Thermal imaging-based core peripheral temperature difference measurement for neonatal monitoring in the NICU

Authors: Nantao Zhang, Xiaoyan Song, Junli He, Fengchao Liang, Jie Yang, Wenjin Wang

Abstract

The core-peripheral temperature difference (CPTD) refers to the difference between the body's core temperature (e.g., chest or abdomen) and peripheral skin temperature (e.g., hands or feet). It serves as a key biomarker for assessing the hemodynamic status of newborns and is an important early warning indicator of potential shock and severe infection. Measurement of CPTD in clinical practice currently requires the use of an infrared spot thermometer to measure the temperature of multiple body parts of a neonate, which is not possible for continuous and fully automatic long-term monitoring. To address these limitations, we propose a thermal infrared (TIR)-based approach that enables non-contact, fully automatic, and continuous CPTD measurement for neonates. The spatial redundancy property of TIR is utilised and combined with a deep learning-based body parsing model to automatically detect different body parts of a neonate, including the chest and limbs (e.g., hand or foot), and measure the temperatures of these two parts to derive their difference as CPTD. Although accurate measurement of the absolute temperature of the neonatal skin is difficult due to the calibration of the TIR camera and environmental influence, the temperature difference between different body parts that emphasizes the spatial contrast at certain moments can be reliably estimated, and it is independent of the subject and environment. In a prospective clinical trial involving 40 preterm infants, our TIR-based CPTD measurement showed a mean absolute error less than 0.3∘ C. Additionally, hand temperatures were, on average, 1.11°C higher than foot temperatures.Hand temperatures also showed a more pronounced response to changes in core temperature, suggesting that they may be better indicators of fluctuations in core temperature. Finally, we investigated the relationship between TIR-based CPTD and infant circulatory disorders. We find that infants with circulatory disorders typically have higher CPTD values, which demonstrates the clinical potential of our methods in reflecting functional limitations of the circulatory system in newborns. To our knowledge, this is the first clinical showcase of using a TIR camera for continuous non-contact CPTD monitoring of preterm infants in the hospital neonatal intensive care unit (NICU), providing important preliminary findings that may enrich the video health monitoring applications in the NICU.

Introduction

Neonatal shock, a leading cause of neonatal mortality, is a clinical syndrome resulting from reduced effective blood volume, leading to microcirculatory insufficiency and multi-organ dysfunction. Neonatal septic shock, triggered by severe infection, is the most prevalent form [1]. This state may lead to hemodynamic dysfunction, manifested by decreased blood pressure, inadequate tissue perfusion, and organ dysfunction [2,3]. Current tools for monitoring shocks can be categorized into invasive methods (e.g., pulmonary artery catheter, central venous pressure, arterial catheter) and non-invasive methods (e.g., ultrasonography, fluid challenge response) methods [4]. Invasive techniques often carry risks of complications and have limited applicability in neonatal care [5–7]. Non-invasive methods, though less invasive, lack the ability to provide fully-automatic, continuous and real-time assessment [8]. Core-Peripheral Temperature Difference (CPTD), defined as the temperature difference between the core and peripheral regions of the body, can serve as a valuable biomarker for monitoring the shock [9]: (1)CPTD=Tcore−Tperipheral.

Figure 1 illustrates the physiological mechanisms underlying the increase in CPTD during shock. Activation of the sympathetic nervous system (SNS) and renin-angiotensin-aldosterone system (RAAS) leads to the release of ligands (catecholamines and angiotensin II). These ligands bind to cellular receptors in peripheral tissues (e.g., hand) and core organs (e.g., heart), inducing peripheral vasoconstriction, increased heart rate, and enhanced cardiac contractility. As a consequence of reduced blood perfusion to peripheral tissues, peripheral temperature decreases while core temperature remains relatively stable. This disparity in temperature between core and peripheral regions results in a pronounced increase in CPTD, reflecting the body’s compensatory response to shock [10–12].

According to the literature and clinical standards, CPTD is typically considered to be within the range of 3°C to 4° C to ensure the stability of peripheral circulation in neonates [10]. When CPTD exceeds 5° C, it is regarded as a clinical threshold, which may signal potential complications such as inadequate peripheral circulation, hypovolemia, or shock. Real-time CPTD monitoring is thus crucial for various clinical applications (e.g., early recognition of shock, assessment of peripheral perfusion insufficiency, monitoring of sepsis, and management of extracorporeal membrane oxygenation), particularly for neonatal monitoring in the neonatal intensive care unit (NICU). Due to the immature development of the circulatory system in preterm infants, which can lead to inadequate blood perfusion and shock, continuous CPTD monitoring is indispensable in the NICU for timely detection and prevention of deteriorations.

