key: cord-0954739-ltjl3kj7
authors: Karni, Noam; Klein, Hadar; Asseo, Kim; Benjamini, Yuval; Israel, Sarah; Nimri, Musa; Olstein, Keren; Nir-Paz, Ran; Hershko, Alon; Muszkat, Mordechai; Niv, Masha Y
title: Self-rated smell ability enables highly specific predictors of COVID-19 status: a case control study in Israel
date: 2020-12-28
journal: Open Forum Infect Dis
DOI: 10.1093/ofid/ofaa589
sha: 4bebdf57384f8e58a96031ad58aa57cf03802416
doc_id: 954739
cord_uid: ltjl3kj7

BACKGROUND: Clinical diagnosis of COVID-19 is essential for detection and prevention of COVID-19. Sudden onset of taste and smell loss is a hallmark of COVID-19, and optimal ways for including these symptoms in the screening of patients and distinguishing COVID-19 from other acute viral diseases, should be established. METHODS: We performed a case-control study on patients that were PCR-tested for COVID-19 (112 positive and 112 negative participants), recruited during the first wave (March 2020 – May 2020) of COVID-19 pandemic in Israel. Patients reported over by phone their symptoms and medical history and rated their olfactory and gustatory abilities before and during their illness on a 1-10 scale. RESULTS: Changes in smell and taste occurred in 68% (95% CI 60%-76%) and 72% (64%-80%), of positive patients, with 24 (11-53 range) and 12 (6-23) respective odds ratios. The ability to smell was decreased by 0.5±1.5 in negatives, and by 4.5±3.6 in positives. A penalized logistic regression classifier based on 5 symptoms has 66% sensitivity, 97% specificity and an area under the ROC curve of 0.83 (AUC) on a hold-out set. A classifier based on degree of smell change only is almost as good, with 66% sensitivity, 97% specificity and 0.81 AUC. The predictive positive value (PPV) of this classifier is 0.68 and negative predictive value (NPV) is 0.97. CONCLUSIONS: Self-reported quantitative olfactory changes, either alone or combined with other symptoms, provide a specific tool for clinical diagnosis of COVID-19. A simple calculator for prioritizing COVID-19 laboratory testing is presented here.

M a n u s c r i p t Graphical abstract M a n u s c r i p t BACKGROUND In December 2019, the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) was reported in Wuhan, China [1] . The resulting coronavirus disease COVID-19 has become a global pandemic with 16.5 million reported cases as of July 29th, 2020 (World Health Organization, 2020).

Since March 2020, an increasing number of reports regarding taste and smell loss in COVID- 19 infections appeared in preprints [2, 3] and in general press, and it is currently well established that taste and smell loss is common in COVID-19 patients [4] [5] [6] [7] .

In a recent crowd-sourced study, ~7000 app users reported testing positive for COVID -19, with 65% of those reporting that they lost their sense of smell or taste [8] , a three-fold increase in prevalence compared to COVID-19 negatives [9] . The severity of smell and taste loss in COVID-19 patients is striking: these sensory abilities were reduced by -79.7 ± 28.7, -69.0 ± 32.6 (mean ± SD), respectively, as reported by about 4000 participants using a 0-100 visual analog scale (VAS) [6] . A follow-up study suggested that recent smell loss is the best predictor for COVID-19 [10] .

Here we assess the prevalence of different COVID-19 symptoms as well as the degree and additional characteristics of smell and taste changes in PCR-swab tested COVID-19 positive vs COVID-19 negative patients. Importantly, patients were recruited in a manner that did not disclose the underlying chemosensory questions in this study. We used these data to develop a classifier that can prioritize patients for PCR-testing, help epidemiological investigations, and screen large populations. The participants were not aware that the questionnaire will include smell and taste loss symptoms prior to their agreement to partake in the study. Informed consent was obtained from all participants. The study was approved by the Hadassah Medical Center Helsinki Committee (permit number 0236-20-HMO).

