Analysis results from thermal sequence images of the palmar area
As the heat from the skin was transferred and the residual heat on the plate evaporated, the pixel value gradually decreased in the thermal image. When sweating is caused by hyperhidrosis, the skin temperature of the affected area is lowered owing to the evaporation of sweat, and the lower the body temperature, the more rapidly the image changes. Figure 1 compares hand data on both sides (right: Rt, left: Lt) for the abnormal group (Ab) and the normal group (Nor). Analysis using the entire hand area is called whole analysis (WA), and analysis using the local area is called local analysis (LA). Even on the palm, the area and amount of sweating may differ from person to person. Therefore, two methods, carried out on the same person, were analyzed and compared to perform a quantitative evaluation by confirming the location. The LA was analyzed locally by dividing the hand into eight areas, which were the five fingers and three sections of the palm. The results showed data from a central region, including a concave portion of the palm (P), which is the section of the palm that sweats the most, and a second finger (F) region, which is the finger used the most (Figure 1). The x-axis in Figure 1A represents the index of the thermal image frame. The y-axis shows the change in intensity for each frame, which is the value obtained by subtracting the average intensity value of the first frame from the corresponding frame. The WA showed a substantial change in intensity, whereas LA showed a minor change.
Most importantly, in Nor, a range stands out in comparison to Ab. In Figure 1, the blue and the red areas represent Nor and Ab, respectively. The LA in Figure 1B was set identically to the WA in Figure 1A. For thermal imaging, the hands were placed on the plate for 10 s, and then removed. In other words, the imaging starts when the temperature of the hand is sufficiently transferred and immediately after the hand is removed. And, as seen in the figure, from the first frame of the image to before the frame of the box area, the temperature of the hand transferred to the plate evaporates quickly. That is, because the heat transferred to the plate is high, it has a high-intensity value in the grayscale image. As there is no longer a continuous heat transfer, thermal evaporation occurs, and the temperature rapidly decreases and the intensity decreases. The lower the ambient temperature, the greater is the reduction. At the same time, the more sweat there is, the more likely it is that there is water on the plate. This also contributes to the rapid change in evaporation. The evaporation process occurs continuously in which the surface temperature of the water decreases slowly until thermal equilibrium with the surrounding environment is achieved. Subsequently, from the frame corresponding to the box areas, it can be observed that the intensity increases as diffusion occurs and then decreases again. This demonstrates a collapsing linearity, which gradually decreases in value because the continuity between the water and vaporization planes is impaired by diffusion as the water content decreases. The more water there is, the higher is the surface temperature and the slower is the evaporation rate owing to slow diffusion. In the case of Nor with little sweat, a large diffusion was confirmed in the blue box area, and in the case of Ab with excessive sweat, diffusion was confirmed in the red box area (Figure 1A). And Ab was identified in later frames than Nor. The same phenomenon can be observed in all graphs of Rt. When the hand is humid or wet, and sweat is present, evaporation is thought to be limited by slow diffusion. Therefore, in the case of Ab, diffusion can be confirmed from a subsequent thermal image frame. In particular, in the data of Lt in Figure 1A, it can be seen that there were several diffusion sections (blue box areas), which were prominently displayed in Nor and gradually appeared longer. On the other hand, in the case of Ab data of Lt, it can be said that maintaining a constant intensity change from any frame achieves thermal equilibrium with the surrounding temperature. Because it was difficult to confirm the pronounced diffusion phenomenon, the intensity was directly approached and confirmed for each frame, and as a result, it was close to zero. This means that there is little change in the intensity because the transferred heat moves rapidly and the temperature becomes equal to the surrounding temperature.
Figure 1B confirms the intensity difference between the thermal image frames, and the y-axis represents the value obtained by subtracting the average intensity value of the corresponding frame from the previous frame. As the residual heat and sweat evaporate, the amplitude decreases. However, owing to slow diffusion, it is possible to see a section in which the amplitude of the previous frame changes significantly. Similarly, as in the box areas shown, in Nor, the value fluctuates considerably for a particular frame owing to diffusion. In the case of Ab with sweat, the more water it has, the slower the diffusion appears, and it can be seen that the data has a relatively low amplitude. Both groups showed rapid changes in the initial stage; however, in particular, the Ab group had a relatively low body temperature because of sweating. Therefore, the initial range of change is considered to be relatively large. Data analysis of various cases is necessary to generalize this phenomenon.
