Eduard Kieser

ARTICLE IN PRESS JID: JJBE [m5G;July 19, 2017;18:50] Medical Engineering and Physics 0 0 0 (2017) 1–7 Contents lists available at ScienceDirect Medical Engineering and Physics journal homepage: www.elsevier.com/locate/medengphy Investigation of the feasibility of non-invasive optical sensors for the quantitative assessment of dehydration Cobus Visser a, Eduard Kieser a, Kiran Dellimore a,∗, Dawie van den Heever a, Johan Smith b a b Biomedical Engineering Research Group, Department of Mechanical and Mechatronic Engineering, Stellenbosch University, Stellenbosch, South Africa Department of Paediatrics & Child Health, Tygerberg Hospital & Stellenbosch University, Durbanville, South Africa a r t i c l e i n f o Article history: Received 27 September 2016 Revised 7 March 2017 Accepted 20 June 2017 Available online xxx JEL classifications:- Keywords: Hydration assessment Infant dehydration Objective measurement Quantitative markers a b s t r a c t This study explores the feasibility of prospectively assessing infant dehydration using four non-invasive, optical sensors based on the quantitative and objective measurement of various clinical markers of dehydration. The sensors were investigated to objectively and unobtrusively assess the hydration state of an infant based on the quantification of capillary refill time (CRT), skin recoil time (SRT), skin temperature profile (STP) and skin tissue hydration by means of infrared spectrometry (ISP). To evaluate the performance of the sensors a clinical study was conducted on a cohort of 10 infants (aged 6–36 months) with acute gastroenteritis. High sensitivity and specificity were exhibited by the sensors, in particular the STP and SRT sensors, when combined into a fusion regression model (sensitivity: 0.90, specificity: 0.78). The SRT and STP sensors and the fusion model all outperformed the commonly used “gold standard” clinical dehydration scales including the Gorelick scale (sensitivity: 0.56, specificity: 0.56), CDS scale (sensitivity: 1.0, specificity: 0.2) and WHO scale (sensitivity: 0.13, specificity: 0.79). These results suggest that objective and quantitative assessment of infant dehydration may be possible using the sensors investigated. However, further evaluation of the sensors on a larger sample population is needed before deploying them in a clinical setting. © 2017 IPEM. Published by Elsevier Ltd. All rights reserved. 1. Introduction Dehydration is a common consequence of acute diarrhea which is characterized by an excessive loss of body water, accompanied by a disruption of metabolic processes. While dehydration affects individuals of all ages, it is particularly life threatening in infants. This is because of their high turnover of fluids and solutes, which can be as much as three times that of adults due to their higher metabolic rates, increased surface area to volume ratio and higher total body water content [1]. Without appropriate intervention dehydration can cause hypovolemic shock due to loss of blood volume, organ failure and even death [2–3]. Dehydration resulting from acute diarrhea is among the leading causes of infant mortality globally, resulting in over 1.3 million infant deaths annually, with 99% of these occurring in developing countries [4]. The current “gold standards” for prospective dehydration assessment are hydration scales which aggregate various subjective clinical observations into a score used to gauge the level of dehydration of an infant [5–8]. Among the most commonly used clinical ∗ Corresponding author. E-mail addresses:-,-(K. Dellimore). markers of dehydration include: breathing strength and rate, capillary refill, eye appearance, fontanelle appearance, heart rate, level of physical activity, mouth (mucus membrane) dryness, neurologic state, pulse quality, skin turgor, absence of tear production, thirst, temperature difference between core and extremities, as well as urine color and smell [5–10]. The three most widely utilized hydration scales, based on a subset of clinical dehydration markers, include the clinical dehydration scale (CDS), the World Health Organization (WHO) dehydration scale, and the 4-point or 10-point Gorelick hydration scale [6,7]. However, previous work has shown that these “gold standard” clinical hydration scales perform poorly when used in rural and resource constrained settings [6,8], where the overwhelming majority of infant deaths occur due to diarrheal diseases [4]. Among the main causes for this is the poor adherence to established clinical practice which can be attributed to a lack of appropriate training of clinical staff, high patient loads and inadequate facilities; and the high degree of subjectivity in the scoring of many of the clinical dehydration markers [8–9]. This motivates the need for a more objective and quantitative means of assessing infant dehydration, suitable for deployment in rural and resource constrained settings. Ideally this method should be non-invasive, easy-to-use (i.e., require little training), reliable and