Facial skin temperature in acute stroke patients with delirium - A pilot study

Facial skin temperature depends strongly on blood flow in small blood vessels in the skin. These are regulated by the sympathetic part of the autonomic nervous system. Delirium may pathophysiologically be associated to changes in the sympathetic part of the autonomic nervous system. In this observational study, we evaluated the influence of various exogenous and endogenous covariables on the regional facial temperatures in acute stroke patients with and without delirium. Facial thermography (FT) was performed using an infrared digital camera. Screening for delirium was done using the Confusion Assessment Method (CAM). Sixty-four patients were enrolled. Eight patients developed delirium. Sex and body temperature were positively associated to facial skin temperature, and so was ambient temperature but to an overall lesser magnitude. Stroke severity, diabetes, infection, facial palsy, facial sensory deficit, and physical activity did not influence facial skin temperature. Overall, there was no association between facial temperature and the occurrence of delirium except in one facial region, the medial palpebral commissure.


Introduction
Skin temperature is strongly influenced by blood flow in small blood vessels of the skin which is regulated by the sympathetic part of the autonomic nervous system [1]. The small blood vessels consist of arterial and venous vasculature and constrict in response to an increased sympathetic stimulation [2]. Somatic stress therefore results in a decreased skin blood flow and lower skin temperature. There is a bilateral skin temperature symmetry in the face measured at various regions [3][4][5][6][7][8][9][10][11][12][13][14][15][16], which indicates a strong central regulation of the facial skin temperature.
Acute stroke is often complicated by delirium [17] which pathophysiologically may be associated with changes in the sympathetic part of the autonomic nervous system [18][19][20][21][22]. We therefore hypothesized that changes in skin temperature may be linked to development of delirium in patients with acute stroke.
Facial thermography (FT) is a method which can measure skin temperature in various regions of the face by image analysis of an image of the face obtained with a digital infrared camera. It has been used in various clinical research settings, e.g. orofacial pain, hypoglycemia, and herpes zoster [15,23,24]. The primary objective of this study was to investigate the feasibility of FT in a population of acute stroke patients with and without delirium. A secondary aim was to evaluate the potential use of FT as a marker of delirium in the acute phase of stroke. We measured facial temperature with FT in different regions of the face in acute stroke patients with and without delirium and evaluated the influence of various exogenous and endogenous covariables on the regional facial temperatures.
This article is the second article from a study in which we measured sympathetic autonomic activity using three different measurement modalities. The findings from the two other modalities (skin conductance level and pupillary dilation velocity after a light stimulus) are presented in a separate article [41]. We chose to present the findings separately due to the nature of the measurements and since one unified report would leave out important points in order not to be excessively long.

Methods
The study, including inclusion and exclusion criteria and descriptive variables, has been described in details elsewhere [41]. In brief summary, the study prospectively enrolled adult patients with acute ischemic stroke admitted to the acute stroke unit at the Department of Neurology, Nordsjaellands Hospital, Denmark. The study enrolled patients during a period of 14 months in 2018 and 2019.
Facial temperature in regions of interest (ROIs) were measured using a digital infrared thermographic camera (T430sc™ from FLIR® Systems, Inc.) photographing the face in a frontal view (Fig. 1). The distance from camera to face varied between subjects since it was adjusted to make the face fill the whole photographic frame. The images were taken by three observers (JS, AA and LB) depending on observer availability. We measured seven ROIs in each side of the face using locations defined by Haddad et al. [7] (Fig. 1): Supratrochlear (ST), Temporal (TEMP), Lateral palpebral commissure (LPC), Medial palpebral commissure (MPC), Nasolabial (NL), Labial commissure (LC) and Inferior labial (IL). All our ROIs were a 9 pixels large square. The aim was to perform FT of each enrolled patient in both the morning and the afternoon every day during their stay at the acute stroke ward. The ambient temperature was recorded with a digital thermometer (Electrolux® E4RTDR01) at the same time as FT was performed. There was no air conditioning capable of keeping the ambient temperature within a narrow and specified range. Data on the body temperatures were collected from patient charts and were based on routine measurements made by the nursing staff using ear thermometers. We only used measurements made within a maximum of 3 h from the FT measurement.
The thermographic images were analyzed using the software FLIR ResearchIR Max, version 4.40.3.16 © 2007-2016 -FLIR Systems. The ROIs were delineated using the software's built-in marker and the temperature within the specific ROI was then calculated by the software. Two researchers measured 254 images independently of each other (JS and SU) in order to make an interobserver analysis. For the rest of the analyses, only the temperatures extracted by JS were used.
The presence or absence of delirium was assessed with the Confusion Assessment Method (CAM) [25] once per day in weekdays. A CAM assessment was based on observations by nurses, therapists, and physicians in this period and observations by JS from that day. A patient with a positive CAM assessment was considered to have delirium for the entire period until the next CAM assessment. A negative CAM assessment was handled similarly. Stroke severity was assessed with the National Institutes of Health Stroke Scale (NIHSS) [26,27]. Modified Rankin Scale (mRS) was used to evaluate disability at discharge [28].
The stroke subtype was classified using the Oxfordshire Community Stroke Project Classification [29] and the TOAST classification [30]. Data regarding the stroke deficits facials palsy and facial sensory deficits, as well as the side of lateralization of all stroke symptoms were also collected. Physical activity was recorded in activity counts (AC) with accelerometers (Actical Z from Philips Respironics) placed on both arms and legs.

