In this study, we developed a core body temperature estimation formula using ECG signals rather than individual characteristics such as age. Within the heat strain decision aid (HSDA) core body temperature prediction model, environmental conditions, clothing, and metabolic heat production were three important elements that were mutually attributable to core body temperature; for each participant, rises in core body temperature resulting from exercise were estimated based on individual characteristics, clothing, the environment, and activity status. Notably, the HSDA is also used for training purposes in the US military, and is currently being improved through multiple studies; specifically, it is being applied to predict safe work continuation times, suitable rest/water intake times, and even the risk of heatstroke, thus providing a preventive measure against it [14]. The model used in this study considered 16 dependent variables, with height being the primary variable among them (e.g., ambient temperature, clothing characteristics, environmental variables, such as humidity and temperature, physical activity, and dehydration).
Constantly fluctuating data, such as those pertaining to physical activity volumes are ideally collected in real time using wearable devices, which are suitable for daily use and do not greatly restrict movement. As such, some companies have implemented laborer health-management programs by monitoring biodata with wearable sensors that use IoT technologies; this promotes preventive behavior, which is initiated based on real-time warnings in cases where physical conditions are poor and/or the risk of heatstroke is increased. As discussed by Lin et al. [15], however, it is unlikely that verification and proof-of-concept experiments based on physiological characteristics are sufficient for proving the efficacy of these types of systems. While biometric information is easy to obtain using wearable devices, there are also several limitations in terms of the required resources (e.g., power consumption, calculation power, and memory capacity) and communication status, which cannot always be ensured. In addition, power is needed to accomplish the so-called “synchronous” real-time processing, which simultaneously measures multiple biomarkers. Hence, it is important to study estimations that use as few biometrics as possible. From a practical standpoint, the estimation model used in this study is advantageous due to the employment of initial core body temperature values and ECG signals without the need for information pertaining to individual characteristics (e.g., the environment or clothing). In hot environments, the first physical response to heat exposure is dermal vasodilation, which dissipates heat outside the body by increasing the heat transfer rate of the body’s surface through an increase in dermal blood flow. Although this biological response is an efficient and effective response in mildly hot environments, excessive external heat is applied to the body under harsher temperature conditions, thus hindering the heat dissipation function. The blood also tends to accumulate in the extremities due to vasodilation, which causes a drop in both blood pressure and cerebral blood flow, the latter of which may result in dizziness, nausea, and fainting. As such, the body maintains its blood pressure by increasing its heart rate and cardiac output. In other words, heart rate is closely related to the thermoregulatory mechanism. From a physiological point of view, it is reasonable to use SD2 as an index for estimating the deep body temperature model because it shows a significant time-series change in heart rate. If system efficacy is ensured through this proposal, then it may be proactively used in high-risk workplaces as an effective countermeasure against heatstroke. Further, the use of various biometric sensors can not only prevent major health disorders resulting from heatstroke, but should also facilitate a wide range of other precautionary and health measures, including the prevention of industrial accidents due to lowered concentration and increased mental stress, daily health management, and elder care support.
The Poincaré plot analysis can calculate the nonlinear analysis indexes SD, SD1, and SD2 through a scatter diagram of the adjoining RRI intervals. The area of the ellipse valued at the SD (total HRV) correlates with baroreflex sensitivity (BRS), LF, HF, and RMSSD. The standard deviation in the short axial direction of the ellipse is referred to as SD1. This index demonstrates the short-term HRV of the rapid RRI change and is the same as the RMSSD [16]. SD1 is related to the coronary vagus nerve function. Reports have shown that it may be used to indicate exercise intensity (endurance drop) [17]. In contrast, SD2 demonstrates long-term HRV and is considered to be correlated with LF and BRS [18]. In this study, the RRI fluctuation was extreme, as it was used as verification in the exercise load test; compared to the amount of change in SD1, the SD2 index more noticeably captured RRI fluctuations. However, SD1 is still meaningful in cases where there is little overall change in RRI (e.g., desk work or light labor) or if the heart rate recovers in a short period of time. Although reports have shown that the ratio of SD1 to SD2 is useful as an index of sympathetic nerve activity [12], both indexes should be assessed based on physiological findings.