low-cost. http://dx.doi.org/10.1016/j.medengphy-/© 2017 IPEM. Published by Elsevier Ltd. All rights reserved. Please cite this article as: C. Visser et al., Investigation of the feasibility of non-invasive optical sensors for the quantitative assessment of dehydration, Medical Engineering and Physics (2017), http://dx.doi.org/10.1016/j.medengphy- JID: JJBE 2 ARTICLE IN PRESS [m5G;July 19, 2017;18:50] C. Visser et al. / Medical Engineering and Physics 000 (2017) 1–7 Fig. 1. Overview of the four hydration sensors developed to investigate: (a) CRT, (b) SRT, (c, d) skin temperature gradient and (e, f) infrared absorption as a surrogate measure of total water content. Previous attempts at objectifying dehydration assessment using non-invasive techniques have focused largely on bioimpedance analysis (BIA), optical estimation of capillary refill time (CRT) and infrared spectrometry [11–25]. Koulmann et al. [11] investigated the estimation of the total water content using BIA in subjects who were dehydrated by exercise and heat stress, as well as glycerol induced hyper-hydration. They found that BIA could accurately estimate total body water in subjects who have succumbed to heat stress induced dehydration and glycerol induced hyper-hydration, but could only estimate total body water half of the time in subjects with exercise induced dehydration. Subsequently, Röthlingshöfer et al. [12] experimentally and numerically investigated the shift in impedance over varying frequencies in exercise induced dehydrated subjects. They showed that an impedance shift of 5–6% occurred as a result of exercise thereby suggesting bioimpedance as a potential marker for dehydration assessment. Both studies found that several confounding factors can influence BIA measurements leading to potentially unreliable prediction of an individual’s hydration status. Utter et al. [13] found BIA to be sensitive in detecting a 3.5% reduction in body weight due to water loss in a study on the hydration status of collegiate wrestlers. However, they reported that BIA is a late dehydration indicator with respect to plasma and urinary markers. CRT is a widely used dehydration marker since it is a fast and non-invasive method [14,15]. However, Blaxter et al. [15] and Pickard et al. [16] reported that manually performed CRT may be unreliable because it is influenced by many factors including age, ambient temperature, ambient light, pressure application and intra- and inter-observer reliability. Several studies have consequently attempted to quantify CRT objectively [17–21]. Shavit et al. [17] used digital videography to quantitatively measure the nail bed CRT of 83 children to assess the presence of severe dehydration (> or = 5% loss of body weight). They reported that digitally measured CRT yielded a specificity and sensitivity of 0.99 and 0.88, respectively, compared to manual CRT assessment which had a specificity and sensitivity of 0.85 and 0.60, respectively [17,20]. Subsequently, Bordoley et al. [18] and Kviesis-Kipge et al. [19,21] separately investigated optical CRT, but were unable to obtain consistent measurements. A possible drawback of these studies is that they determined CRT at the fingertip which may be a sub-optimal assessment site according to Strozik et al. [22,23] who found that CRT measurements at the sternum and forehead are more consistent than at the extremities. Another non-invasive approach for objective dehydration assessment is near infrared (NIR) videography, which was studied by Attas et al. [24] to determine the hydration state of human skin. Although this work did not focus on the clinical assessment of dehydration, they were able to measure the distribution of moisture in the skin in the wavelength bands at 970 nm, 1200 nm and 1450 nm. Application of this approach to dehydration assessment in the skin is supported by Nachabé et al. [25] who investigated measuring the NIR spectrum of lipid and water phantoms and achieved measurement errors within 5%. They also concluded that NIR spectrometry may be useful for in vivo real-time measurement of dehydration in tissue. This study aims to explore the basic feasibility of prospectively assessing infant dehydration using various non-invasive optical sensors based on the quantitative and objective measurement of several markers of dehydration: (i) skin turgor, (ii) capillary refill time, (iii) skin temperature gradient and (iv) infrared absorption as a surrogate measure of tissue water content. Sensor performance is evaluated by comparing sensor measurements with clinical assessments made during a clinical study on infants (aged 6–36 months) suffering from gastrointestinal distress. 