Loss of Smell and Taste Can Accurately Predict COVID-19 Infection: A Machine-Learning Approach

Abstract

The COVID-19 outbreak has spread extensively around the world. Loss of smell and taste have emerged as main predictors for COVID-19. The objective of our study is to develop a comprehensive machine learning (ML) modelling framework to assess the predictive value of smell and taste disorders, along with other symptoms, in COVID-19 infection. A multicenter case-control study was performed, in which suspected cases for COVID-19, who were tested by real-time reverse-transcription polymerase chain reaction (RT-PCR), informed about the presence and severity of their symptoms using visual analog scales (VAS). ML algorithms were applied to the collected data to predict a COVID-19 diagnosis using a 50-fold cross-validation scheme by randomly splitting the patients in training (75%) and testing datasets (25%). A total of 777 patients were included. Loss of smell and taste were found to be the symptoms with higher odds ratios of 6.21 and 2.42 for COVID-19 positivity. The ML algorithms applied reached an average accuracy of 80%, a sensitivity of 82%, and a specificity of 78% when using VAS to predict a COVID-19 diagnosis. This study concludes that smell and taste disorders are accurate predictors, with ML algorithms constituting helpful tools for COVID-19 diagnostic prediction.

Keywords: COVID-19; SARS-CoV-2; machine learning; prediction model; smell; taste; visual analog scale.

Figure 1

Modelling framework for the analysis of symptom associations and COVID-19 prediction. Data from 777 patients were obtained from different hospitals in the South of Spain. (a) For the analysis of the association between the intensity reported for loss of smell and taste, along with other symptoms, and a COVID-19 diagnosis, a first model was derived using step-wise logistic regression (LR) with a holdout validation scheme, by splitting the sample into a training (75%) and a testing dataset (25%). The performance of the model was assessed through ROC analysis, with AUC, SE, PPV and NPV parameters being calculated for the holdout testing (25%) dataset. (b) For the analysis of the discrimination ability and predictive value of different symptom variable datasets, including categorical (D1), continuous visual analog scales VAS (D2), dichotomized VAS (D3) as well as simplified predictor datasets with a reduced number of symptoms (D4 and D5), a comprehensive 50-fold cross-validation scheme was designed by assessing three different ML algorithms (LR, RF, and SVM). The performance of the models obtained were calculated through the mean AUC, SE, SP, PPV and NPV values over the 50-cross validated estimates obtained for each model tested. LR = logistic regression. RF = random forest. SVM = support vector machine, ROC= receiver operating characteristic, AUC= area under the curve, SE = sensitivity, SP = specificity, PPV= positive predictive value, NPV= negative predictive value.

Figure 2

Violin plots showing the distribution of self-reported VAS scores for symptom intensity. The white dots depict the median value and the vertical gray lines the interquartile range (25th and 75th quantiles). Horizontal black lines represent the mean value. Symptoms accompanied with an asterisk (*) showed significant differences in the Student’s t-test between positive and negative groups: loss of smell, loss of taste, facial pain, cough, dyspnea (p < 0.0001), and nasal obstruction (p = 0.0311). Nasal discharge (p = 0.1481) and age (p = 0.0628) were not significantly different between groups.

Figure 3

Step-wise logistic regression model obtained under BIC criterion: (a) Forest plots for association between symptoms and COVID-19 diagnosis. Note that for this analysis, VAS numeric variables reporting the intensity of symptoms were dichotomized into two categories: being higher or lower than the corresponding VAS cutoff point. VAS cutoff points for each symptom were previously calculated using ROC analysis and are listed in Table 2. Error bars denote 95% CIs; (b) ROC curve for prediction of COVID-19 in the (25%) holdout testing dataset. AUC, SE, SP, PPV, and NPV mean values are shown together with their 95% CIs. The confidence region of the ROC curve is depicted in blue.

Figure 4

Comparison of accuracy for different predictor datasets and ML algorithms: the boxplots show the distribution of (a) AUC, (b) SE, and (c) SP parameters for different predictor datasets (D1–D5, see description in Table 4) and ML algorithms (LR, RF, and SVM). The red line represents the mean value (calculated over the 50-fold cross-validation estimates), and the edge boxes the 25th and 75th percentiles. The whiskers represent the minimum and maximum data values not considered outliers, and the outliers are plotted individually using the ‘+’ symbol in red.