar to a Centor score of > 3 (80% vs 85.5%, respectively), and exceeds the sensitivity, PPV and NPV of both Centor > 3 and clinician predictions. (Table 2)

Conclusion: This study demonstrates feasibility of an AI algorithm to predict strep throat in pediatric patients using a smartphone image. The current image-based AI algorithm is comparable to Centor scores and clinician predictions for sensitivity, specificity, PPV and NPV, which has implications as a clinical decision support tool. Future studies will focus on strategies to enhance AI performance such as an increased sample size or incorporating Centor criteria into the predictive model.