Although CPTD holds potential clinical value, its widespread use in clinical practice has been limited due to various constraints, such as inconvenient handling. Currently, CPTD measurements primarily rely on infrared spot thermometers to measure temperatures in the patient’s core and peripheral areas. This method is inefficient and not fully automated, requiring caregivers to manually take measurements for spot-checks. In addition, traditional assessment methods also include thermistor probes or manual observations. Thermistor probes may cause skin irritation, particularly in premature infants, while also increasing the risk of infection. Furthermore, thermistor probes cannot provide global temperature information at different body parts. Manual observation is inaccurate and susceptible to subjective interpretation.

To overcome these limitations, we investigated the feasibility of using thermal infrared (TIR) cameras for CPTD estimation, demonstrating the potential for non-contact, global, and continuous monitoring. The key advantage of TIR cameras lies in their ability in performing non-contact and continuous temperature measurements; however, a significant drawback is the accuracy of measuring absolute temperature, as it is difficult to obtain accurate absolute temperature readings without proper calibration in specific environments. In this study, we focused on measuring CPTD, which represents the relative difference between core body temperature and peripheral skin temperature. This approach effectively mitigates the potential impact of inaccurate calibration. Since the method only measuring the temperature difference between the core and peripheral regions, it avoids potential issues related to environmental temperature fluctuations and device calibration in the measurement of absolute temperature.

In terms of ease of use, cost, and reliability, TIR-based method offers several advantages over traditional methods. TIR cameras are easier to use, eliminating the need for probe placement and calibration, which allows for faster and non-contact temperature measurements. Although the initial equipment cost of TIR cameras is higher, their overall cost is lower in the long-term application, as they do not rely on consumables like traditional probes. In a study by Buono et al., the TIR-based method demonstrated high consistency in measurements when compared to traditional thermistor probe methods [13]. Furthermore, TIR-based method can integrate temporal temperature changes, enhancing the stability and reliability of measurement.

All objects with temperatures above absolute zero (0 K, or −273.15° C) emit infrared radiation [14]. A TIR camera captures this infrared radiation from the object’s surface using an array of infrared sensors, converts it into an electrical signal, and generates a thermal image representing the temperature distribution [15]. The imaging principle of TIR cameras is primarily based on Planck’s Law of Blackbody Radiation. Planck’s law describes the spectral distribution of radiant energy emitted by a blackbody at a specific temperature T , as expressed by the following equation [16]: (2)Mbb(T)=2hc2λ5⋅1exp⁡(hcλkT)−1, where Mbb(T) represents the radiant energy emitted per unit time, per unit surface area, and per unit wavelength interval by a blackbody at temperature T . The equation includes Planck’s constant ( h ), the speed of light ( c ), Boltzmann’s constant ( k ), the absolute temperature ( T ), and the wavelength of the radiation ( λ ). Real objects, however, emit radiation ( M(T) ) that is typically lower than blackbody radiation and can be expressed (3)M(T)=ϵ⋅Mbb(T), where ϵ denotes the emissivity of the object. Human skin typically has an emissivity of approximately 0.98 [17]. The signal detected by the TIR camera includes not only the radiation from the target object ( Mobj ) but also environmental radiation ( Mo ) and atmospheric radiation ( Matm ). The detector signal can be expressed as [18]: (4)s(T)=K[ϵ×Patm×Mob(T)+(1−ϵ)Patm×Mo(To)+(1−Patm)Matm(Tatm)], where the detector signal is influenced by several factors, including the detector’s sensitivity factor ( K ), the atmospheric transmission coefficient ( Patm ), the target object’s temperature ( T ), the atmospheric temperature ( Tatm ), and the ambient temperature ( To ). The TIR camera uses these electronic signals to create a pseudo-color images, where colors represent temperature variations. Hotter areas are typically depicted in bright or red hues, while colder areas are shown in dark or blue tones. Various color palettes can be used to enhance visualization of temperature distribution, but the colors represent the spatial temperature contrast, rather than absolute temperature values.

The core research question of this study Can thermography cameras provide a non-contact, continuous, and reliable method for assessing neonatal peripheral circulation by measuring CPTD? We hypothesize that, by integrating deep learning techniques, TIR cameras can achieve automated CPTD detection in NICU settings with high consistency and clinical applicability compared to traditional approaches using infrared spot thermometers for manual checks.