Statistics
Comparisons of patients with and without delirium were made for all descriptive variables using either Welch's t-test, Wilcoxon rank sum test or Fisher's exact test, as appropriate. Facial palsy and facial sensory deficits were included as variables since they might affect measurements of the facial temperature. Aphasia was included since it might affect the CAM assessments. The possible influence of the facial motor and sensory deficits was analyzed using linear mixed models for each ROI per side of the face. The respective ROIs were set as outcome and the respective symptom as a fixed effect. The intercept for each subject and the measurements occasion were random effects. Similar models were made with the Oxford categories of stroke instead of the facial deficits.
We constructed linear mixed models to analyze the association between the temperature in each ROI and a given covariable at the time when a given thermographic image was taken. Continuous covariables were age, NIHSS, physical activity, ambient temperature, and body temperature. Categorical covariables were sex, diabetes mellitus, the use of beta-blockers, and possible infection (clinical judgement of treating physician). The ROI temperature was set as outcome and the covariable was set as a fixed effect. The intercept for each patient, the measurement occasion and the side were set as random effects.
A model not containing any of the covariables but only delirium was also made, i.e., whether a patient had a positive or a negative CAM score in relation to a specific FT measurement occasion. This model was supplemented with a series of models also containing the covariables as fixed effects.
Linear mixed models were made for each ROI in the interobserver analyses. ROI temperature was set as outcome. Observer was set as a fixed effect. Random effects were the intercept for each subject, measurement occasion and side. Measurement occasion was a construct of days since enrolment and the time of the day (i.e., morning or afternoon).

Results
Sixty-four patients were studied. An overview of the population is shown in Table 1. This table is also presented in the other article from this study [41. The population is grouped into patients with delirium and patients without delirium. Only eight patients developed delirium during the observation period. Table 2 shows the results from the mixed models analyzing the influence of the covariables on facial temperature measurements. There was no effect of age on the ROI temperatures. Although both the models with NL and LC had P values just below 0.05, their 95% CIs included 0 when rounded to two decimals. Sex was associated with skin temperatures in some ROIs (MPC, NL, and IL), and in all three instances male sex was associated with a higher temperature. In some of the ROIs (TEMP, ST, LPC, and IL), the skin temperature was significantly associated with ambient temperature whereas this was not the case in the other ROIs. Ambient temperatures were in the range of 20.7 • C to 28.9 • C (mean = 23.6 • C, SD = 1.4 • C). There was a significant positive association with body temperature in all ROIs except for IL. Body temperatures were in the range of 35.9 • C to 37.7 • C (mean = 36.7 • C, SD = 0.4 • C).

Effect of covariables on facial temperature measurements
None of the other covariables had statistically significant effects on the ROI temperatures, except that use of beta-blockers was associated with a slightly lower skin temperature in a single ROI (IL) compared to non-users of beta-blockers (difference of − 0.79 • C, 95% CI (− 1.39 to − 0.2), P value = 0.10).
The ROI skin temperatures were not influenced by the stroke symptoms facial paresis or facial sensory deficits. The Oxford categories showed no distributional difference with regard to the side of the stroke symptoms. In a few of the ROIs statistical significance was reached, but no systematic influence of the stroke categories on the ROI temperatures was evident. Results are available in Supplemental Table 1 and   Supplemental Table 2.

