An olfactory self-test effectively screens for COVID-19

The study addresses the need for wide-scale, rapid, low-cost COVID-19 screening methods that do not require handling physical specimens. Existing alternatives like AI-based cough/voice analysis and online symptom checkers can misclassify individuals with non-COVID illnesses and fail to detect individuals with asymptomatic COVID-19. Given that olfactory dysfunction is a prevalent and early symptom of COVID-19, the authors hypothesized that a sensitive, objective olfactory perception-based measure—specifically, the olfactory perceptual fingerprint (OPF)—could detect subtle, even subconscious, olfactory alterations. They developed an online tool for participants to rate intensity and pleasantness of household odorants and tested whether these perceptual measures could indicate COVID-19 status at both population and individual levels, including asymptomatic cases and symptomatic non-COVID cases.

Prior work established olfactory loss as a common and early marker of COVID-19 and validated symptom-based predictors. Large consortia (e.g., GCCR) showed smell and taste impairments strongly associate with COVID-19. Symptom checkers can predict infection but struggle with asymptomatic infections and symptomatic non-COVID illnesses. AI cough/voice tools are proposed but face similar limitations. Olfactory testing has established clinical tools (e.g., UPSIT, Sniffin’ Sticks), and the concept of olfactory perceptual fingerprints (OPFs) leverages perceptual distances or descriptor-based summaries to capture individual olfactory profiles. Pleasantness and intensity are primary dimensions of olfactory perception and are relatively stable across cultures. This background motivated applying OPFs to COVID-19 screening.

Design and participants: An online tool (www.smelltracker.org) was built in Drupal and translated into 15 languages. Recruitment was via media outreach by an international consortium; participation was anonymous and approved by the Wolfson Hospital Helsinki Committee (#0066-20-WOMC). Initial dataset: 12,020 participants (March 25–September 23, 2020). A later independent asymptomatic-only dataset added 1,464 participants, totaling 13,484 participants overall. Demographics: 7,189 women (mean age 44.32 ± 14.28), 4,831 men (mean age 45.23 ± 15.29), from 134 countries. COVID-19 self-reported status: positive (C19+), negative (C19−), unknown (C19-UD). Symptom checklist included fever, cough, dyspnea, tiredness, aches, runny nose, sore throat, loss of smell, loss of taste, or no symptoms. Procedure: On first use, participants selected five odorants from 71 common household options across five categories (two with lower trigeminal components, three higher). They then rated each odorant for intensity (very weak–very strong) and pleasantness (very unpleasant–very pleasant) on 0–100 visual analogue scales. The same odorants were prompted on subsequent uses. After ratings, participants reported COVID-19 test status and current symptoms. Participants received feedback graphs and later (post-study) an estimate of resemblance to C19+ or C19− based on OPF. Data: Initial 12,020 participants provided 171,500 ratings for 60 odorants (11 odorants unrated). Across the entire study, 13,484 participants (462 C19+) provided 178,820 ratings. Statistical analysis: Analyses were conducted in Matlab. For initial comparisons, 23 odorants with ≥25 C19+ ratings were retained. Normality was limited; therefore, two-sided Kolmogorov–Smirnov tests compared intensity and pleasantness distributions (C19+ vs C19−). For individual odorants, non-parametric Chi-squared tests compared C19+ and C19−; effect sizes were estimated using Eta squared. Country-level time series: daily cases (Johns Hopkins) were smoothed with a 7-day moving average; national inverse-intensity signals (100 – mean intensity) were smoothed with a 5-day moving average and cross-correlated with cases across lags (−14 to +14 days) to select maximal correlation; Pearson correlations were reported. ROC analyses used standard methods with bootstrapped confidence intervals (1,000 iterations) and non-parametric AUC comparisons (StAR). OPF construction: Descriptor-based OPFs used relative differences from group means for each odorant–descriptor pair (intensity and pleasantness), averaged across the participant’s five odorants to yield a 2D representation. For asymptomatic classification, a linear SVM was trained on balanced subsets and tested on independent subsets; repeated 500 times with unique participants in test sets. Implementation: Feedback presented resemblance to C19+ vs C19− as percentages to encourage formal testing while avoiding diagnostic claims within regulatory constraints.