In this study, we leveraged deep learning to automate the identification of the chest-abdomen (core region) and hand (peripheral region) in neonatal TIR videos for continuous CPTD monitoring. Specifically, we fine-tuned a high-resolution Transformer architecture to enhance pose estimation of neonatal TIR images by decreasing the patch size and step size and increasing the number of Transformer layers at higher resolution stages. The model successfully identified the hand and chest-abdominal regions of newborns with high accuracy, achieving average precision scores (AP) of 76% and 85.5%, respectively, enabling precise calculation of temperature difference. To minimize the negative impact of calibration errors, such as baseline offsets, we opted to calculate the relative temperature difference instead of the absolute temperature difference. A prospective study involving 40 preterm infants in a NICU department demonstrated the feasibility of continuous, fully-automatic CPTD monitoring. The estimated MAE of the CPTD were 0.2969° C. Additionally, we studied the relationship between core temperature and hand/foot temperatures. The results indicate that hand temperature can more reliably reflect the changes in core temperature than foot temperature. Finally, we analysed the relationship between CPTD and neonatal circulatory disorders, and showed that infants with circulatory disorders had higher CPTD. The application of our method will contribute to a more timely and effective monitoring of CPTD in preterm infants in NICU, guaranteeing a more dedicated monitoring of their physiological changes in their most vulnerable stage.

The remainder of this paper is organized as Section 2 reviews the applications of TIR cameras in NCIU. Section 3 outlines the materials and methods of our study. Section 4 details the execution process of clinical experiments. Section 5 discusses the experimental results. Finally, Section 6 concludes the paper, summarizing our findings and contributions.

Related works

Since the first showcase of RGB cameras to measure skin perfusion changes in 2000, non-contact camera monitoring techniques have made significant progress in patient monitoring in healthcare. Before reviewing the clinical research on contactless vital signs monitoring, we recommend consulting the Elsevier book Contactless Vital Signs Monitoring [19], which provides a comprehensive overview of the principles and advancements in the field of non-invasive remote monitoring of physiological parameters. The book offers valuable insights into the state-of-the-art technologies of contactless monitoring systems, serving as an essential reference for understanding the latest developments in this field. Recently, Liao et al. proposed a heart rate variability monitoring technique based on RGB cameras for assessing the maturity of the autonomic nervous system in neonates and validated its feasibility in NICUs [20]. Zeng et al. achieved high-precision detection of cardiorespiratory parameters using an RGB camera, and the results met the standards defined by the ANSI/AAMI EC13:2002 standard [21]. Ye et al. used an RGB camera for continuous, contact-free blood oxygen saturation monitoring in the NICU. Their clinical trial on 22 preterm infants showed that the MAE was less than 4% [22]. Huang et al. demonstrated through a multicentre clinical trial that an RGB camera-based infant sleep-wake monitoring technique with consistent depth representation constraints achieved more than 85% accuracy and generalisation performance across different ages and environments [23]. Zhu et al. proposed an RGB camera-based sleep position detection method that accurately identifies sleep postures, even when the body is occluded, by using body rolling motion and head orientation estimation in video. This method can be applied to clinical monitoring in sleep medicine [24]. Wang et al. introduced a contactless intelligent monitoring system that introduces a near-infrared camera combined with physiological signals such as heart rate and respiration rate to enhance the accuracy of sleep stage classification through the fusion of multidimensional data [25].

TIR cameras offer distinct advantages over regular RGB cameras by capturing thermal radiation instead of visible light, allowing them to generate images based on an object’s temperature distribution. Herman et al. demonstrated the potential of using TIR cameras for early melanoma diagnosis [26]. TIR cameras have been used to monitor cardiorespiratory signals, as shown by Cho et al., who used a TIR camera to measure respiratory rate [27]. In neonatal populations, Beppu et al. employed thermal infrared cameras to monitor newborn skin temperature, helping to regulate body temperature and prevent hypothermia [28]. Barson et al. detected inflammatory lesions in infant chest-abdomen using thermography [29]. TIR cameras are also recommended for identifying external risks associated with sudden infant death syndrome [30]. Lyra et al. tracked respiratory patterns in neonates by monitoring subtle temperature changes around the nostrils and mouth, providing a non-contact method for respiratory rate monitoring, especially useful for premature infants at risk of respiratory distress [31]. Additionally, Ornek et al. developed a TIR camera-based system to monitor neonatal health by analyzing thermal symmetry and localized temperature anomalies (e.g., caused by necrotizing enterocolitis or respiratory distress), enabling real-time detection of infections for early clinical intervention [32].