2. Methods 2.1. Hydration sensors Four different non-invasive hydration sensors were investigated to determine their ability to quantitatively assess dehydration severity: 1 Capillary refill time (CRT) sensor was designed to quantify the capillary refill test performed by clinicians. The sensor consists of a silicone pressure pad that is mounted on a swing arm, over a Honeywell FS1500 force sensor which is connected to a controller via an ADS115 16-bit analog to digital converter (ADC) as shown in Fig. 1(a). When the swing arm is locked into place it can be used to apply a blanching pressure for a prescribed time period (e.g., 6 s) while the measured force output is shown on a 1.2 inch tri-color 8 × 8 LED matrix display to ensure that the applied force falls within a prescribed range (2.7–3.3 N). This range was determined by measuring the force applied by an experienced physician. After applying the force, the swing arm is withdrawn to allow the 5MP Raspberry-Pi camera (set to 1024 × 768 pixels at 30 fps) to start recording the refilling of the capillaries. Four white LEDs are used to ensure appropriate illumination, while power is provided by two, series connected, 3.7 V, 20 0 0 mAh lithium polymer cells controlled by a LM7805 linear regulator. To enable the camera to focus on near-field Please cite this article as: C. Visser et al., Investigation of the feasibility of non-invasive optical sensors for the quantitative assessment of dehydration, Medical Engineering and Physics (2017), http://dx.doi.org/10.1016/j.medengphy- ARTICLE IN PRESS JID: JJBE [m5G;July 19, 2017;18:50] C. Visser et al. / Medical Engineering and Physics 000 (2017) 1–7 Table 1 Summary of subject characteristics. Cohort A (n = 4) B (n = 10) Sensors tested Weight at admission (kg) Weight at discharge (kg) Age (months) Gender (% female) Hydration status at admission Mild (<2%) Moderate (2–5%) Severe (>5%) Length of hospital stay (h) ISP 6.6 ± 1.6 7.0 ± 0.8 9.3 ± 3.0 100.0 CRT, SRT, STP 7.8 ± 1.93 8.3 ± 2.1 11.2 ± 6.8 40.0 1 2 1 31.8 ± 14.8 2 4 4 39.4 ± 22.2 CRT, capillary refill time; ISP, infrared spectrometry; SRT, skin recoil time, STP, skin temperature profile. objects, a 10 mm diameter plano-convex lens with a focal length of 30 mm was installed in front of the camera. 2 Skin recoil time (SRT) sensor was developed using a Raspberry-Pi camera (OmniVision OV5647) which was set to record at a resolution of 640 × 480 pixels and 90 fps, as shown in Fig. 1(b). The camera was mounted on a movable head with two degrees of freedom to facilitate operation. The camera’s focal length was also shortened using a plano-convex lens with a focal length of 30 mm and a very narrow field of view and focus range. A directional lighting system was implemented using two LEDs to facilitate correct orientation and distance. The sensor is designed to rest on the patient’s sternum while the physician performs a conventional skin recoil test on the patient’s abdomen. 3 Skin temperature profile (STP) sensor utilized a FLIR-E60 thermal camera (FLIR Systems, Wilsonville, OR, USA) to measure skin surface temperature over a prescribed path along the subject’s body, as shown in Fig. 1(c) and (d). The thermal camera has a sensitivity of 0.05 °C, an accuracy of less than 2% of the measurement value and acquires images with a resolution of 320 × 240 pixels. 4 Infrared spectrometry (ISP) sensor was implemented using two infrared LEDs (Roithner LaserTechnik, LED1300-series and ELD-) which produce light at 1300 nm and 1480 nm as illustrated in Fig. 1(e) and (f). These two wavelengths were selected to ensure that water absorption coefficients with a sufficiently large difference in magnitude were used. The shorter wavelength was chosen to maximize the skin penetration depth of the infrared light, while the longer wavelength was selected because of its sensitivity to water at various concentrations. The LEDs were used in conjunction with an IPD14-12-5T photodiode (Roithner LaserTechnik, Austria) capable of measuring the spectrum from- nm. The LEDs and photodiode were mounted to ensure that the sensing area of the photodiode and the light inducing area of the LEDs lay in the same plane. This ensured that when pressed onto the skin all background light is cutoff from the photodiode and the majority of light measured is captured from the two infrared LEDs. Data are acquired through the use of a current amplifier and the ADC of a microcontroller (Arduino UNO32, Digilent Inc., USA). To ensure the safety of the device the circuitry was encapsulated in silicone rubber (Mold Max® 30, Smooth-on, USA). 3 Ambulatory Admission Unit of Tygerberg Hospital, Western Cape Province, South Africa after obtaining ethical committee approval. Informed consent was obtained from the parents of each infant prior to study enrolment. Subjects were eligible for enrolment if they were admitted with gastrointestinal distress, presented signs of mild to severe dehydration as determined by an attending physician, and were 6–36 months old. In total only 11 patients were enrolled in the study (numbered 0–10) as summarized in Table 1. This low recruitment occurred due to the sparse number of infants admitted with acute dehydration during the ethically approved study timeframe. The SRT, CRT and STP sensors were tested on subjects 1–10 (cohort B) while the ISP sensor was only tested on subjects 0–3 (cohort A). Once enrolled all patient information was recorded along with a “snapshot” of the patient condition (taken every 4–6 h following enrollment until discharge) which comprised: 1 A clinical assessment by a physician in which standard scoring of clinical dehydration markers was performed. Scoring was done on a 3-point scale, according to the attending physician’s best clinical estimate. The clinical assessment included assessment of manual CRT and SRT, eye and fontanelle appearance, mucus membranes dryness, neurological state (i.e., mood, restlessness or lethargy, etc.), tear production and pulse quality. 