The interviews were carried out over the phone. The questionnaire is based on questions compiled by physicians and scientists in the Global Consortium for Chemosensory Research, GCCR [6] . The full questionnaire is included in Supplementary data and has five parts: 1) General information (e.g., age, gender); 2) Medical history (e.g., medical conditions, medications, changes in taste/smell in the past, pregnancy, contact with a confirmed patient); 3) Current illness: 23 physical signs and symptoms, including binary question (yes/no) on smell, taste and chemesthesis (cooling, burning, tingling sensation), PCR swab test results and dates, date of exposure to confirmed COVID-19 patient, subjective recovery feeling; 4) Smell: Participants were instructed to rate their sense of smell/ taste and the degree of their nose blockage on a scale from 1-10 (1 corresponding to "no sense of smell" and 10 to excellent sense of smell) and similarly rate the ability to breathe through the nose A c c e p t e d M a n u s c r i p t before/during/after illness. Blocked nose rating was used to test the plausible hypothesis that it causes a change of smell; 5) Taste (e.g., rating of taste ability before/during/after illness, as described for smell), experience of strange/bad taste in the mouth, change in sensitivity to irritants (chemesthesis) and change in basic taste modalities -sweet, salty, sour, bitter, each elicited by non-volatile compounds via specific receptors or channels expressed in dedicated taste receptor cells [11] . The fifth basic taste modality, umami or savory, was not used because it does not have a Hebrew translation. "Other" taste was available as an additional optional answer. Data was kept in Compusense Cloud on-line software (Compusense Inc., Guelph, ON, Canada).

Log-odds for the individual symptoms were calculated over the full dataset.

Confidence intervals and p-values for the log-odds were estimated from the glm function using the logistic link implemented in the statistical software R (https://www.r-project.org/). Classifiers were trained from the reported symptoms to evaluate the separation between COVID-19 positive and COVID-19 negative patients. The classifiers were trained on a random subset of 2/3 of the data (the training set, 148 samples), and evaluated on the remaining samples (the test set). Sampling of the train and test sets was stratified by COVID-19 status. We trained the classifier on the full symptom matrix: All symptoms of question 23 in the questionnaire (see Supplementary) were included, except "no symptoms" or "other".

All eye symptoms were combined to "Eye symptoms". Also added were quantitative questions for taste, smell, and nose blockage (rating before the illness minus rating during the illness, questions 31, 35, 37, 38, 40, 41 in the questionnaire) and chemesthesis (question 45) "Coated tongue", "Dizziness", "Ears pressure", "Eye burn", "Eye discharge", "Hearing A c c e p t e d M a n u s c r i p t change", "Lacrimation", and "Vision changes" were removed since these symptoms were reported by less than 10% of the subjects. The classifiers were trained as penalized logistic regressions, using the elastic net algorithm (α = 0.5 implemented in the glmnet package in the R environment. This regression method encourages sparse coefficient vectors, meaning that it is suitable in situations where only few coefficients are non-zero. The regularization parameter (lambda) was initially set using cross-validation, but then increased until the model included no more than six symptoms. For classifiers based on a single symptom, no regularization was used.

Classifiers were evaluated using the hold-out test set. The score from the classifier was thresholded at zero, so that patients with score exceeding 0 were called positive by the classifier. Sensitivity (predicting COVID-19 positives correctly) and specificity (predicting COVID-19 negatives correctly) metrics were calculated from the following formulas:

in which TP (TN) are the COVID-19 positives (negatives) classified correctly, and FN (FP) are the COVID-19 positives (negatives) classified incorrectly. Due to our balanced sample, accuracy is the average of sensitivity and specificity. We further computed the accuracy metrics that account for the expected proportion of positive cases in the tested population, namely the positive predictive value (PPV) and the negative predictive value (NPV). The scores obtained from the logistic classifier (s) were translated into probability to be positive (P) by adjusting for the proportion of COVID+ out of the tests (π) according to the following formulas:

We take π to be 0.08 because that was the proportion during data collection.