Researchers have also utilized TIR cameras to assess peripheral circulation status. Jorge et al. developed a non-contact method to evaluate peripheral arterial hemodynamics using a TIR camera and validated it in an ice-water stimulation experiment [33]. Notably, the validation in this study was limited to healthy adults and has yet to be performed in more complex clinical scenarios, such as neonatal populations. The relative temperatures in this study were calculated based on each region’s absolute temperature before stimulation, rather than on the temperature difference between core and peripheral regions. Additionally, the investigators manually labeled the core and peripheral regions without employing an automated method for region-of-interest detection. Beppu et al. proposed a novel method, combining the YOLOv5 object detector and a decision tree, to automatically identify six body parts (head, torso, left/right arm, and left/right leg) in TIR images of newborns for temperature monitoring [28]. Their method leverages the relative positional relationships between body parts to improve detection accuracy. However, it is important to note that the performance of this approach is highly sensitive to the posture of the subject being imaged, as the relative positions of body parts can vary significantly across different poses. Furthermore, this study is limited to the measurement of absolute temperature and does not focus on CPTD.

In conclusion, although TIR cameras have demonstrated potentials in various fields within NICU environments, to our best knowledge, no studies have yet utilized TIR cameras for monitoring CPTD as a measure of peripheral perfusion and circulatory status. Our research innovatively applies TIR cameras to monitor CPTD, evaluating circulatory function by analyzing spatial temperature differences.

Materials and methods

This study was conducted in the NICU of Nanfang Hospital, Southern Medical University, China and Shenzhen University General Hospital, China. As thermal images exclusively capture infrared radiation, the study does not have the issue of privacy invasion. The study does not have disturbance or alternation to the patients’ routine care procedures. As a result, the clinical trial was approved by the Ethics Committee of the Southern University of Science and Technology (No. 2022-007-03), Nanfang Hospital of Southern Medical University (No. NFEC-2022-100) with the written informed consent of the infant’s legal guardian. The study period spanned from February 2024 to August 2024. We employed a longitudinal study designed to continuously monitor and record changes in CPTD of infants over time.

The experiments utilized a mini640_09110X SWR SA thermal infrared camera (Iray, China) equipped with a vanadium oxide uncooled infrared focal plane sensor, a 9.1 mm focal length lens, an F1.0 aperture, a 48.7∘

×

38.6∘ field of view (FOV), and a thermal sensitivity of ≤ 50 mK. The camera was externally triggered at a constant frame rate of 15 Hz, capturing image data at the resolution of 512 × 640 pixels. A BNT410CN Infant Infrared Spot Thermometer (Braun, Germany) was used to record the CPTD reference manually. The spot thermometer offered a temperature measurement accuracy of ±

0.2∘ C within the range of 35∘ C to 42∘ C.

To investigate the correlation between our measurements and neonatal states, 40 infants with severe conditions were selected for monitoring. Inclusion criteria encompassed critically ill and preterm infants with a gestational age of 29 to 40 weeks, regardless of postnatal age. Some of the included patients suffered from diseases that may affect the circulatory system. Therefore, we divided the patients into two major groups according to disease group with normal circulation and group with abnormal circulation (see Table 1). The group with normal circulation includes some digestive problems (e.g., vomiting, gastroesophageal reflux (GER)), mild anaemia (MA), Neonatal Hyperbilirubinemia (NNH), etc., and these disorders have less direct impact on the heart and vascular system. The group with abnormal circulation mainly includes those diseases that may directly or indirectly affect the heart, circulation, or vascular system of the newborn, such as congenital heart disease (e.g., atrial septal defect (ASD), patent ductus arteriosus (PDA)) and haematological disorders (e.g., neonatal anaemia (NA), ABO haemolytic disease (ABO HD)).

The experimental setup, as depicted in Fig. 2, placed the infant in a supine position within an open incubator, allowing for unrestricted movement. The TIR camera was mounted on a stable trolley at a suitable distance from the infant to ensure that the entire body, or at least the exposed chest-abdomen and hands, were captured. Video recordings were conducted for a duration of two minutes. Data collection was performed in a temperature-controlled and stable NICU environment, effectively minimizing the impact of environmental factors such as temperature, humidity, and airflow. According to the hospital where the clinical trial was organized, the NICU room temperature was maintained at 25° C. Consequently, we assumed a relatively stable patient body temperature throughout the recording session. An infrared spot thermometer was used to measure the temperature of the patient’s hand and chest-abdomen at both the beginning and end of the recording. This approach enabled the calculation of respective mean values for the hand and chest-abdomen during the recording period, thereby mitigating the impact of sensor noise and yielding more stable and reliable reference.

