Body temperature (BT) is strictly regulated ^{1, 2, 3} and is a key vital sign. In critically ill patients, abnormal BT is associated with adverse clinical outcomes. ^{4, 5} Moreover, BT abnormalities are used to identify specific phenotypes of immune reaction to infection ⁶ and to evaluate response to interventions (eg, antibiotic or anti-inflammatory therapies). ⁷ Accordingly, prompt recognition and monitoring of BT perturbations is fundamental in the management of traumatic brain injury ⁸ and cardiac arrest ⁹ and also as an early warning sign of critically ill patients in general.

The pulmonary artery catheter is considered the gold standard for the measurement of core temperature, ¹⁰ but its use is limited and declining. ^{11, 12} Compared with pulmonary artery catheter, temperature-sensing intra-urinary bladder catheters, oesophageal and nasopharyngeal probes may represent less invasive methods of core BT measurement. However, their use is limited in critical care practice due to urine output dependence and easy displacement. Thus, non-invasive BT measurement methods (axillary thermometer, tympanic infrared and temporal scanner) have become common. ^{2, 13, 14, 15} But their accuracy in estimating core BT in ICU patients is uncertain. ^{16, 17} Moreover, a recent survey ¹⁸ highlighted a lack of agreement on the preferred BT measurement methods between doctors and nurses and lack of protocol governance among Australian and New Zealand ICUs.

Accordingly, we sought to investigate the accuracy of the axillary chemical dot, tympanic infrared and temporal scanner methods in estimating core temperature in critically ill patients admitted to the ICU. The primary hypothesis was that all three non-invasive BT measurement methods would show low accuracy in estimating core BT compared with invasive methods and that these accuracies would differ from each other. The secondary hypothesis was that accuracy of each non-invasive BT measurement method would be influenced by clinical variables.

The Ethics of Human Research Committee of the Austin Hospital approved the study and waived the need for informed consent (LNR/17/Austin/315) due to its implied observational nature.

We performed a prospective observational study in critically ill patients admitted to the ICU of two tertiary metropolitan hospitals (Austin Hospital and Warringal Hospital, Melbourne, VIC). We included patients who had a temperature-sensing intravascular (mostly pulmonary artery catheter) or intra-urinary bladder catheter, which was placed before the inclusion in the study by the treating physician. We excluded patients aged less than 18 years, patients with skin infection, and pregnant women

Characteristics (manufacturer and other commercial details) of the BT measurement methods used in the study are reported in the Online Appendix, eTable1. All BT measurement methods were calibrated according to hospital protocols and manufacturer’s specifications. We classified BT measurement methods as:

Patients’ demographic data (age, sex, height and weight), cardiovascular comorbidities and main pathology at ICU admission were recorded. Concurrent haemodynamic parameters (blood pressure and heart rate), medications (sedatives, opioids, vasopressors and inotropes), fluid balance, artificial organ supports (mechanical ventilation, renal replacement therapy and extracorporeal membrane oxygenation), and warming or cooling interventions were recorded at each BT measurement time point.

The primary outcome of the study was the assessment of accuracy of non-invasive BT measurement methods in estimating core temperature and the relationship between axillary chemical dot, tympanic infrared and temporal scanner methods by the evaluation of:

Specifically, clinical accuracy was defined as bias between each method and the others within ± 0.2°C. ¹⁶ Clinical adequacy was set at LoA of 1°C compared with core temperature.

A secondary outcome of the study was the evaluation of clinical conditions or interventions, which predicted bias and LoA between invasive and non-invasive BT measurements.

Continuous data are presented as mean (standard error; SE) or median (interquartile range; IQR), categorical data are summarised as number (percentage). After assessing the distribution of the data with the kernel density plot and the Kolmogorov–Smirnov test, the relationship between invasive and non-invasive methods was investigated with the Bland–Altman approach, by plotting differences versus means of paired core temperature (gold standard) and non-invasive (index methods) measurements. ^{21, 22, 23, 24, 25}

To account for repeated measurements of BT in the same individual and adjust for significant explanatory clinical variables, mixed linear modelling was performed and patients’ demographic characteristics, comorbidities, medications and extracorporeal organ supports were included in a multivariable model. Then, bias and LoA were adjusted for significant predictors of BT temperature. Finally, a multivariate analysis accounting for repeated measures was performed with the aim to explore the relationship between variables which were independently associated with BT and bias of non-invasive BT measurement methods in estimating core temperature. We analysed data using Stata/SE 13 (StataCorp, College Station, TX, USA) and SAS version 9.4 (SAS Institute Cary, NC, USA) while graphs were plotted using MS Excel 2017 (Microsoft, Seattle, WA). To increase the robustness of this study, a reduced two-sided P value of 0.01 was used to indicate statistical significance.