The ROC curve corresponds to true-positive and false-positive rates for different values of the threshold; the curve and the area under the ROC curve (AUC), which measures the degree of separability between positive and negative scores, were estimated using the pROC package [12] .

The sample size was calculated to allow detecting differences in smell-loss or taste-loss prevalence between COVID-19 positive and negative populations. Based on previous research [e.g. [13] ], we used conservative estimates of 60% prevalence in the positive population and 35% prevalence in the negative population. Power was estimated by Monte Carlo simulations, namely repeatedly (b=1000) resampling from two Binomial distributions corresponding to the positive population and the negative population. Assuming 100

individuals are assigned to each group, and a two-sided t-test is used, the probability of detection (power) is 92%. To be on the conservative side, we used somewhat larger samples.

Completed questionnaires were obtained from 112 COVID-19 positives and 112 COVID-19

negatives. The median age of the respondents was 35 ± 12 years for positives and 37 ± 12 years for negatives (mean ± SD) years. The positives group included more men (64%), while the negatives group was more balanced (48% males). Seven patients classified as hospitalized (received respiratory support during their hospitalization and / or were A c c e p t e d M a n u s c r i p t hospitalized in the intensive care unit) and the rest 217 were classified as ambulatory patients.

Signs and symptoms that appeared in the binary part of the questionnaire (Supplementary Material, question 23 of the full questionnaire) and were found to occur in at least 10% of the positive patients are summarized in Table 1 . A few symptoms, including dry cough and sore throat, were prevalent in COVID-19 positives, but even more so in the negatives control sample.

Smell change, taste change, change in chemesthetic ability (perceiving spicy, tingling or cooling sensations) and muscle ache were significantly more prevalent in COVID-19 positive as compared to COVID-19 negative patients (68%, 72%, 31%, 62% vs. 8.0%, 18%, 6%, 34%, respectively) ( Table 1) . Other CDC recognized symptoms [14] , such as lack of appetite, fever, and diarrhea were approximately twice or three times more common among positives than negatives.

Nausea or vomiting, although considered a COVID-19 symptom, were not more common among COVID-19 positive as compared to COVID-19 negative patients. By contrast, lack of appetite, despite not being included as an "official" CDC symptom [14] , was found to be significantly more common in COVID-19 positive patients.

Taste and smell changes often, but not always, occur together: Figure A c c e p t e d M a n u s c r i p t 63% of the positive patients reported impairment of at least one of the four taste modalities (sweet, salty, sour, and bitter) compared to only 10% among the negative patients ( Figure   2B ). 6% of positives and 2% of negatives added comments about taste changes as free text.

Additional 31% of positives and 88% of negatives did not report any taste-related changes.

In COVID-19 positive patients with taste impairment, all four taste modalities were usually impaired.

In addition to the binary questions, the participants were asked to rate their smell and taste senses before and during their illness on a 1-10 scale.

As seen in Figure 3 , the change in smell and taste ability during disease, compared to a self- Combination of symptoms were next checked for ability to differentiate COVID-19 positive from negative diagnosis. To that end, several classifiers were trained based on 66% of the sample and evaluated on 34% that was kept as a holdout set. The process of selection of descriptors is outlined in Figure 4A . Relevant symptoms (n=30) were included as possible descriptors for the classifiers, and the elastic-net penalization was increased until no more A c c e p t e d M a n u s c r i p t than six symptoms were included in the model (number limited for practicality). The effect of excluding or including a particular symptom was evaluated in order to understand the importance of separating taste from smell and using binary vs. quantitative measures for each ( Figure 4A ).

The results of the evaluation on the holdout set are summarized in Supplementary Table S1 , and classifiers 1-3 can be seen in Figure 4B . Classifiers that did not use chemosensory symptoms had poor performance (AUC 0.60, black curve, classifier 1, and additional classifiers (Table S1 ). Adding the quantitative smell-change symptom (maroon curve, Classifier 2) is sufficient to outperform all other classifiers (AUC 0.83). Remarkably, using quantitative smell-change as a sole symptom (magenta curve, Classifier 3) resulted in a classifier that was nearly equally effective as Classifier 2 (AUC 0.81).