Implementation

Data pre-processing

Raw temperature data stored in the binary format needed to be parsed and transformed for visualisation. To enhance the visualisation, the data for each patient was normalised and a "jet" colour mapping was applied to generate a pseudo-colour image dataset. Next, we created a dataset called DL dataset for model training and validation. The dataset was constructed by randomly selecting 1 frame from every 15 frames in each subject’s data.

In the DL dataset, the "COCO Annotator" tool was employed to manually annotate the Ground-Truth (GT) keypoints. COCO Annotator is an open-source image annotation tool designed for creating and manipulating image datasets, particularly those adhering to the COCO (Common Objects in Context) format [34]. While the standard COCO format defines 17 commonly used human keypoints, privacy concerns and clinical requirements led to a modified labeling scheme. Specifically, the labeling of head keypoints (e.g., nose, ears) was reduced, while the labeling of fingers, toes, chin, chest, and navel was increased (Fig. 3(a)). By annotating these 21 keypoints, the newborn’s posture can be roughly described [35]. Figure 3(b) illustrates a labeled example, showing all keypoints of an infant and their corresponding point connections.

The hand region was defined by the keypoints of the wrist and middle finger. By connecting the keypoints of the middle finger (R5, L5) and the wrist (R4, L4), a line segment was formed. A square centered on the midpoint of this line segment, with the length of the line segment as the side length, defined the hand region. The foot region is then defined in the same way with key points at the toe (R9, L9) and the ankle (R8, L8). The chest-abdominal region was delineated by a rectangular area connected by four left shoulder (R2), right shoulder (L2), left hip (R6), and right hip (L6). In this study, we used two different approaches to select regions of interest (ROIs). When assessing TIR-based CPTD and reference value agreement, we selected the hand with the higher confidence level as the final detection to ensure as much continuity of monitoring as possible. This is because medical interventions, postural changes, or occlusion may occur during filming, making it impossible to film both hands simultaneously. Instead, when investigating the relationship between core temperature and hand/foot temperatures, we screened the images of each newborn and retained only those images in which both the hand and foot were clearly visible. When calculating CPTD, we used the average temperature of both hands or feet as a proxy for peripheral temperature.

Key point detection

In this study, the AggPose model proposed by Cao et al. [35] is used for infant keypoint detection. The structure of the AggPose model is mainly based on the visual Transformer architecture, which removes the traditional convolutional layers and contains the following key (1) Overlapping Patch Embedding: The input image is segmented into overlapping pixel squares (i.e., patches, with a size of 7 × 7 pixels and a step size of 4 pixels) to extract fine-grained spatial features; (2) Multi-stage Aggregation Vision Transformers (AViT) by stacking multiple layers from high to low resolution, each layer consists of multiple AViT Blocks, which contain Efficient Attention and Multilayer Perceptron (MLP) layers inside them for further extraction and expression of features; (3) Cross-resolution MLP information fusion between different resolution layers through the MLP layer to achieve multi-scale feature integration. After cross-layer feature aggregation, the model generates the final prediction of human body key point locations. The authors first pre-train the AggPose model on the COCO dataset, loading the Mix Transformer parameters pre-trained on ImageNet to enable the model to learn generic human pose features. The authors then fine-tuned the infant posture RGB dataset they constructed to improve detection accuracy. According to Dosovitskiy et al. [36] a smaller Patch and step size enables finer sampling and improves the model’s ability to capture image details. Therefore, we reduced the Patch size to 5 × 5 pixels and the step size to 3 pixels, allowing the image to be divided into more overlapping chunks, further enhancing the detail capturing ability. In order to take full advantage of the high-resolution detail information from the 5 × 5 pixel patch and the 3-pixel step size setting, we increased the number of Transformer layers in the high-resolution stage (from 3 to 6 layers in stage 1) to better capture the feature distribution of the TIR image and improve the recognition performance.

Training and validation of the model

The DL dataset underwent leave-one-out cross-validation (LOOCV), where the data from one patient in each fold served as the validation set, and the rest were used for training. Transfer learning was employed using the pre-trained model weights provided by AggPose. Given that these weights were pre-trained on an RGB image dataset with 256x192 pixels, adjustments were made using an adaptive pooling layer. The torchvision transforms module served as a data augmentation tool, enabling various geometric transformations such as rotation, scaling, and translation. Since thermal images are not dependent on visible light color information, color space transformations were unnecessary.