- Group-level olfactory changes: C19+ participants exhibited significantly altered intensity (D = 0.45, p = 3.96e−233, corrected) and pleasantness (D = 0.31, p = 2.36e−114, corrected) distributions versus C19− and C19-UD. - Odorant-specific differences: For the 23 odorants with ≥25 C19+ raters, intensity ratings differed significantly between C19+ and C19− for all odorants (all χ2 ≥ 7.6, all p < 0.0058, all Eta2 > 0.08). Pleasantness differed for 17/23 odorants (all χ2 > 8.69, all p < 0.0188, Eta2 > 0.02). - Population-level correlations: In countries with ≥250 respondents and ≥10 diagnosed participants (n = 8), inverse intensity ratings significantly tracked daily COVID-19 case counts in 7/8 countries after optimal lag adjustment (Pearson r, FDR-corrected): Israel r = 0.68; Sweden r = 0.55; Portugal r = 0.47; Brazil r = 0.39; UK r = 0.26; Japan r = 0.25; USA r = 0.25; France r = −0.14 (ns). - Individual-level single-odorant classification: • Basil: AUC = 0.91. At TPR 62%, FPR 5% (62% sensitivity, 95% specificity, p = 0.03 corrected). At TPR 79%, FPR 13% (79% sensitivity, 87% specificity, p = 0.0043 corrected). • Cumin (C19+ vs C19− and C19-UD): AUC = 0.83; at TPR 77%, FPR 16% (77% sensitivity, 84% specificity, p < 0.00001 corrected). • Olive Oil: AUC = 0.79 (n = 5,167 total; 120 C19+). At TPR 77%, FPR 28% (77% sensitivity, 72% specificity, p < 0.00001 corrected). - Comparison to symptom-based classification: • Same Olive Oil cohort: olfaction AUC 0.79 vs symptom-based AUC 0.77 (AUC diff = 0.02, p = 0.29). • Symptomatic-only subset (n = 2,627; 115 C19+): olfaction AUC 0.70 vs symptoms AUC 0.59 (AUC diff = 0.11, p = 0.00022). Example operating point: TPR 75%, FPR 40% (75% sensitivity, 60% specificity). - Asymptomatic classification using OPF: • Initial dataset: OPF AUC 0.74 in asymptomatic participants (training 23 C19+ / 23 C19−; testing 10 C19+ / 250 C19−; repeated 500 times); symptoms AUC = 0.5 by definition. At TPR 70%, FPR 31% (70% sensitivity, 69% specificity, p < 0.001). • Independent later asymptomatic dataset (03/25/2020–04/10/2021; 114 C19+, 1,350 others): OPF AUC 0.66; at TPR 67%, FPR 36% (67% sensitivity, 64% specificity, p < 0.001), significantly better than symptoms (AUC 0.5). - Overall: Olfactory self-testing indicates COVID-19 at population and individual levels, including asymptomatic cases, and can outperform symptom checkers in key scenarios. The abstract highlights a representative operating point of 79% sensitivity and 87% specificity.

The findings support the hypothesis that olfactory perception, quantified through intensity/pleasantness ratings and OPFs, can indicate COVID-19 status without requiring physical samples. At the population level, inverse intensity ratings closely tracked national case rates across multiple countries, suggesting utility as an adjunct to symptom-tracking for public health monitoring. At the individual level, single odorants (e.g., Basil, Olive Oil) provided substantial discriminative power, and OPF-based classifiers detected COVID-19 even among completely asymptomatic individuals, where symptom checkers inherently fail. In symptomatic individuals without COVID-19, olfactory testing also reduced false positives compared to symptom checkers. The approach leverages cross-culturally stable dimensions of olfactory perception and relative scoring to mitigate cultural variability. Implementation through smelltracker.org provides rapid, scalable screening feedback, guiding individuals toward confirmatory testing while complementing existing tools. Mechanistic considerations discussed include peripheral and central pathways for COVID-19-induced olfactory dysfunction, but definitive mechanisms remain unresolved.

An online, household odorant-based olfactory self-test can effectively screen for COVID-19 at both population and individual levels, including asymptomatic cases where symptom checkers underperform. Single-odorant intensity ratings yield strong discrimination in some cases, and OPF-based analysis provides broader robustness and reduced susceptibility to user bias. The tool offers a rapid, low-cost, and scalable complement to symptom checkers and could inform public health responses. Future work should (1) conduct clinical validation against gold standards and WHO Target Product Profiles, (2) expand datasets for asymptomatic and vaccinated populations, (3) refine algorithms across cultures and odorant selections, (4) evaluate longitudinal testing regimens to enhance accuracy and containment strategies, and (5) explore mechanistic links between specific odorant perceptual changes and COVID-19 pathophysiology.

- The study is a basic science effort rather than a clinical validation; it does not address WHO Target Product Profiles or gold-standard diagnostic benchmarks. - Several analyses rely on subsets (e.g., specific odorants), reducing statistical power. - Self-selected participants may introduce sampling bias (e.g., country, sex distribution), though OPF’s relative metrics likely mitigate some effects. - COVID-19 test statuses were self-reported without formal verification; misreporting would add noise and likely weaken observed effects. - RT-PCR is an imperfect reference; predictions target RT-PCR results, not definitive infection status. - Lack of timing relative to diagnosis/symptom onset limits interpretability, given RT-PCR sensitivity declines over time. - The tool is unsuitable for assessing ongoing infection risk in those with persistent olfactory deficits post-COVID-19. - Vaccination status was not captured during data collection; evolving dynamics may alter olfactory loss patterns and tool utility. - Single-odorant tests are susceptible to user interference; OPF reduces but does not eliminate this limitation.