Missing data
A total of 274 infrared images were taken giving a total of 548 potential readings per ROI because each ROI is located in both the right and the left side of the face. The maximum number of infrared images that could have been taken was 321. However, in 47 instances a image was not taken either due to absence of patient or researcher. Not all ROIs could be read in all images, see Table 3. Only LPC and MPC could be read from all images.
The models with NIHSS, activity, ambient temperature, or body temperature did not utilize all the thermographic measurements. One patient was not NIHSS scored. Nine patients did not have accelerometer recordings. A corresponding ambient temperature measurement was missing for 20 infrared images. A total of 115 images lacked a corresponding measurement of body temperature.

Interobserver analysis
In Table 4 we show the results from the interobserver analyses. There was a small but significant effect of the observer in the ROIs TEMP, LPC, MPC and IL. Only 254 of the total 274 images were available for interobserver analysis.

Facial temperature during delirium
The results of the models analyzing the association between temperature in each ROI are shown in Table 5. There was no difference between facial temperature measurements while delirium was present and measurements obtained while delirium was absent except in MPC where skin temperature was significantly lower during delirium (difference of − 0.40 • C, 95% CI (− 0.72 to − 0.08), P value = 0.014).
These models were supplemented with models containing both delirium and each of the covariables in turn. These models yielded   overall the same results for the covariables as in the models not containing delirium as a fixed effect. The presence of delirium remained significantly associated with the skin temperature in MPC with effect sizes and CIs of similar magnitude, no matter the covariable. Skin temperature in ST became significantly associated to the presence of delirium in the model with ambient temperature, but there was no association between delirium and temperature in ST in all other models. The results from the supplemental models are available in Supplemental Table 3 and aggregate FT data in relation to delirium can be seen in Supplemental Table 4.

Discussion
The facial skin blood flow is regulated by the sympathetic autonomic nervous system and increased sympathetic stimulation decreases the skin blood flow [1,2]. The blood flow through the facial skin may very well be the major contributor to the skin temperature [31]. This makes decreased skin temperature a potential marker of increased somatic stress in the non-febrile patient. Skin temperature may also be affected by other factors. In our study, we incorporated various endogenous and exogenous covariables into linear mixed models in order to evaluate their influence on the regional facial temperature ( Table 2). The covariables were chosen because they either theoretically could influence skin temperature or were significantly different between patients who developed delirium and those who did not (i.e., age and NIHSS). Sex, body temperature and ambient temperature had a significant influence on several ROIs.
Male sex was associated with an increased skin temperature in three ROIs (MPC, NL, and IL). Males have previously been demonstrated to have higher cheek skin temperatures than females [5]. This difference may theoretically be caused by the generally larger body mass of males compared to females or to a higher metabolic rate in men. We have not found any previous studies examining this explicitly. Our results nevertheless indicate that a sex effect should be considered in future facial thermography studies.
Body temperature was positively associated with facial temperature in all but one ROI (IL). The range of skin temperature increase in these regions was 0.29 to 0.65 • C per one-degree Celsius increase in body temperature. These results indicate a consistent association between facial temperature and body temperature which hence also should be taken into account when performing studies using FT. To our knowledge no previous studies have examined the relationship between facial skin temperature and body temperature in non-febrile patients in a clinical setting. A relationship between facial skin temperature and body temperature has been demonstrated in healthy volunteers in e.g., an exercise setting [32], but such a relation was not found in a non-exercise setting [5].
The influence of ambient temperature was less consistent, and a significant association was only found in four of the seven ROIs. Although we have not found any studies documenting a direct relation between ambient temperature and facial temperature, we decided to temporarily stop data collection during an exceptionally hot summer period where room temperature often exceeded 29 • C assuming that high ambient temperatures would impact the reliability of the measurements. There seems to be a tendency in thermographic studies in general to keep the ambient temperature within relative narrow intervals [33,34], potentially diminishing the effect of ambient temperature on skin temperature. A previous study found no association between ambient temperature and cheek skin temperature, but the   ambient temperature range was not mentioned [5]. To the best of our knowledge our study is the first to study facial skin temperatures in an ambient temperature span as wide as 20.7 • C to 28.9 • C (mean = 23.6 • C, SD = 1.4 • C).

