# Infectious disease diagnosis by artificial intelligence (AI): Differences in patient backgrounds and symptoms between antigen test positives and novel AI-powered pharyngeal endoscopy test positives

## TL;DR

This study compares patients who tested positive for infectious diseases using a traditional antigen test versus a new AI-powered endoscopy system, finding differences in symptoms and testing timing.

## Contribution

The study introduces a novel AI-powered endoscopy system for infectious disease diagnosis and highlights its distinct clinical implications compared to conventional antigen testing.

## Key findings

- AI-positive cases tested earlier after symptom onset compared to antigen-positive cases.
- AI-positive cases were more frequently pediatric and showed higher rates of cough and fever.
- The AI system misdiagnosed 23% of COVID-19 cases as influenza due to overlapping symptoms and pharyngeal features.

## Abstract

This study aimed to identify differences in patient background and symptoms between individuals who tested positive using conventional rapid antigen (Ag) tests and those who tested positive using a novel artificial intelligence (AI)–powered pharyngeal endoscopy system. A total of 813 patients underwent both influenza/COVID-19 Ag testing and AI-powered endoscopic testing. We analyzed differences in patient characteristics and symptoms between the two test-positive groups. AI testing showed an overall percent agreement of 62% (95% confidence interval [CI] 58–66%) (442/713), a positive percent agreement of 47% (95% CI 40–53%) (125/269), and a negative percent agreement of 71% (95% CI 67–76%) (317/444) compared with Ag testing. Compared with Ag-positive cases, AI-positive cases exhibited a shorter interval from symptom onset to testing; median 18 hours (Interquartile range [IQR] 10–27) for AI+ and Ag-, 24 hours (IQR 18–41) for AI+ and Ag + , and 27 hours (IQR 17–47) for AI- and Ag+ (p < 0.001). In analyses comparing the AI+ and Ag- vs. AI- and Ag + , AI+ and Ag- were more frequently paediatric (<15 years old) (odds ratio [OR] 3.4 (95% CI 1.6-7.2), p = 0.001), tested earlier after symptom onset (<24 hours) (OR 2.6 (95% CI 1.3-4.9), p = 0.005), had contact with infected individuals (OR 4.6 (95% CI 2.2-9.3), p < 0.001), cough (OR 11 (95% CI 4.7-27), p < 0.001), and fever (≥38.0°C) (OR 5.6 (95% CI 2.8-11), p < 0.001), but showed lower frequencies of gastrointestinal symptoms (OR 0.2 (95% CI 0.05-0.9), p = 0.04). Notably, the AI system misdiagnosed 23% (23/99) of COVID-19-positive patients as influenza-positive, likely due to follicular lesions on the pharyngeal wall—a key feature used by the AI system for diagnosing influenza. These findings demonstrate the impact of differences in diagnostic methodologies between conventional Ag testing (which detects pathogen viral load) and novel AI testing (which assesses host immune response to viral infection) on the clinical characteristics of test-positive patients.

This cross-sectional study investigated differences in patient backgrounds and symptoms between influenza-positive cases diagnosed by rapid antigen tests and those identified by a novel artificial intelligence (AI)-powered endoscopy system recently introduced in Japan. Compared with antigen test positives, AI test positives were tested earlier after symptom onset and were more frequently paediatric and symptomatic. These findings suggest potential utility of the AI testing system at the point-of-care, particularly for patients tested during the early phase of illness, when antigen test sensitivity is reduced. These findings highlight the impact of differences in diagnostic methodologies—namely, conventional rapid antigen testing, which detects pathogen viral load, versus the novel AI approach, which evaluates host immune response—on the clinical characteristics of test-positive patients. In addition, we observed cases in which the AI system misdiagnosed COVID-19 as influenza, likely due to symptomatic similarities between the two infections. Together, these findings underscore the promise of AI-based diagnostic systems for infectious diseases, while emphasizing the need for further deep learning-based improvements to enhance differentiation among respiratory pathogens.