A Three-Tier Artificial Intelligence Model for COVID-19 Triage Using Pharyngeal Images (Preprint)
Bibliographic record
Abstract
Background: SARS-CoV-2 remains a common cause of acute respiratory illness; however, symptom-based triage poorly discriminates it from other febrile conditions. A recently developed artificial intelligence (AI)-powered pharyngeal camera acquires pharyngeal images and clinical data to assist in influenza diagnosis; leveraging this workflow, we evaluated an adjunct AI algorithm (COVID-19-AI) that reports high, medium, or low suspicion to guide whether SARS-CoV-2 testing should subsequently be performed. Objective: This study aimed to report diagnostic accuracy outcomes and clinical utility of the COVID-19-AI as a triage support tool. Methods: We conducted a performance evaluation using a prospectively collected multicenter dataset from 26 Japanese institutions between December 2023 and March 2024. Patients with suspected influenza or COVID-19 were eligible. The COVID-19-AI algorithm, a stacked ensemble of a Swin Transformer and boosting models, was developed using pharyngeal images combined with routine clinical variables from 2133 patients, and it produced a 3-tier output. Classification thresholds were predefined to optimize clinical rule-out and rule-in utilities. Diagnostic performance was assessed in 696 independent patients against centralized reverse transcription polymerase chain reaction-confirmed SARS-CoV-2 infection under 2 prespecified operating criteria: inclusive (high or medium=positive and low=negative) and strict (high=positive and medium or low=negative). A subanalysis stratified accuracy by time from symptom onset (12-hour bins to 72 hours). Results: Among 696 analyzed participants (all Asian), 247 (35.5%) had reverse transcription polymerase chain reaction-confirmed SARS-CoV-2 infection. The COVID-19-AI categorized 12.4% (n=86), 72.8% (n=507), and 14.8% (n=103) patients as high, medium, and low suspicion, respectively. Under the inclusive criteria, sensitivity of COVID-19-AI was 93.9% (95% CI 90.4%-96.4%), specificity was 19.6% (95% CI 16.1%-23.5%), and negative predictive value was 85.4% (95% CI 77.6%-91.3%). Under the strict criteria, sensitivity was 24.7% (95% CI 19.6%-30.4%), specificity was 94.4% (95% CI 92.0%-96.3%), and positive predictive value was 70.9% (95% CI 60.7%-79.8%). Across 12-hour onset strata, sensitivity under the inclusive criteria remained ≥92.0% and specificity under the strict criteria remained ≥83.3%; no pronounced temporal trend was observed. Additionally, an integrated model using both pharyngeal images and clinical variables (area under the receiver operating characteristic curve [AUROC] 0.78) outperformed models using only clinical variables (AUROC 0.75) or images alone (AUROC 0.71); feature importance analysis further confirmed that pharyngeal image information was the most influential individual predictor, providing greater predictive value than any single clinical variable. Conclusions: Embedded within AI-powered pharyngeal camera workflows, a 3-tier AI suspicion output enables complementary operating behaviors-high sensitivity to rule out COVID-19 (inclusive criteria) and high specificity to support immediate infection control measures (strict criteria), although these criteria involve inherent trade-offs with low specificity and low sensitivity, respectively. Performance stability across onset times suggests robustness to symptom chronology, offering a standardized tool for clinical triage.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.002 | 0.005 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.003 | 0.002 |
| Open science | 0.002 | 0.001 |
| Research integrity | 0.002 | 0.001 |
| Insufficient payload (model declined to judge) | 0.008 | 0.002 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".