*3.3. Classification of Healthy Control Subjects and Influenza Patients*

SVM established a classification model using three vital signs, including HR, RR and temperature, estimated by RGB and IRT sensors. The vital signs were measured for 22 healthy control subjects and 28 influenza patients (45.5 years of average age) diagnosed as influenza using virus isolation from all 41 patients with influenza-like symptoms. Figure 11a illustrates the distribution of the vital signs (22 blue dots: healthy control subjects, 28 red dots: influenza patients) and the separating hyperplane obtained by SVM using all data. SVM classification using the three vital signs achieved more accurate screening than fever-based classification (Figure 11b). Figure 11c presents the result obtained through leave-one-out cross-validation. The sensitivity, specificity, NPV and PPV were 85.7%, 90.1%, 83.3% and 92.3%, respectively. The fever-based screening using an electric thermometer was adopted to compare SVM classification. The sensitivity and specificity were 60.7% and 86.4%, respectively.

**Figure 11.** Classification model based on Support Vector Machine (SVM). (**a**) SVM classification. (**b**) Confusion matrix.

#### **4. Discussion and Conclusions**

The outbreak of 2019-nCoV was first reported in Wuhan, China, in December 2019 and was confirmed to have spread to more than 110 countries as of March 2020. When such a novel virus outbreaks, enhanced public health quarantine and isolation is essential. For this purpose, we developed a multiple vital sign measurement system for the mass screening of infected individuals in places of mass gathering. In this study, we focused on developing our system to measure three vital signs, to achieve automation, stability and swiftness for practical use in real-world settings. From a technical perspective, we proposed specific signal and image processing methods for highly reliable vital sign measurements and compared them with conventional methods (Tables 1 and 2). Tapered window, RGB signal reconstruction and MUSIC were applied for HR measurement. Automatic ROI tracking using sensor fusion and nasal or oral breathing selection using SQI and MUSIC were applied for HR measurement. The proposed method showed agreement with their reference devices (HR: [−10.4, 12.6] bpm, RR: [−2.97, 3.67] bpm, temperature: [−0.449, 2.56] ◦C). The reliability and stability of our system on vital sign measurement were significantly improved.


**Table 1.** Comparison of proposed RGB signal reconstruction method with conventional green trace method on HR measurement.

**Table 2.** Comparison of proposed Nasal/oral SQI method with conventional nasal alone method on RR measurement.


Moreover, we tested multiple vital sign-based screening in a laboratory and a clinic. The proposed method's sensitivity and specificity (85.7%, 90.1%) were found to be higher than those of fever-based screening (60.7%, 86.4%). The tendency of the three vital signs measured by healthy control subjects and influenza patients is shown in Figure 12. The medians of facial skin temperature of influenza patients and healthy control subjects were 37.3 and 35.5 ◦C, respectively. The medians of HR of influenza patients and healthy control subjects were 99.3 and 76.4 bpm. The medians of RR of influenza patients and healthy control subjects were 18.9 and 14.0 bpm. Each vital sign of patients with influenza was found to be elevated. This contributed to improvement in SVM classification based on the three vital signs.

**Figure 12.** Box plot of vital signs between influenza patients and healthy control subjects. (**a**) Facial skin temperature. (**b**) HR. (**c**) RR.

However, the proposed method has some limitations. The ROI detection of sensor fusion may fail when the background has the color of skin or hair. In terms of the classification test based on SVM, the facial skin temperature may include the influence of the ambient environment. The measurement environment at a laboratory is different from that at a clinic, even at the same ambient temperature. This causes a difference in facial skin temperature regardless of the seasonal influenza. Therefore, we need to develop environment-invariant temperature estimation using an IRT.

In conclusion, we proposed automatic, stable and rapid HR, RR and body temperature measurements using an RGB-thermal sensor and its application for the screening of infectious diseases. This method introduces (1) the sensor fusion approach for the detection of detailed facial landmarks in a thermal image, (2) HR estimation, which introduces tapered window, signal reconstruction and MUSIC and (3) RR estimation, which implements nasal or oral breathing selection using SQI and MUSIC. Moreover, we demonstrated a classification model based on SVM using healthy control subjects and patients with seasonal influenza. The results indicate that the proposed method is indispensable for the high performance of contactless multiple vital sign measurements for infection screening.

**Author Contributions:** Conceptualization, G.S., S.A. and T.M.; methodology, T.N., T.M., H.L., M.K., T.K. and G.S.; software, T.N., G.S.; validation, T.N., G.S., S.A. and T.M.; formal analysis, T.N., G.S.; investigation, G.S.; resources, G.S.; data curation, T.N., G.S., S.A.; writing—original draft preparation, T.N., G.S.; visualization, T.N., G.S.; supervision, G.S.; project administration, G.S.; funding acquisition, G.S. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research was funded by in part by the JSPS KAKENHI Grant-in-Aid for Scientific Research (B) under Grant 19H02385, The Okawa Foundation for Information and Telecommunications and in part by the National Science Foundation Program of China under Grant 61801149.

**Conflicts of Interest:** The authors declare no conflict of interest.