Adding taste change to Classifier 2 did not improve its performance, as it resulted in AUC of 0.82 (Classifier 7, Table S1 ). Taste change as a sole descriptor resulted in AUC of 0.75 (Classifier 15, Table S1 ) and as an added descriptor to other "Basic" symptoms, in AUC of 0.76 (Classifier 13, Table S1 ). Thus, while there is a high correlation (0.82) between quantitative changes in smell and quantitative changes in taste, the smell change descriptor outperforms the taste descriptor. Using the quantitative smell and taste descriptors resulted in higher AUC's than binary (yes/no) descriptors of these changes. For example, a binary smell descriptor used as a sole descriptor resulted in AUC of 0.78 (Classifier 16, 

We have established the prevalence and degree of decrease in taste and smell in patients who were eligible to receive PCR swab tests during the COVID-19 pandemic and found significant differences in PCR-positive and PCR-negative patients.

The change in smell ability is not related to nasal obstruction, as nose blockage was low, as shown before [6, 15] . Taste and chemesthesis changes strongly correlate with smell change (in agreement with [6] ). Taste changes are more common than smell changes in negatives, and chemesthesis changes are less common than taste and smell changes in positives, leading to odds ratios for these chemosensory modalities that are high, but lower than for olfaction.

All taste modalities in COVID-19 patients were impacted together (or not at all). This is of interest for understanding the pathophysiology of the disease: a recent study suggests CoV-2 infection of non-neuronal cell types expressing ACE2 and TMPRSS2 as the mechanism underlying COVID-19 related anosmia [16] , but the reason for COVID-19 ageusia is less clear [11, 17] . Our results support the idea of impairment of supporting cells or tissues, rather than of taste receptors cells Type 2, which express bitter, sweet and umami taste receptors or Type 3 which express sour sensing channels [18] . Routine addition of taste questions for patients screening is not warranted, as these did not contribute to the classifier performance. Nevertheless, patients with prior conditions of impaired olfaction (estimated 5% of population [19] ) require a suited classifier.

We present three versions of classifiers: the first one is based on four yes/no questions (muscle ache, lack of appetite, fever, sore throat) and the quantitative smell change. The second version uses only the quantitative smell change and has a similar AUC. The third one (suited for participants with pre-existing olfactory impairment) is based on five yes/no questions (muscle ache, lack of appetite, fever, sore throat, breath difficulty) and

quantitative taste change. The probabilities for COVID-19 based on these calculators are available via GitHub.

Our best-preforming classifier (Classifier 2, AUC 0.83) used quantitative smell change, muscle ache, lack of appetite, fever, and (negative contributing) sore throat. Performed in parallel to Gerkin et al. [10] , our study similarly included both binary and quantitative questions regarding taste and smell as two separate indicators. Our results for positive patients are in overall agreement with Parma et al. [6] and Gerkin et al. [10] . This is striking in 

The method of patient recruitment is one of the limitations of this study: social media-based recruitment may limit participants' representation as it targets mostly younger patients, with internet access and social media accounts. Word of mouth recruitment was used as well and contributes as well to creating a sample that is not necessarily representative of the general population. Male and female patients were not fully matched across positives (64% males) and negatives (48% males), in accord with higher % of males (56%) among

False negative results of the RT-PCR test have been reported to occur in ~30% of COVID-19

patients [21] . At the 3 weeks follow-up, none of the negative patients who were not recovered at the time of their initial questionnaire (~50) had repeated RT-PCR test, so false negatives could not be ruled out. Serology tests were performed for a sample of 5 negative patients who reported taste and/or smell impairments. All 5 had negative serology using the LIAISON® SARS-CoV-2 S1/S2 IgG assay [22] , confirming these patients are likely true negatives.

Importantly, our classifiers are not SNOUT ('Sensitive test when Negative rules OUT the disease') but can definitely be referred to as 'Specific test when Positive rules IN the disease' (SPIN).