2 Patient data collection by a nurse and a digital assessment using the sensors. The data collected by the nurse included blood pressure, pulse rate, body temperature and body weight. The digital assessment was performed on the stomach of each subject with the SRT and ISP sensors, on the chest for the CRT sensor and on the face for the STP sensor. 2.3. Data analysis 2.2. Testing procedure The sensor data were analyzed and processed offline using Python and MATLAB® (Mathworks Inc., Natick, Massachusetts USA). The camera-acquired data (from the CRT, SRT and STP sensors) necessitated the application of image processing and optimization techniques in order to extract the desired markers. In the case of the CRT sensor, exaggeration of the blanching site and flow tracking methods (i.e., Shi–Tomasi corner detection and the Lucas– Kanade method [30]) were used to improve the estimation of the CRT from the acquired videos. For the SRT sensor a Python application was developed to scan through the STR videos frame-by-frame to record the recoil time. The thermal images were imported into a Python-based application that allowed manual marking of the path along which the temperature profile should be extracted. A curve was then automatically fitted to the profile and the maximum gradient of the fitted curve was returned as the hydration marker. For the ISP sensor the photocurrent produced by the photodiode receiving the IR light is directly proportional to the measured voltage, which in turn is linearly related to the intensity of the reflected diffused light. Tissue water content is therefore qualitatively inferred from the infrared absorption following the Beer–Lambert law [31], which dictates that the absorption of light is proportionally related to the concentration of the absorbing medium. This follows the approach used by Nachabé et al. [25]. To evaluate the sensors’ performance in assessing dehydration status the quantitative estimates of the CRT, SRT, maximum temperature gradient and reflected diffuse light intensity were correlated with the percentage weight loss computed based on post illness weight change at discharge (i.e., the reference for dehydration severity level). This was computed using the following expression: Following initial hardware validation of the four sensors via in vitro testing and preliminary tests on adults [28,29], an in vivo clinical study on infants was conducted at the Pediatric dehydration level weight at discharge − weight at measurement = ∗ 100 weight at discharge A more detailed description of the design and implementation of the prototype hydration sensors can be found in Kieser [26] and Visser [27]. (1) Please cite this article as: C. Visser et al., Investigation of the feasibility of non-invasive optical sensors for the quantitative assessment of dehydration, Medical Engineering and Physics (2017), http://dx.doi.org/10.1016/j.medengphy- ARTICLE IN PRESS JID: JJBE 4 [m5G;July 19, 2017;18:50] C. Visser et al. / Medical Engineering and Physics 000 (2017) 1–7 Table 2 Summary of the coefficients for the linear regression that was performed on the standardized data for the 11 markers of dehydration (excluding the ISP sensor). Marker Coefficient p-value 95% CI+ 95% CI− Device CRT Device SRT Maximum temperature gradient Age normalized heart rate Age normalized respiratory rate Core body temperature Eye appearance Mucous membrane dryness Fontanelle appearance Presence of tears Neurological status - −- −0.0959 −- - −- −- −0.231 −0.298 −0.026 −0.228 −0229 −0.420 −0.079 - CI, confidence interval; CRT, capillary refill time; ISP, infrared spectrometry; SRT, skin recoil time. The sensor specificity and sensitivity were also calculated for various dehydration threshold levels to produce receiver operating characteristic (ROC) curves, which were compared to the results obtained from the application of three conventional clinical hydration scales (i.e., the CDS, WHO and Gorelick scales). From the ROC curve the optimal specificity and sensitivity were determined (based on equal weighting) and used as an indication of the devices’ feasibility to quantitatively assess dehydration. Additionally, to explore whether a sensor fusion approach would provide better predictive performance than each individual sensor, a linear regression model was developed using the results from the CRT, SRT and STP sensors along with eight other clinical markers of dehydration recorded in the patient