Training was conducted on a GPU for up to 200 epochs. The initial learning rate was set to 0.001 and dynamically adjusted using the ReduceLROnPlateau the learning rate was decreased by a factor of ten if the validation loss failed to improve within 10 consecutive epochs, and the minimum learning rate was established at 0.00001 to prevent excessively low learning rates. To mitigate overfitting, loss values were monitored throughout training, and early stopping was implemented to terminate training at an appropriate time. Following training, AP and average recall (AR) were calculated using IoU to evaluate the overlap between predicted and actual regions, focusing primarily on hand and chest-abdominal regions. AP and AR are metrics commonly used to evaluate the performance of object detection and image segmentation models. AP assesses the model’s overall accuracy across various IoU thresholds, accounting for the proportion of true positive detections. AR evaluates the model’s coverage capability across different IoU thresholds, reflecting the model’s ability to identify all actual objects. We considered IoU thresholds ranging from 0.50 to 0.95, calculating the average at intervals of 0.05 to obtain both AP and AR values.

Extraction of CPTD

The process of extracting signals from TIR videos requires the conversion of raw temperature data into grey scale frames. To remove background noise, a threshold of 191.25 was applied to the TIR data, excluding pixels with grey values below this threshold (i.e., the lower 75% of the grey range) and retaining only high grey value areas to highlight postural information about the neonate. Subsequently, model inference acquires key points of the body to identify ROIs to extract core-periphery temperatures. To mitigate the impact of anomalous temperature readings on the CPTD calculation, outliers were identified and removed prior to analysis. Outliers were defined using a three-tiered (1) temperatures outside the physiological range of 32–42 °C, (2) abrupt temperature changes exceeding 2 °C between consecutive measurements, and (3) values deviating significantly from the local mean within a 30-frame sliding window in the time domain. Specifically, temperatures that deviate by more than three times the standard deviation from the window’s mean (i.e., outside the range of mean ± 3 standard deviations) were classified as outliers. Interpolation using cubic spline methods was employed to replace these outliers with estimated values, ensuring data continuity for subsequent analysis.

To evaluate the validity of the temperature detection method and analyze potential errors, mean absolute error (MAE) was calculated between the relative temperature difference obtained by the model and that measured by the infrared spot thermometer. The MAE is defined as the average of the absolute differences between predicted and actual values and is commonly used as a measure of accuracy between predicted and observed data. The CPTD extracted from the video represents a continuously varying time series. To compare with the reference value, the mean value of the CPTD time series for each neonate, denoted as Tm , was calculated. Let Tm,i represents the temperature difference of the i -th neonate predicted by the model, and Tr,i represents the actual temperature difference of the i -th neonate measured by the infrared spot thermometer. With N subjects, MAE were calculated as (5)MAE=1N∑i=1N|Tm,i−Tr,i|.

To visualize the consistency between the proposed method and the reference, Bland-Altman plots and correlation analyses were employed. In order to investigate the relationship between core temperature and hand/foot temperatures, visual analyses were performed using scatter plots.

Results and discussion

Evaluation of model performance

Accurate automated detection of core and peripheral sites is crucial for successful CPTD measurement. Fine-tuning the AggPose detection model effectively adapted it to the neonatal TIR dataset. Table 2 presents the average AP and AR statistics for hand and chest-abdominal detection in the 40 infants studied, comparing the baseline model with the fine-tuned model. The baseline model achieved a mean AP and AR of 64.1% and 70.4% for hand detection, and a mean AP of 73.3% and AR of 75.2% for chest-abdominal detection. After fine-tuning, the model’s performance improved, with a mean AP of 76.0% and AR of 76.6% for hand detection, and a mean AP of 85.5% and AR of 84.3% for chest-abdominal detection. Compared to the baseline model, the fine-tuned model showed improvements of 11.9% and 6.2% in AP and AR for hand detection, respectively, and 12.2% and 9.1% in AP and AR for chest-abdominal detection, indicating the effectiveness of the fine-tuning in enhancing the model’s performance.

Fig. 4(a) presents the heatmap distribution of AP and AR for each sample. Analysis reveals that folds 9, 15, and 30 exhibit the highest mean AP, while folds 4 and 10 demonstrate the lowest mean AP. Visualizing the heatmap results for these cases (Fig. 4(b)) provides insights into individual characteristics that may explain the lower AP. For example, in fold 4, the model struggled to identify the hand regions due to the left hand being obscured by the infant’s head and the unclear boundary of the right hand. In fold 10, the infant’s hands were covered by a blanket due to unrestricted movement. Conversely, the infants in folds 9, 15, and 30 maintained more stable postures, with clearly visible hand and chest-abdominal regions free from significant obstructions, enabling the model to accurately recognize them.