Both the HRV index used in the frequency analysis and that according to the RRI frequency analysis were composed of an LF component (low frequency component 0.04–0.15 Hz band) and an HF component (high frequency component 0.15–0.4 Hz band); it is thought that the LF component is mainly composed of Mayer wave-related sinus arrhythmia, which is derived from blood pressure, while the HF component is composed of respiratory sinus arrhythmia, which is derived from respiratory activity. The HF component decreases in conjunction with the inhibition of parasympathetic nerve activity due to autonomic nerve disturbances and mental load. In this context, the LF/HF ratio to the LF component is considered to be an index of sympathetic nerve activity [9][19]. Castrillón et al. [20] have examined this index as an indicator of post-exercise recovery. It is important to note that most of the previous studies on this index have been validated by physiological measurements taken during desk work or in the context of standing up from the supine position, not during exercise, as in this study [19][21]. There also exist some negative findings regarding this index [22], so there is room for debate about the measurement conditions and other issues. Although these indexes capture gentle fluctuations between 0.04–0.4 Hz, they are not useful for the real-time detection of status during exercise loads. In this study, we were able to verify that both the LF and HF components disappeared during exercise loads; during those in which RRI becomes noticeably short with rapid respiration, we also believe that heart rate fluctuations shift to components other than the frequency bands thereby defined. While the well-known HRV index according to frequency analysis is now popular in simple stress measurement applications, it is crucial to ensure proper handling based on appropriate mathematical/physiological findings.
Regarding the estimation model proposed in this study, it is necessary to avoid actual measurements that tend to be higher than estimated values when targeting heatstroke prevention. Although the allowable range for the core body temperature estimation error is difficult to determine, a limit on the order of 0.5°C is probable if the properties of core body temperature are considered. Focusing on S09-B (Supplementary Fig. 1(j)), wherein the lowest core body temperature was estimated relative to the actual measurement, the participant in question was a very strong athlete, with an exercise intensity of 157 W (on par with hard labor) and maximum oxygen intake of 60%; this participant was obviously accustomed to exercise, as was also indicated through the interview survey. During exercise, the heart rate rises to supply oxygen to the body; however, physical training increases cardiopulmonary function, thus allowing the same exercise to be performed at a lower heart rate. In addition, as post-exercise heart-rate recovery (HRR) is correlated with the physical activity Baecke score, HRR is reportedly a useful index for exercise habits [23]. Although a previous study among athletes found a correlation between post-one-minute HRR and age [24], it is possible to use a post-three-minute recovery index to evaluate exercise adaptability. Considering daily exercise habits, S09 appears to have experienced a different heart rate variation and recovery trend than other participants; his Poincaré plot index also exhibited markedly gentle RRI changes. As such, the difference in this heart rate response trend is thought to have caused the estimation error.
This study also had some limitations. For one, the proposed estimation model was targeted at core body temperature increases during exercise loads in hot environments (35°C, 50%), in which case various environmental conditions remain unverified. Caution should therefore be taken when interpreting the results. Two, participants wore shorts when biometrics were assessed during the experiment. However, previous investigations using the HSDA model have implemented five types of protective clothing when engaging in treadmill exercises [25]. As the use of clothing substantially increases core body temperature, future studies should therefore verify the proposed model under different conditions (e.g., different clothing types and room temperatures).
Regarding ECG signals, there are many other indexes apart from those investigated in this study [26], including those that demonstrate vagus nerve activity in time regions such as pNN50 [27] and the regularity (complexity) of time-series data (e.g., approximate entropy and sample entropy) via nonlinear analysis [28][29][30]. Although these indexes do not necessarily capture all the different types of phenomena, it is still possible to identify relationships and differences by contemplating indexes from a mathematical perspective. In other words, it is possible to achieve highly accurate estimation results through a diversified approach that decomposes the same biodata into multiple indexes, which may then be compared based on their unique differences and characteristics. Of course, it is also possible to add other physiological responses to phenomena that cannot be captured by ECG signals. Thus, this study should be considered an initial step, in which case we plan to conduct additional examinations.