snapshots including normalized heart rate, normalized respiratory rate, core body temperature, eye appearance, fontanelle appearance, mucous membrane dryness, presence of tears, and neurological status; with the percentage weight loss due to water loss used as the output of the regression. It should be noted here that the ISP sensor data were omitted from the regression model due to the small test population (four infants). The results obtained from the normalized linear regression analysis on the sensor data and the clinical markers, presented in Table 2, were used to select the predictors for a second regression model based on a sensor fusion approach, which was then implemented and tested. The first regression analysis returned a coefficient for each predictor as well as a corresponding p-value. In general predictors with a large coefficient and small p-value explain more variation and correlate better to the outcome. The parameters which had the largest coefficients and smallest p-values were then included in the second sensor fusion based regression model. When deriving the final regression model, the unstandardized values of the input and output variables were used so the algorithm could be implemented without any scaling or unit conversion. 3. Results 3.1. Clinical evaluation of sensor performance Fig. 2(a) shows the digitally measured CRT plotted as a function of the observed dehydration level as determined by post illness weight gain. From the figure it can be seen that digitally measured CRT did not correlate with the observed hydration level in our study (p = 0.388, R2 = 0.023). The average CRT was 2.00 ± 1.04 s. As can be seen from Fig. 2(b) the digitally measured SRT correlated with the observed hydration level with an R2 = 0.247 (p < 0.001). It is also interesting to note that the average SRT for all the patients was around 0.079 ± 0.038 s. Fig. 2(c) shows that only a weak correlation (p = 0.04, R2 = 0.096) could be observed between the maximum temperature gradient and the observed dehydration. Fig. 2(d) shows the grouped results for the ISP sensor for the wavelengths 1300 nm (‘࢞’) and 1480 nm (‘◦’). Both wavelengths exhibit a positive trend with a significance smaller than 0.05 (p = 0.011 and p = 0.0 02 at 130 0 nm and 1480 nm, respectively) with correlations of R2 = 0.34 and R2 = 0.48, at 1300 nm and 1480 nm respectively. Measurements at 1480 nm have a lower intensity and are better correlated than that at 1300 nm. The results for the normalized linear regression of all the available predictors (excluding the ISP sensor, due to its small population size) are summarized in Table 2. The R2 value for the complete model (based on all 11 markers) is 0.478. The results from the regression analysis in Table 2 were then used to compare the predictors and to select the digital sensors to be included in the sensor fusion regression model: Hydration marker = −1.58 + 60.6 × SRT +6.7 × Maximum temperature gradient (2) Fig. 3(a) shows a scatter-plot of the hydration marker as calculated by the sensor fusion regression model in Eq. (2). With an R2 value of 0.244 and a p-value of 0.0012 it can be seen that the correlation between the sensor fusion marker and the reference weight gain is better than any other input parameter, however, it is only marginally better than that of the SRT sensor by itself. Fig. 3(b) shows the corresponding ROC curves for detecting 5% and 10% dehydration. The areas under the ROC curve for detecting both 5% and 10% dehydration are 0.86 and 0.87, respectively. Fig. 3(c) shows the best sensitivity and specificity combination for each of the sensors as well as the clinical scales (CDS, Gorelick and WHO), for detecting the presence of dehydration greater than or equal to 5%. The sensitivity and specificity of the fusion model is 0.90 and 0.78 respectively, slightly better than for the SRT sensor (0.80 and 0.84 respectively). Additionally, the CRT sensor exhibited a sensitivity and specificity of 0.60 and 0.50 respectively while the STP exhibited 0.90 and 0.50. The specificity and sensitivity achieved by the ISP sensor for cohort A are also shown for each wavelength investigated. The figure shows that a higher overall sensitivity and specificity were obtained at 1480 nm (sensitivity: 1.0, specificity: 0.86) than at 1300 nm (sensitivity: 1.0, specificity: 0.79) and the SRT sensor. The best performing clinical dehydration scale was the Gorelick scale (sensitivity: 0.56, specificity: 0.56). 