Assessment of relative temperature differences

The predicted hand and chest-abdominal regions were used to estimate CPTD from infrared TIR frames. Figure 5 illustrates the continuous monitoring of CPTD in 40 preterm inpatients. Overall, the TIR-based CPTD measurement exhibited good agreement with the reference, with minimal serial deviations. However, the TIR-based CPTD measurement values were slightly higher. This can be attributed to the camera’s wider measurement range, capturing both the high-temperature chest-abdomen and the low-temperature hand, leading to a larger measured temperature difference. As observed in Fig. 4, the 4th and 10th folds displayed the lowest AP, influencing the measurement of their temperature difference. Occlusion and movement are likely to cause the camera to capture other parts of the body, making it difficult to obtain the true body temperature of the hand. The mean value of CPTD was calculated for all frames of each infant and compared with the reference value, and the MAE between TIR-based and reference-based CPTD measurements is 0.2969∘ C, indicating high accuracy and stability of CPTD monitoring of preterm infants.

Fig. 6 presents Bland-Altman plots and linear correlation plots for all subjects whose CPTD measurements were analyzed. The results demonstrate a strong correlation between the proposed TIR-based method and the reference method (i.e., the infrared spot temperature gun method). The Bland-Altman plots reveal an average difference of 0.22∘ C between the two methods, with 95% limits of agreement ranging from -0.30∘ C to 0.74∘ C. These findings indicate a high degree of agreement between the two methods.

Figure 7 presents a scatter plot illustrating the relationship between core temperature and hand/foot temperature. The plot reveals that, under equivalent core temperatures, hand temperatures consistently exceed foot temperatures by an average of 1.11∘ C. Additionally, the fitted curves indicate that as core temperature increases, both hand and foot temperatures also increase. However, the hand temperature demonstrates a more pronounced response to the changes in core temperature (slope of 0.37), while the foot temperature response is relatively subdued (slope of 0.15). This difference may be due to the hand’s proximity to the heart and more active circulation, aligning with the findings of Johnson et al. [37] and Rowell et al. [38], who noted that high-perfusion areas (such as the hand) are more influenced by fluctuations in core temperature. In contrast, the greater distance of the foot from the heart and the reduced blood flow rate–referring specifically to the volume of blood delivered per unit of time–make its temperature more vulnerable to external environmental factors. This leads to a slower response to changes in core temperature, a phenomenon also discussed in cardiovascular research by Gonzalez-Alonso et al. [39]

We calculated the mean CPTD values of the normal and abnormal circulation groups separately (Fig. 8). The results show similar performance between the two measurement methods, both showing that CPTD in the normal group is significantly lower than in the abnormal group. To further investigate these findings, we conducted an analysis of variance (ANOVA) to examine the significance of difference between CPTD values obtained by the compared methods (TIR-based method vs. reference method) with respect to different circulatory statuses (normal circulation vs. abnormal circulation). As shown in Table 3, the one-way ANOVA results show that in the abnormal circulation group, the difference in CPTD values measured by the reference method and the TIR-based method is not significant (F = 1.688, p-value = 0.203). Similarly, in the normal circulation group, the difference between the two methods is neither significant (F = 1.838, p-value = 0.182). This indicates that within the same circulation condition, whether normal or abnormal, the TIR-based method and the reference method maintain high consistency, further supporting the reliability of the TIR-based method for measuring CPTD values. However, when comparing CPTD values under different circulation conditions, both the TIR-based and the reference method show significant differences. Under the reference method, the CPTD values differ significantly between the abnormal circulation and normal circulation groups (F = 65.48, p-value = 8.63e-10); similarly, under the TIR-based method, the CPTD values in the abnormal circulation group are significantly higher than those in the normal circulation group (F = 102.81, p-value = 2.32e-12). These results indicate that different circulatory statuses have a significant impact on CPTD values, and the TIR-based method has high consistency with the reference method in distinguishing the circulatory status of newborns. In conclusion, the TIR-based method demonstrates measurement consistency close to that of the reference method in ANOVA and a strong ability to distinguish different circulation states of infants.