Interobserver analysis
The interobserver analysis showed small but significant systematic differences between the two observers regarding the ROIs TEMP, LPC, MPC and IL. Although the observers had previously trained together in correctly identifying the ROIs, the location of some ROIs may be difficult to identify and mark up with the software. The ROIs TEMP and IL both lie in facial regions without clear anatomical landmarks that can be used as points of reference making small differences between observers' identification of these two ROIs more likely. The two palpebral ROIs LPC and MPC are, in contrast, located in direct vicinity to clear anatomical landmarks, i.e., respectively the lateral and the medial palpebral commissures. However, the temperature gradients in the immediate vicinity of both of these regions may be relatively steep and thus small differences in the placement of the software ROI marker may explain the interobserver differences implying a limitation in their use. However, the distance from zero was not greater than 0.34 • C for any of the confidence intervals and the size of the differences for TEMP, LPC and MPC seems negligible and not clinically relevant. MPC had an interobserver difference of − 0.07 • C, 95% CI (− 0.14 to − 0.004), which is magnitudes less that the difference between measurements with delirium present and delirium absent, difference of − 0.40 • C, 95% CI (− 0.72 to − 0.08) in the unadjusted model for MPC. There are no other studies reporting on interobserver analyses in these specific ROIs. A previous study found a small interobserver effect for a ROI involving the major part of the cheek [5] and another study also reported a small interobserver effect for the forehead and cheek combined [16]. We did not make any intraobserver analyses but we acknowledge the possibility that significant intraobserver differences exist and may confound the interobserver analyses. To the best of our knowledge, no other studies have examined a potential intraobserver effect for digital infrared facial thermography in these ROIs.

Facial temperature during delirium
We also examined the possible association between alterations in facial skin temperature and delirium in the acute phase of ischemic stroke. We hypothesized that since the skin is under autonomic control and delirium is associated with autonomic dysregulation, the facial skin temperature might reflect the occurrence of delirium.
We found a general lack of association between skin temperature in a given ROI and delirium analyzed using the linear mixed models except for the skin temperature in the region of the medial palpebral commissure and possibly the supratrochlear region. However, skin temperature in the ST region was only statistically significant in the model where ambient temperature was set as a covariable, not in any of the other models with ST, making the finding of a true and clinically meaningful association very doubtful. There may very well be no biological association between the ST skin temperature and delirium.
The association of skin temperature in the ROI MPC and delirium seems more robust. Significant associations were found in an unadjusted model with delirium as the sole fixed effect and in several adjusted models with relevant covariables added in turn as a fixed effect in conjunction with delirium. It is of note that the effect sizes and CIs for delirium were of similar magnitude in all models.
MPC is in a region of particular thin skin and this may explain an association between delirium and the MPC temperature. We speculate that the thinner skin allows the effect of an increased somatic stress response (i.e., an increased sympathetic activity) to be detectable by infrared thermography in this region since somatic stress causes the small blood vessels of the skin to constrict resulting in decreased skin blood flow and lower skin temperature. We are aware that blood flow is not the only factor contributing to skin temperature and that the magnitude of the contribution of blood flow is unclear. On one hand, a previous work showed that skin blood flow is the major determinant of skin temperature [31], while another study has shown that the skin blood flow is not the major determinant of skin temperature and reported that the thermal conductivity of the skin and underlying tissue contribute to a larger degree than skin blood flow [35].
Regardless of these considerations our results suggest that routine infrared temperature measurement in the MPC may be useful in a clinical setting for warning of a delirium under development, but this needs to be explored in further studies. A single measurement of MPC will probably not be useful to detect delirium, but serial measurements could reveal a change in the MPC temperature indicative of delirium. This would be in line with the understanding that delirium is associated with a somatic stress state [18][19][20][21][22].
Based on the results in this pilot-study, we have been able to calculate an estimate of a required sample size needed to convincingly demonstrate changes in the skin temperatures of the MPC ROI in association to delirium. We used a power calculation method described elsewhere [36]. Alpha level was set to 0.05, the power level set to 0.8 and delirium incidence set to 12.5%. We set the effect size to be 0.4 • C and the variance in the MPC skin temperatures to be 0.59 • C 2 (the effect size for delirium and the cumulative variance from the random effects in our model for MPC with delirium as a single fixed effect and the intercept for each patient, the measurement occasion, and the side of the face as random effects). A sample size of 265 patients would be needed to convincingly dismiss or demonstrate that the presence of delirium is associated to an average decrease of 0.4 • C in skin temperature in the MPC facial region.