4. Discussion The results presented in Fig. 2 show the performance of the four hydration sensors obtained from the clinical tests on infants. The poor correlation (R2 = 0.023, p = 0.388) seen in Fig. 2(a) for the CRT sensor can partially be explained by the restlessness of the infants during testing which introduced significant artifacts into the data. This also makes it difficult to compare the performance of the Please cite this article as: C. Visser et al., Investigation of the feasibility of non-invasive optical sensors for the quantitative assessment of dehydration, Medical Engineering and Physics (2017), http://dx.doi.org/10.1016/j.medengphy- JID: JJBE ARTICLE IN PRESS [m5G;July 19, 2017;18:50] C. Visser et al. / Medical Engineering and Physics 000 (2017) 1–7 5 Fig. 2. Raw hydration markers plotted with respect to observed dehydration level as determined by post-illness weight gain for the four hydration sensors: (a) CRT, (b) SRT, (c) STP, (d) ISP. CRT sensor to that reported previously in the literature [9,10]. Fig. 2(b) shows a fairly strong correlation between the SRT sensor readings and observed dehydration (relative to correlations reported for other hydration markers in the literature [32]). It is interesting to note that the recoil times were usually short (trecoil < 0.1 s) relative to conventional manually recorded recoil times which are captured on a scale that only discriminates between instant, less than 2 s and more than 2 s. There are two likely explanations for this phenomenon. First, infants have very elastic skin, especially when not severely dehydrated. Second, it is possible to observe different phases during one skin recoil event, in the frame by frame analysis. In the first distinct phase the skin moves normally with respect to the underlying tissue, in the second phase the skin moves laterally with respect to the underlying tissue. In this study the SRT was defined as the duration of the first phase because it could be reliably captured from the acquired images. However, in conventional manually performed SRT the clinicians typically use both phases when estimating the skin recoil time since it is challenging for a human observer to react in the very short time period corresponding to the first phase of skin recoil. Another possible explanation is that it is easier to perceive the skin motion rather than the shape of the skinfold in such a short time when observing with the naked eye. The positive relationship between maximum temperature gradient and dehydration level in Fig. 2(c) supports the initial hypothesis that higher temperature gradients can be expected for more dehydrated patients, however, the correlation is fairly weak (R2 = 0.096). This may be attributed to confounding factors such as weather and climate control. The ISP data shown in Fig. 2(d) are consistent with theory since the intensity of both wavelengths increases with the onset of dehydration. This is indicative of the absence of chromophores linked to reduced blood flow in the skin due to vasoconstriction in an attempt to maintain homeostasis. The correlation coefficients in Table 2 can be used to compare the predictive value of the various markers to one another (in our study). According to the output from our regression equation the predictors that performed best in the infant study are, in descending order: SRT, heart rate, neurological status and maximum temperature gradient. This finding corresponds favorably with the results reported by Duggan et al. [32] in a study on 135 infants (ages 3–18 months) in which the prognostic value of various clinical indicators of infant hydration was investigated. They used post-illness weight gain as the reference and multiple regression to rank the value of the various predictors. They found SRT (R2 = 0.14, p = <0.001), altered neurologic state (R2 = 0.062, p = 0.002), sunken eyes (R2 = 0.023, p = 0.052) and dry mucous membranes (R2 = 0.017, p = 0.082) to be the clinical signs that best corresponded with dehydration level. The R2 value for their total regression model was 0.244 and the p-value for their regression model was <0.001. This is similar to what was found with the sensor fusion model presented in this study, with a R2 value of 0.244 and p-value of 0.0012. In this sensor fusion model the SRT and STP markers were combined into a single predictor. Fig. 3(a) shows the output of the derived regression model for each measurement case plotted with respect to the reference dehydration level at each time interval. From Fig. 3(a) and (b) one can observe a strong correlation between the output of the fusion model and the reference hydration state. As seen in Fig. 3(c) the combined predictor performed better than either of its input predictors in terms of sensitivity and Please cite this article as: C. Visser et al., Investigation of the feasibility of non-invasive optical sensors for the quantitative assessment of dehydration, Medical Engineering and Physics (2017), http://dx.doi.org/10.1016/j.medengphy- JID: JJBE 6 ARTICLE IN PRESS [m5G;July 19, 2017;18:50] C. Visser et al. / Medical Engineering and Physics 000 (2017) 1–7 Fig. 3. (a) Predicted dehydration from the fusion algorithm. (b) The fusion algorithm ROC curve. (c) Comparison of sensor sensitivity and specificity performance with the CDS, WHO and Gorelick scales. specificity. However, the combined predictor performs only marginally better than the SRT sensor. This can be explained by the following (1) the combined predictor of a linear regression will always have a better or equal correlation to the reference than any of its constituent