The higher CPTD of the neonates in the abnormal group reflected the declined performance of their circulatory systems. To better understand this phenomenon, we reviewed the specific physiological mechanisms underlying the sick infants in the abnormal circulatory function group. In newborns with congenital heart defects such as ASD and PDA, abnormal shunts within the heart or large blood vessels alter the normal path of blood flow. For example, ASD causes oxygenated blood to flow from the left atrium to the right atrium [40], while PDA causes abnormal flow between arteries and veins [41]. This abnormal shunt prevents efficient oxygenated blood from being adequately transported throughout the body, especially to peripheral sites, leading to a decrease in the efficiency of circulation and the formation of a classic macrocirculatory disorder (i.e., obstruction of blood circulation in the arterial and venous vasculature). This obstruction results in significantly lower peripheral temperatures compared to core temperatures, increasing the temperature difference between the core and periphery [42]. In addition, neonates with blood disorders (e.g., NA and ABO HD) have a reduced number of red blood cells, resulting in a decrease in the oxygen-carrying capacity of the blood [43,44], which results in hypoxia of the peripheral tissues, a decrease in the metabolic rate, and a consequent decrease in the temperature, which leads to an increase in the CPTD value [45]. Aside from macrocirculatory disturbances, microcirculatory disorders are also an important factor in elevated CPTD. Microcirculation in neonates is critical for blood transport to peripheral capillaries, and in the presence of ABO HD, blood viscosity increases and the resistance to blood flow in the microvasculature increases, limiting the ability of blood to reach peripheral sites [44]. The restricted microcirculation further exacerbates the temperature difference between the core and the periphery, resulting in significantly higher CPTD values.

The results of this clinical trial demonstrated the feasibility of using TIR camera combined with image processing techniques for continuous CPTD monitoring of infants. With an estimated MAE of less than 0.3∘ C, the method shows promising potential for non-contact assessment of peripheral circulation. Limitations of this study include the (1) this study was not a multicentre clinical study with patient data from different regions, which may make it difficult to directly generalise to other neonatal care settings. Therefore, we will conduct a multicentre study in near future to verify the applicability and reliability of the method in a wider range of healthcare scenarios, (2) the incubator’s thermal insulation requires the TIR camera to be placed inside to capture temperature data, posing a potential falling hazard. Mechanical instability (e.g., loose mounting brackets) or accidental collisions during clinical procedures (e.g., emergency interventions or routine adjustments) could cause the camera to detach. Additionally, infant privacy concerns must be considered. Despite these challenges, the TIR camera offers significant benefits for NICU infant monitoring and clinical it improves efficiency by enabling rapid, non-contact temperature measurements, reduces the risk of cross-contamination by minimizing direct contact with newborns, significantly improves the comfort of infant monitoring, and supports continuous monitoring in 24 hours/7 days, providing timely feedback for early warning and intervention, especially for critically ill preterm infants. With further development, TIR camera is expected to enhance the NICU monitoring, offering a more efficient, safe, and reliable method for circulatory monitoring.

Motivated by the findings of this study, which highlight the potential of TIR cameras in assessing circulatory status, we plan to further investigate the temporal dynamics of temperature and its correlation with blood perfusion and the circulatory system. Specifically, we will design long-term, multi-stage monitoring experiments to record temperature variations in both the core and peripheral regions. We will analyze how these temporal dynamics can be combined with the spatial distribution of CPTD to provide a more comprehensive assessment of infant circulatory status. Additionally, we found that monitoring peripheral circulation could be enhanced by incorporating the perfusion index (PI), a crucial indicator of circulatory health. Therefore, we plan to explore the opportunity of combining TIR cameras with RGB/NIR cameras. RGB/NIR cameras can provide direct signals related to blood perfusion, while TIR cameras reflect implicitly the outcome of perfusion, which is complementary to RGB/NIR cameras. By combining these two types of sensors, we can more comprehensively assess the circulatory status of neonates.

Conclusions

This study presents a novel approach to contactlessly and continuously monitor CPTD of neonates using a TIR camera. The proposed method addressed the limitations of current CPTD measurement approaches requiring manual spot-checks. Deep learning based infant body landmark detection automates the identification of core and peripheral sites of neonates, facilitating accurate calculation of their relative temperature difference. A clinical trial included 40 preterm infants in a NICU validated the feasibility of this approach. The results show that the MAE between the TIR-based CPTD measurements and those obtained using the infrared spot temperature gun is less than 0.3∘ C. In addition, we investigated the relationship between core temperature and hand/foot temperatures. The results showed that hand temperatures were, on average, 1.11∘ C higher than foot temperatures and that the hand responded more promptly to changes in core temperature, indicating that hand temperature could be a more sensitive surrogating indicator of changes in core temperature. Finally, we also found that CPTD was higher in infants with circulatory disorders, demonstrating the clinical value of TIR-based measurements in assessing circulatory function in neonates. Our research highlights the significant potential of CPTD for non-contact neonatal monitoring and early warning of adverse events in NICUs.