A convenience sample of 50 consecutive critically ill patients (n = 35, 70% male) was included in the study. Demographic and clinical characteristics of all patients are shown in Table 1. Median age was 68 years (IQR, 59–74 years) and median APACHE III score was 40 (IQR, 33–49). The most frequent reason for ICU admission was cardiac surgery (n = 44, 88%). All patients were sedated (N = 50, 100%), most were mechanically ventilated (n = 47, 94%) and 13 patients (26%) were receiving vasopressor/inotropic support at study inclusion.

Over 375 observations in 50 patients, core temperature measured by invasive BT measurement method ranged between 33.9°C and 39°C (Online Appendix, eTable 2). Non-invasive BT measurement methods systematically underestimated core temperature (Table 2). On average, the LoA between invasive and non-invasive BT measurements methods was about 3°C, with the temporal scanner methods showing the worst performance in estimating core temperature (bias, 0.66°C; LoA, −1.23°C, +2.55°C), followed by tympanic infrared (bias, 0.44°C; LoA, −1.73°C, +2.61°C) and the axillary chemical dot method (bias, 0.32°C; LoA, −1.64°C, +2.28°C). The axillary chemical dot and tympanic infrared had the lowest degree of bias and the largest LoA, which approximated 4°C. However, none of the non-invasive BT measurement methods showed clinically acceptable accuracy.

The regression model accounting for repeated measures of BT is presented in the Online Appendix, eTable 3. An independent and negative association was found between BT and propofol infusion, mechanical ventilation, neuromuscular blocker agents, renal replacement therapy, aortic surgery and active warming. After adjusting for significant covariates (Online Appendix, eTable 4), bias and LoA between invasive and non-invasive BT measurement methods were reduced (Online Appendix, eTable 5). However, none of the non-invasive BT measurement methods showed clinically acceptable accuracy in estimating core temperature. The corresponding Bland–Altman plots are represented in Figure 1, A–C. Moreover, we reported the deviation of each non-invasive BT measurements from invasive temperature (Online Appendix, eFigure 1, A–C).

The regression model exploring the relationship between explanatory variables of BT and bias between invasive and non-invasive BT measurement methods is presented in Table 3. None of the clinical variables included in the analysis showed significant association with bias either between core temperature and tympanic infrared or temporal scanner.

Axillary chemical dot failed to measure BT in 16 observations and was reported by nurses as non-readable. Tympanic infrared failed once in measuring BT (Online Appendix, eTable 2).

We assessed the accuracy of the axillary chemical dot, tympanic infrared and temporal scanner methods to estimate core temperature in ICU patients. Each of the non-invasive BT measurement methods showed wide LoA and inadequate clinical accuracy in estimating core temperature among an average of more than 300 paired observations. Moreover, each non-invasive BT measurement method showed low agreement with other non-invasive methods. Finally, adjusting for explanatory variables of BT such as propofol infusion, mechanical ventilation, neuromuscular blocker agents, renal replacement therapy, aortic surgery and active warming only improved their accuracy to a limited degree.

Previous international guidelines published in 2008 ²⁶ suggested avoidance of axillary methods, temporal scanner and chemical dot thermometers in ICU patients. However, such recommendations were only supported by observational studies ^{27, 28, 29, 30, 31} with controversial findings and methodological problems. Specifically, there was lack of control for possible confounders such as device calibration, investigator skill assessment, method used, failure to use the Bland–Altman approach in estimating bias and LoA, and lack of adjustment for multiple within-subject measurement. ³² Thus, these recommendations do not appear to have been widely adopted.