predictors; and (2) the coefficient of the SRT in the normalized regression model is significantly larger than that of any of the other predictors so the SRT predictor therefore dominates the fusion model. The area under the ROC curves for the fusion model is also slightly higher than that of the SRT sensor for similar reasons. High specificity and sensitivity were found with the ISP sensor at both wavelengths which implies or comparable or better performance than the CDS, Gorelick and WHO scales. This result is not statistically significant due to the small test population of the ISP sensor and therefore cannot be considered conclusive. The ISP results are nevertheless reported to motivate further research in the field. Overall the results presented here seem quite promising. The digital assessment of SRT and ISP in our study outperformed all the individual markers for infant dehydration assessment that were reviewed by Steiner [20]. The STP sensor achieved similar sensitivity and specificity as clinically measured CRT, which was the highest performing marker in the Steiner review (sensitivity: 0.65 specificity: 0.85). Furthermore, the STP–SRT fusion model outperformed the “gold standard” CDS, WHO and Gorelick hydration scales as implemented by the physicians during the study. These clinical hydration scales performed better in the reviews by Pringle [6] and Jauregui [7] in terms of sensitivity and specificity, however, failed to achieve comparable performance to the fusion model in the current study. This suggests that these hydration sensors may show promise for dehydration assessment in resource constrained settings pending further clinical evaluation. Several challenges were encountered during the clinical evaluation of the sensors which may have influenced the quality of the data obtained. In particular, the uncontrolled movement of infants during sensor testing introduced many artifacts into the acquired data. This was mitigated partially by implementing supplementary, quality enriching algorithms and repeated measurements, and by performing measurements when the mother was soothing the infant. Another major challenge was the irregularity of measurements taken by the clinical staff during their rounds. This occurred partly due to high patient loads and consequent patient prioritization, with higher risk patients seen more often than lower risk ones. Finally, the sparse availability of candidates led to a small sample size of infants for evaluation of the sensor performance which makes only a basic assessment of the sensors’ feasibility possible. 5. Conclusion The basic feasibility of assessing infant dehydration using various non-invasive optical sensors based on the quantitative and objective measurement of four markers of dehydration: (i) skin turgor, (ii) capillary refill time, (iii) skin temperature gradient and (iv) infrared absorption as a surrogate measure of tissue water content, was investigated in this study. A fusion algorithm based on two of the sensors (SRT, STP) produced a sensitivity and specificity (0.90, 0.78) greater than that of the Gorelick (0.56, 0.56) and CDS (1.0, 0.2) scale. It also produced a greater sensitivity and comparable Please cite this article as: C. Visser et al., Investigation of the feasibility of non-invasive optical sensors for the quantitative assessment of dehydration, Medical Engineering and Physics (2017), http://dx.doi.org/10.1016/j.medengphy- JID: JJBE ARTICLE IN PRESS [m5G;July 19, 2017;18:50] C. Visser et al. / Medical Engineering and Physics 000 (2017) 1–7 specificity with regards to the WHO scale (0.13, 0.79). Overall, the results from this study suggest that objective and quantitative assessment of infant dehydration may be possible. However, further evaluation of the sensors on a large sample population is needed before deploying them in a clinical setting. Conflict of interest Kiran Dellimore is employed by Philips Research Europe. Ethical approval Ethical approval was obtained from the Stellenbosch University Human Research Ethics Committee (Ref#: S13/10/204— “Quantitative Hydration Sensor Development Infant Testing”). Acknowledgments This study was supported by Philips Research Asia philips funding is S003021 and the Technology and Human Resources for Industry Programme (THRIP) funding is IPR TP-. Gratitude is also expressed to the late Prof. Cornie Scheffer and Dr. Jos van Haaren. References [1] Manz F. Hydration in children. J Am Coll Nutr 2007;26(Suppl 5):562S–569S. [2] Carcillo JA, Tasker RC. Fluid resuscitation of hypovolemic shock: acute medicine’s great triumph for children. Intensive Care Med 2006;32(7):958–61. [3] Gutierrez G, Reines H, Wulf-Gutierrez ME. Clinical review: hemorrhagic shock. Crit Care 2004;8(5):1. [4] Santosham M, Chandran A, Fitzwater S, Fischer-Walker C, Baqui AH, Black R. Progress and barriers for the control of diarrhoeal disease. Lancet 2010;-):63–7. [5] Friedman JN, Goldman RD, Srivastava R, Parkin PC. Development of a clinical dehydration scale for use in children between 