LA, the result of the local analysis, includes results for P, which is the mid-palm area, and F, which is part of the second finger. This can also be interpreted in the same manner as described above. During evaporation, diffusion was confirmed in the frame corresponding to the box areas, and in the case of Ab, diffusion was confirmed in a later frame. A thermal equilibrium state with almost no change in intensity was also confirmed. Interestingly, in the result of the F region for LA (Rt) in Figure 1, the Ab data appeared to have achieved almost thermal equilibrium because the intensity was almost zero within the thermal image frame; however, it was confirmed that it had a non-zero intensity value from the 42nd frame onwards. Thermal energy overlaps at certain boundary points while slowly evaporating and diffusing slightly from the central area where most heat is conducted to the edge of the finger. In addition, it can be seen that the drastic change due to evaporation at the beginning is relatively smaller in the F data than that in the P data. Because the finger is a peripheral body part, it cools faster than the palm, and this data confirms that it has a lower body temperature than it.
Figure S2 shows the data of the other subject for the results shown in Figure 1. Initially, it shows a drastic change due to evaporation, followed by a diffusion section from the frame corresponding to the box areas. In graph A, the average intensity change increases and then decreases; in graph B, for the same section, the amplitude increases more sharply than that in the previous frame. It was confirmed that Ab appeared at a lower rate than Nor.
Hyperhidrosis is often accompanied by cold hands and feet because the temperature of local areas, such as the hands, becomes cold as a result of sweating. Rheumatoid arthritis (RA) may be the underlying cause of cold hands and feet. When diagnosing the thermography of RA, the temperature difference between the fingertips (second, third, and fourth) and the palm area is obtained and compared for both sides9. In the case of the palm, the central region is checked; however, owing to the characteristics of our experiment, analysis is performed from the transferred heat based on the thermal image. Therefore, the average intensity of the four palm regions and the intensity difference between the three fingertips were obtained. The corresponding areas when handling the WA data are shown in Figure S1A. Figure 2 shows the temperature difference between the fingers and palms on both sides. In the case of Nor, the temperature difference was not large. Owing to evaporation, the temperature difference decreased. On the other hand, Ab had a large temperature difference between the fingers and palms; therefore, a rapid temperature decrease was observed in the beginning, indicating that the temperature of the hand was low at the time of imaging.
The results obtained from the subjects of the Figure S2 data can be seen in Figure S3. Nor did not show a significant difference because the temperature difference between the fingers and the palm was similar; however, in the case of the fourth finger on both sides, a large temperature difference was initially observed. Ab also showed a large initial change, indicating that the temperature of the finger was lower than that of the palm, more than Nor.
Figure 3 shows the intensity changes extracted from the second finger and mid-palm regions, which are representatively used when evaluating hyperhidrosis using the LA method. LA uses each mask image for a palm region separated into five fingers and three palm sections. Three intensity values were extracted from each mask image to confirm these changes (Figure S1B). Figure 3 also confirms the intensity change, which decreases owing to evaporation in both regions. The intensities of the three spots show evaporation at different values and different rates. Furthermore, it can be seen that the Ab spots show a sharp change at the beginning compared with those of Nor. The blue and red box areas indicate the same frame as in Figure 1, where prominent diffusion phenomena in Nor and Ab are shown. If the value does not change after a certain time, the image intensity is zero, which is considered to be in thermal equilibrium with the ambient temperature. However, in Ab of Figure 3A, the intensity of three spots gradually increased from the 41st–43rd thermal image frames, which are considered to be affected by the diffusion of the surrounding area.
Similarly, the results for the other study subject are shown in Figure S4. The blue and red box areas in this figure indicate the same frame as in Figure S2, where prominent diffusion phenomena in Nor and Ab are shown. In each dataset, it can be seen that the intensity changes because of evaporation, and diffusion occurs at a slow rate. In the case of Ab, a sudden change in the initial stage can be observed owing to water. In particular, the constant section, which can be confirmed in Ab of B, is in a state in which the surface temperature is in thermal equilibrium with the ambient temperature. In fact, the intensity is zero from the 5th, 7th, and 11th thermal image frames for each spot.
