Fig 1.
Flow chart diagram of the study design.
Best infrared parameters: MaxE (maximum temperature of the Eyes); mE (mean temperature between the eyes); MaxL (maximum temperature of the lateral of the nose).
Table 1.
Demographic and symptom profiles of subjects.
Symptoms with significant statistics differences between infected and non-infected for COVID-19 are in bold.
Fig 2.
Thermal image profile of the analyzed patients’ faces.
(A) The dotted line describes the region with the greatest significant difference between patients tested positive (PCR+) and negative (PCR-) for COVID-19; (B) Regions with greater sensitivity and specificity in the evaluation of the thermal image: MaxE (maximum temperature of the Eyes), MaxL (maximum temperature of the lateral of the nose), mE (mean temperature between the eyes), Nose (mean temperature of the nose surface), Mouth (mean temperature of the mouth) and (C) Evolution of facial thermal profile of a subject during the infection period (according to sequential PCR results), showing that IRT captured fever (T > 37.5°C) only in the first day of infection.
Fig 3.
Boxplots of the IRT parameters between PCR+ and PCR-.
The means were significantly higher for all IRT parameters, except TIC and Nose (see Table 2 for t-test results). Infected patients presented an overall higher face temperature then non-infected, however rarely within fever state (dashed line represents the temperature of 37.5°C). **** p<0.0001; ns–no statistical significance.
Table 2.
Means±SD, t-test results and ROC analysis results between infected (PCR+) and non-infected (PCR-) patients for COVID-19.
Parameters with best performance based on AUC comparisons after ROC analysis are presented in bold.
Fig 4.
Receiver operator characteristics (ROC) curve with the best IRT parameters.
Performance of each parameter is compared through their respective areas under the curve (AUC), 95% confidence intervals are also provided. Arrows indicate the criterion cutoff points. MaxE (maximum temperature of the Eyes), MaxL (maximum temperature of the lateral of the nose), mE (mean temperature between the eyes).
Table 3.
Criterion validity for all infrared parameters evaluated in this study.
The cutoff values represent the best trade-off between sensitivity and specificity in the 136 patients analyzed in the emergency room of HU-USP Hospital, São Paulo, Brazil.
Fig 5.
ROC curves and confusion matrix of predictive models for COVID-19.
(a) Logistic regression with max temperature of eye inner canthus (TIC) as the only predictor and (b) Random Forest model (mtry = 4, min_n = 3, trees = 1000), combining all thermal parameters as predictors. RF model showed vastly improved performance, evidencing that considering one single temperature spot (common practice in the infrared solutions currently) is problematic for COVID-19 detection.
Fig 6.
Feature importance analysis based on permutation technique (extracted from the Random Forest Model).
The most important features responsible for model classification were MaxL and MaxE.