Our sample was composed of slightly to moderately ill patients. It should also be kept in mind that our data is specific to Israeli patients and reflect to some degree the criteria for A c c e p t e d M a n u s c r i p t PCR tests eligibility during the recruitment (fever and dry cough were sufficient for PCR test but change of smell and/or taste alone was not).

The resurging pandemic puts the clinic and public health authorities in a scenario not usual for modern medicine -namely, the limited resources require or may require in the future, prioritization of testing and treatment. The fact that our sample contained PCR-positive and PCR-negative ambulatory patients, all suspected to have COVID-19 prior to PCR testing, enabled the development of symptoms-based classifiers.

Our results suggest that ranking of the ability to smell before and during the illness, is an excellent practical approach to identify COVID-19 positive patients offering reasonably high predictive capability (Specificity 97%, Accuracy 82%). Additional classifier is available for patients with prior olfactory impairments (Supplementary Material Figure S1 ).

Based on the classifiers developed in this work, we propose a simple calculator that can be used to prioritize testing (available at https://github.com/KimAsseo/Hadassah_COVID-19).

Additionally, high-performance classifier may potentially capture false negative PCR tests results of high scored individuals. The current study provides a practical tool for assessing potential COVID-19 patients.

M a n u s c r i p t M a n u s c r i p t Table of mean ± SD for COVID-19 positives and negatives in general and for those reporting changes of taste or smell. Scores for taste, smell, and nose blockage were evaluated on a 1-10 scale. P values for the difference in the magnitude of change between COVID-19 positives and negatives was calculated using a two-sided t-test. Different combinations of symptoms established better classifiers than Classifier 1, those using quantitative questions exhibiting better performance than those using binary ones. 

A Novel Coronavirus from Patients with Pneumonia in China

Anosmia and Ageusia as Initial or Unique Symptoms after SARS-COV-2 Virus Infection

Anosmia and dysgeusia in patients with mild SARS-CoV-2 infection

Smell dysfunction: a biomarker for COVID-19

COVID-19 and anosmia in Tehran

More than smell -COVID-19 is associated with severe impairment of smell, taste, and chemesthesis

Olfactory Dysfunction in COVID-19: Diagnosis and Management

Real-time tracking of self-reported symptoms to predict potential COVID-19

a preregistered, cross-sectional study

The cell biology of taste

pROC: an open-source package for R and S+ to analyze and compare ROC curves

Loss of smell and taste in combination with other symptoms is a strong predictor of COVID-19 infection

Symptoms of Coronavirus | CDC

Self-reported loss of smell without nasal obstruction to identify COVID-19. The multicenter CORANOSMIA cohort study

Non-neuronal expression of SARS-CoV-2 entry genes in the olfactory system suggests mechanisms underlying COVID-19-associated anosmia

COVID-19 and the Chemical Senses: Supporting Players Take Center Stage

Taste buds: cells, signals and synapses

Smell and taste disorders

COVID-19 diagnostics in context

S1/S2 IgG The fully automated serology test for the detection of SARS-CoV-2 IgG Antibodies

We thank the Global Consortium for Chemosensory Research (GCCR) team for inspiring the questionnaire used in this research. We thank Yehuda Tarnovsky for help with patient's recruitment. Graphical abstract was created with BioRender.com.

A c c e p t e d M a n u s c r i p t M a n u s c r i p t The classifier using "smell" and "taste" as separate descriptors, rather than "smell or taste" as a single joint descriptor showed better performance. The "Basic" + smell only descriptor outperformed the "Basic" + taste only descriptor, resulting in Classifier 2. Finally, the smell only descriptor was tested alone without all other "Basic" symptoms, resulting in Classifier 

The study was conducted in accordance with Helsinki committee and the required ethics approval was granted (reference number HMO-0236-20).

Written informed consents for publication of patients' clinical details were obtained from the patients.A c c e p t e d M a n u s c r i p t

Authors declare no conflicts of interest.

The authors declare that they have no competing interests. A c c e p t e d M a n u s c r i p t