Analysis of the questionnaire results of subjects
Two types of questionnaires were filled out during the experiment. One was the existing questionnaire used to diagnose hyperhidrosis (hyperhidrosis evaluation questionnaire, HDSS), and the other was a questionnaire (palmar hyperhidrosis condition questionnaire, PHCQ) developed by combining the existing hyperhidrosis questionnaire items with items most associated with hyperhidrosis symptoms. Each set consisted of nine and six items, respectively. The purpose of the PHCQ was to examine the condition of the subjects' hands and ascertain whether sweat, which was the experimental condition of this study, was well-formed. It was further determined that the results of this experiment would be considered reliable if a clear state change was observed. Therefore, corrected item-total correlation coefficients (CITC) were calculated to evaluate whether each questionnaire had internal consistency, that is, for which conditions it was asked consistently.
The CITC results for HDSS are shown in Table 1. There are a total of nine items (H1 to H9); however, H5, H6, and H9 have zero variation, that is, all subjects surveyed had the same score. Therefore, the calculation was omitted and calculated as the result of type A. Usually, a correlation value of 0.7–0.8 (item-adjusted total correlation) indicates an acceptable Cronbach’s alpha, and a value lower than this is an unreliable scale. Therefore, re-evaluation was performed by removing items with low correlation to identify and evaluate the internally matching questionnaire items. Type B was the result of re-evaluation by removing H2 and H8, which showed the lowest correlation in Type A, with 0.0232 and 0.0719, respectively. The total Cronbach's alpha for Type A also increased from 0.6303 to 0.7692. Type B was also re-evaluated by removing H7, which showed the lowest correlation with 0.3314, and the result was Type C. Finally, the total Cronbach's alpha value improved to 0.8936, and it was confirmed that the subjects of this study had particularly good internal consistency for H1, H3, and H4 items.
The CITC results for the PHCQ are shown in Table 2. There are a total of six questionnaire items. The left and right hands were investigated (DL1–6, DR1–6, WL1–6, and WR1–6) for the usual hand condition (dry, D) and the altered hand condition (wet, W). This questionnaire was designed with reference to the existing questionnaire for diagnosing hyperhidrosis to verify whether the experimental factors considered in this study were well-formed and the results analyzed from such data were reliable. In both the D and W cases, 0.8210 and 0.8540 were found to have good internal consistency. Therefore, it can be confirmed that all items are worthy of retention, which means that the results of this study, conducted for an objective evaluation of hyperhidrosis, can be trusted.
Classification of hyperhidrosis
In this study, the presence or absence of sweat, which is a symptom of hyperhidrosis, was binary classified, and the experiment was conducted assuming that Nor was the normal group and Ab was the abnormal group. Among the six machine learning classifiers used for classification (ECOC-SVM, decision tree, discriminant analysis, K-nearest neighbors, random forest, and AdaBoost), the result of K-nearest neighbors showed the best performance with an accuracy of approximately 98%, and the random forest results showed the second-best performance with an accuracy of approximately 91% (Table 3). The Matthews correlation coefficient (MCC) suggests that the closer it is to 1, the better is the prediction of both classes, even if the data are unbalanced. For all metrics, the K-nearest neighbors showed the best results. K-nearest neighbors are considered to have shown excellent results because classification is performed using only the data. In the case of random forest, which showed suboptimal performance, a bagging method that generates multiple data samples through random sampling and trains them differently showed sufficiently easy and intuitive prediction performance. However, further improvement of performance by strengthening randomness was limited because similar or a small number of model types were trained. In the case of other models, such as ECOC-SVM, decision tree, discriminant analysis, and AdaBoost, performance can be improved if more diverse feature data are trained. Figure 4 shows the receiver operating characteristic (ROC) curve for performing cross-validation on the trained classifier to evaluate the classifier performance equally. The area under the curve (AUC) for all classifiers were above 0.8, indicating good classification performance.
Hyperhidrosis evaluation system
In this study, an application system and protocol capable of objective analysis were established for the accurate diagnosis of palmar hyperhidrosis (Figure S5). When basic personal information of a patient, such as name, date of birth, and gender, is input into this system, a coded folder is automatically created, and accessibility is improved by including a function to automatically collect thermal images through the corresponding path. In addition, it is possible to check the graph results analyzed from the collected thermal images and the history of accumulated patient data simultaneously; thus, a comparative analysis is possible, and the prediction result of the presence or absence of hyperhidrosis can be confirmed.