Limitations
Two important weaknesses of this study are the small number of patients and missing observations. Only 64 patients were studied and only 8 developed delirium. Our study should therefore be viewed as a pilot-study. A larger study is needed to confirm our results.
In all of the models, a maximum number of two fixed effects (e.g., delirium and age) and three random effects were employed. The changes in skin temperature in each ROI is analyzed in multiple similar models where covariates are used in turn. This approach does not carry the same risk of overfitting any particular model as an approach where all covariates are incorporated into a single adjusted model for each ROI.
The ROIs TEMP, ST, LC, and IL have a high number of missing observations compared to the other ROIs. These ROIs were difficult to measure reliably because hair and beard covered the skin. If hair or beard was present in a ROI on a given photo, we omitted to make a reading of the skin temperature in that ROI explaining the missing observations. There may be a potential confounding effect as only males had beards and therefore more frequently had missing observations from the three ROIs in the lower face, NL, LC, and IL. In both NL and IL males had significantly higher skin temperatures than females. It is not possible to dismiss or to confirm a confounding effect.
The thermal images were done by three observers depending on observer availability. The distance from patient to camera was not equidistant in regard to metric length, instead each image was taken in a way trying to make the face fill up the whole frame. This of course introduces an unknown amount of variance which may have a systematic dependence on the observer as a factor alone (i.e., an interobserver effect) and on the individual observers' ability to consistently align the camera so that the face would take up the whole frame (i.e., an intraobserver effect).
The approach of trying to have the face take up the whole frame instead of placing the patients equidistant in absolute length (e.g., always with a distance of 1 m from face to camera) of course means that the absolute skin area for each ROI had small variations across images.
This introduces an unknown amount of variance across all ROIs of the same type. It is unknown whether this caused any systematic effect.
It is also a limitation of the study that as many as 115 infrared images lacked a corresponding measurement of body temperature. We planned to use the routine body temperature measurements done by nurses in the ward and discovered too late that many of these routine measurements were not done at the planned time. We did not use measurements of body temperature more than 3 h apart from the infrared images and since many routine body temperature measurements scheduled to be done in the afternoon were actually done later in the evening many body temperature measurements were not usable. The body temperature measurements were also done by several nurses and with several thermometers, introducing and unknown amount of variance in the body temperature data.
Activity data was missing from 9 patients as they had not worn the accelerometers. It cannot be refuted that these missing data may have introduced a bias into our results.
In this study we used the CAM to assess for delirium. We chose to define delirium as a positive CAM. This of course introduces a limitation as the CAM has not been validated in a stroke setting [37,38]. The presence of aphasia can impair assessments of disorganized thinking or inattention, two items in the CAM. We did not exclude patients with aphasia. However, patients with aphasia, or other neurological deficits impairing communication, so severe that they could not give informed consent to participate in the study were not eligible for the study due to lack of consent. This means that patients with communication deficits, who were still able to communicate meaningfully, were eligible for the study. How this affected the CAM assessments is unknown.
It should be appreciated that the delirium incidence of 12.5% in our study is relatively low compared to delirium incidences found in stroke populations in three systematic reviews [17,39,40]. The patients in our study had relatively minor strokes based on their NIHSS scores which may in part explain why the delirium incidence was relatively low.
Another weakness of our study is that we did not adjust for multiple testing in the models. A simple Bonferroni correction is illustrative. If we set the number of tests to 8 (distinct covariable models for each ROI), the P value would only be significant if it is less than 0.0063, leaving only the ambient temperature adjusted model with MPC as significant regarding an association with delirium.

Conclusion
In summary, we measured facial temperature with infrared digital thermography in different regions of the face in acute stroke patients with and without delirium and evaluated the influence of various exogenous and endogenous covariables on the regional facial temperatures. Sex and body temperature both had significant association to facial skin temperature and should both be considered in future research using infrared facial thermography. We also found a relationship between skin temperature in some facial regions and the ambient temperature, but of an overall lesser magnitude than for body temperature.
Overall, there was no association between facial temperature and the occurrence of delirium except in one facial region, the medial palpebral commissure. The study has limitations and the association between the skin temperature in the medial palpebral region and delirium could be due to multiple testing on a low number of patients in the study. Further studies will be needed to explore whether FT is clinically feasible as a tool for early detection of sympathetic autonomic dysfunction in other categories of hospitalized patients with delirium.

Funders
Dr. Stokholm has worked on the study as a PhD-student funded by the following: