*3.2. RR and Body Temperature Measurements Using IRT at a Laboratory and Clinical Settings*

Figure 8 shows an example of the signal selection applied by the proposed method, which is detailed in Section 2. The mean and minimum temperature changes in each ROI are shown in Figure 8b,d. To determine the respiration signal from four signals, we calculated the SQI parameters, which included the PSD, ACR and CPSD of each signal (Figure 8c,e). Using the SQI parameters, we chose the respiration signal.

**Figure 8.** Determination of respiration signal applying nasal and oral breathing decision based on SQI. (**a**) Thermal facial image with ROI. (**b**) Mean and minimum temperature fluctuations in nasal area. (**c**) SQI parameter obtained by power spectral density (PSD), autocorrelation (ACR) and cross-power spectral density (CPSD) of nasal temperature changes. (**d**) Mean and minimum temperature fluctuations in oral area. (**e**) SQI parameter obtained by PSD, ACR and CPSD.

To evaluate the nasal or oral breathing decision based on SQI and MUSIC, we compared the proposed method with the raw temperature change in the nasal area applied to FFT, which is a general method for estimating RR using IRT. The ground truth of RR was measured using the respiratory effort belt. We performed 15 s measurement four times and obtained 88 pairs of RRs from 22 healthy control subjects, including 6 subjects with nose clip for instructing subjects to mouth breathing. A comparison of RR estimation is shown in Figure 9. Figure 9a shows the Bland–Altman plot of nasal temperature change. The 95% limits of agreement ranged from -7.60 to 7.99 bpm (standard deviation σ = 3.98) and the RMSE was 3.98. Figure 9c shows the scatter plot of nasal temperature change; the Pearson correlation coefficient was 0.53. Figure 9b shows the Bland–Altman plot of the proposed method. The 95% limits of agreement ranged from -2.97 to 3.67 bpm (standard deviation σ = 1.68) and the RMSE was 1.73. Figure 9d shows the scatter plot of the proposed method; the Pearson correlation coefficient was 0.87. The results showed that the proposed method can reduce the 95% limits of agreement from [−7.60, 7.99] bpm to [−2.97, 3.67] bpm.

**Figure 9.** Bland–Altman plots and scatter plots of RR obtained by infrared thermography (IRT) sensor and respiratory effort belt. (**a**) Bland–Altman plot of nasal temperature change under the application of FFT. (**b**) Bland–Altman plot of the proposed method applying nasal or oral signal selection using SQI and MUSIC. (**c**) Scatter plot of nasal temperature change under FFT application. (**d**) Scatter plot of the proposed method.

Facial temperature, which is estimated by ROI detection using sensor fusion, was also evaluated. The ground truth of the temperature was measured using an electric thermometer. From all subjects, which included 22 healthy control subjects and 41 patients with influenza-like symptoms, a comparison of temperature estimation is shown in Figure 10. Figure 10a shows the Bland–Altman plot of temperature. The 95% limits of agreement ranged from -0.45 to 2.56 ºC (standard deviation σ = 0.77) and the RMSE was 1.30. Figure 10b shows the scatter plot; the Pearson correlation coefficient was 0.71.

**Figure 10.** Bland–Altman plots and scatter plots of body temperature obtained by IRT sensor and electric thermometer. (**a**) Bland–Altman plot. (**b**) Scatter plot.