These shortcomings were further highlighted in an initial systematic review. ¹⁶ In a further systematic review of 75 studies in adult and paediatric patients from different clinical settings (only 45% ICU), investigators concluded that peripheral thermometers do not have clinically acceptable accuracy, although high quality data were limited and study heterogeneity was significant. In a more recent systematic review of 13 studies in adult ICU patients, Cutuli and colleagues ³² found that the majority of studies did not control for clinical confounders ³³ or used suboptimal statistical methods. ^{21, 22, 23, 24} Because of such limitations and the wide types of non-invasive thermometry used to assess BT, both systematic reviews ^{16, 32} could not synthesise the data available and/or obtain conclusive results. Given the above problems, we aimed to overcome such methodological issues.

Our findings are similar to and extend those of Moran et al, ²⁰ who investigated the accuracy of tympanic infrared and axillary glass mercury thermometers in estimating core temperature in a cohort of critically ill patients. In their study, the accuracy of the axillary glass mercury method was superior to the tympanic infrared method. However, the authors did not specify whether non-invasive BT measurement methods had been calibrated before the commencement of the study and set tympanic infrared in “core mode”. In our study, the tympanic infrared devices were unadjusted to assess core temperature according to the manufacturer algorithm ⁴ (ear mode +1.04°C). However, such conversion would have only influenced the direction of the bias, while the magnitude of inaccuracy would have remained unchanged.

In contrast with our observations, Fulbrook and colleagues ³⁴ reported that axillary chemical dots slightly underestimated core temperature compared with the tympanic infrared method. However, the authors only included a small number of patients and observations. More recently, Farnell et al ²⁵ reported axillary chemical dot failures in seven of 160 BT measurements in 25 critically ill patients, which confirms the rate of “non-readable” strips seen in our study. In addition, axillary chemical dots appeared to show lower accuracy than the tympanic infrared method in estimating core temperature. However, as the authors did not account for repeated within-subject measurements, the robustness of such findings is unclear. Finally, Myny et al35 argued in favour of a better accuracy for the temporal scanner in estimating core temperature in 57 critically ill patients. Yet, the number of observations was small and the participation of only one operator increased observer bias and limited external validity.

The low accuracy of axillary chemical dot, tympanic infrared and temporal scanner methods in estimating core temperature implies that these methods are clinically inadequate for ICU patients. Such implication is of importance for research in ICU patients with specific conditions (eg, suspected infection), for whom BT may represent an inclusion criterion. Moreover, the wide LoA imply that “average” adjustments upward of 0.5 or 1 degree are not helpful. In addition, the impact on bias and LoA of clinical explanatory variables implies that any link between these techniques and core temperature will markedly vary according to clinical circumstances. Finally, the high rate of failure for the chemical dot approach is of concern and implies limited applicability. In their aggregate, our findings imply that current non-invasive temperature measurement techniques are inappropriate and potentially misleading in many ICU patients.

Our study has several strengths. First, we investigated the accuracy of the most widely used non-invasive BT measurement methods in ICU. Second, we used the Bland–Altman approach accounting for repeated within-subject measures, which allowed us to overcome methodological issues raised by previous systematic reviews. ^{16, 32} Third, we performed regression analysis to identify predictors of BT, which allowed us to explore the impact of clinical variables on temperature estimation. Nonetheless, our study carries some limitations. It was not blinded and on-duty bedside nurses may have been biased in their data collection. However, nurses were asked to measure BT as they do in daily clinical practice and we estimate that more than 100 nurses were involved in the study measurements, which gives our findings a robust level of external validity and makes systematic bias unlikely. We tested only few non-invasive BT measurement methods and our results do not apply to other types of thermometry such as the newest zero heat flux. ^{36, 37} However, such thermometers are of limited daily use because of their cost. We found wider LoA compared with previous articles in this field, ³² supporting the limited use of non-invasive BT measurement methods in the ICU. However, this finding may be explained by the design of our study, which controlled for clinical confounders and adopted adequate statistical method to account for significant explanatory variables and multiple within-subject measurements. Finally, we merged different invasive BT methods into a single variable; however, previous evidence ²⁰ investigated the accuracy of temperature-sensing urinary catheters in estimating core temperature and confirmed their negligible bias and LoA.

Non-invasive body temperature measurement techniques are inaccurate, show bias and have wide LoA in relation to core temperature in ICU patients. Given the diagnostic and therapeutic importance of this vital sign in critically ill patients, the use of axillary chemical dot, tympanic infrared and temporal scanner methods appears suboptimal in the ICU.

Acknowledgements: We thank the ICU nurses who work at the Austin Hospital and Warringal Private Hospital. This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.