1 and 36 months of age. J Pediatr 2004;145(2):201–7. [6] Pringle K, Shah SP, Umulisa I, Mark Munyaneza RB, Dushimiyimana JM, Stegmann K, et al. Comparing the accuracy of the three popular clinical dehydration scales in children with diarrhea. Int J Emerg Med 2011;4(1):1. [7] Jauregui J, Nelson D, Choo E, Stearns B, Levine AC, Liebmann O, et al. External validation and comparison of three pediatric clinical dehydration scales. PLoS One 2014;9(5):e95739. [8] Falszewska A, Dziechciarz P, Szajewska H. The diagnostic accuracy of clinical dehydration scale in identifying dehydration in children with acute gastroenteritis a systematic review. Clin Pediatr 2014;53(12):1181–8. [9] Armstrong LE. Assessing hydration status: the elusive gold standard. J Am Coll Nutr 2007;26(Suppl 5):575S–584S. [10] Popov T. Review: capillary refill time, abnormal skin turgor, and abnormal respiratory pattern are useful signs for detecting dehydration in children. Evidence Based Nurs 2005;8(2):57. 7 [11] Koulmann N, Jimenez C, Regal D, Bolliet P, Launay JC, Savourey G, et al. Use of bioelectrical impedance analysis to estimate body fluid compartments after acute variations of the body hydration level. Med Sci Sports Exerc 20 0 0;32(4):857–64. [12] Röthlingshöfer L, Ulbrich M, Hahne S, Leonhardt S. Monitoring change of body fluid during physical exercise using bioimpedance spectroscopy and finite element simulations. J Electr Bioimpedance 2011;2(1):79–85. [13] Utter AC, McAnulty SR, Riha BF, Pratt BA, Grose JM. The validity of multifrequency bioelectrical impedance measures to detect changes in the hydration status of wrestlers during acute dehydration and rehydration. J Strength Cond Res 2012;26(1):9–15. [14] Saavedra JM, Harris GD, Li S, Finberg L. Capillary refilling (skin turgor) in the assessment of dehydration. Am J Dis Child 1991;145(3):296–8. [15] Blaxter LL, Morris DE, Crowe JA, Henry C, Hill S, Sharkey D, et al. An automated quasi-continuous capillary refill timing device. Physiol Meas 2015;37(1):83. [16] Pickard A, Karlen W, Ansermino JM. Capillary refill time: is it still a useful clinical sign? Anesth Analg 2011;113(1):120–3. [17] Shavit I, Brant R, Nijssen-Jordan C, Galbraith R, Johnson DW. A novel imaging technique to measure capillary-refill time: improving diagnostic accuracy for dehydration in young children with gastroenteritis. Pediatrics 2006;118(6):2402–8. [18] Bordoley A, Irwin R, Kalyani V, McDonald C, Standish D. D1git: automated, temperature-calibrated measurements of capillary refill time. University of Pennsylvania, Philadelphia, USA. Unpublished Results. [19] Kviesis-Kipge E, Curkste E, Spigulis J, Gardovska D. Optical studies of the capillary refill kinetics in fingertips. In: World congress on medical physics and biomedical engineering; 2009. p. 377–9. [20] Steiner MJ, DeWalt DA, Byerley JS. Is this child dehydrated? JAMA 2004;291(22):2746–54. [21] Kviesis-Kipge E, Curkste E, Spigulis J, Eihvalde L. Real-time analysis of skin capillary-refill processes using blue LED. In: SPIE photonics Europe; 2010. p. 771523. [22] Strozik KS, Pieper CH, Roller J. Capillary refilling time in newborn babies: normal values. Arch Dis Child Fetal Neonatal Ed 1997;76(May 3):F193–6. [23] Strozik KS, Pieper CH, Cools F. Capillary refilling time in newborns–optimal pressing time, sites of testing and normal values. Acta Paediatr 1998;87(3):310–12. [24] Attas EM, Sowa MG, Posthumus TB, Schattka BJ, Mantsch HH, Zhang SL. Near-IR spectroscopic imaging for skin hydration: the long and the short of it. Biopolymers 2002;67(2):96–106. [25] Nachabé R, Hendriks BH, Desjardins AE, van der Voort M, van der Mark MB, Sterenborg HJ. Estimation of lipid and water concentrations in scattering media with diffuse optical spectroscopy from 900 to 1600 nm. J Biomed Opt 2010;15(May (3)):037015. [26] Kieser E. Hydration sensor development for use on infants in underserved areas. Stellenbosch University; 2015. [27] Visser C. A micro approach to quantitative dehydration sensor development. Stellenbosch University; 2015. [28] Kieser E, Dellimore K, Scheffer C, Visser C, Smith J. Development of diagnostic sensors for infant dehydration assessment using optical methods. In: Conf Proc IEEE Eng Med Biol Soc; 2015. p. 5537–40. [29] Visser C, Scheffer C, Dellimore K, Kieser E, Smith J. Development of a diagnostic dehydration screening sensor based on infrared spectrometry. In: Conf Proc IEEE Eng Med Biol Soc; 2015. p. 1271–4. [30] Lucas BD, Kanade T. An iterative image registration technique with an application to stereo vision. IJCAI 1981;81(1):674–9. [31] Swinehart DF. The Beer–Lambert law. J Chem Educ 1962;39(7):333. [32] Duggan C, Refat M, Hashem M, Wolff M, Fayad I, Santosham M. How valid are clinical signs of dehydration in infants? J Pediatr Gastroenterol Nutr 1996;22(1):56–61. Please cite this article as: C. Visser et al., Investigation of the feasibility of non-invasive optical sensors for the quantitative assessment of dehydration, Medical Engineering and Physics (2017), http://dx.doi.org/10.1016/j.medengphy-