2.1.

Experimental Setup

The diffused reflectance spectroscopy-based absorbance setup (patent pending, 467/KOL/2009) for monitoring the spectral response of the conjunctiva is represented in Fig. 1(a). A white light source (Model No. LS-450) and a spectrograph (Model No. STS-VIS) with wavelength resolution of 0.47 nm (both are from Ocean Optics, Florida) were used in our study. Lab-grade optical fibers from Ocean Optics were used for the transmission and collection of light to and from the sample (conjunctiva). The light from the source is transmitted through the six surrounding fibers [Fig. 1(a), excitation fiber] and is incident on the conjunctiva while the single fiber, in the middle of the probe [Fig. 1(a), detection fiber], collects the diffused light and sends it back to the spectrograph. The corresponding spectral response as generated in the spectrograph is then transferred to a laptop computer through a USB interface where it is processed in our developed software. The wavelength calibration of our setup has been established with a He-Ne laser (632.8 nm), fluorescent lamp, and emission/absorption of a number of dyes including aqueous bilirubin solution, as shown in Fig. 1(b).^17,18 The comparative spectral response of a normal volunteer and a jaundice patient is represented in Fig. 1(b). A distinct difference in their spectral appearance is visible; the contribution of yellow pigment deposited in the conjunctiva of the jaundice patients is higher compared to the normal volunteer.

Fig. 1

(a) Schematic representation of our working device. The light from the source is transmitted through the six excitation fibers of the excitation arm and incident on the subject (conjunctiva). The diffused light is collected by the detection fiber and transmitted through detection arm to the spectrograph. The spectral response corresponding to the conjunctiva is processed and generated in the laptop computer. (b) The comparative spectral response of conjunctiva of a normal volunteer and jaundice patient has been represented. An absorption spectrum of aqueous bilirubin solution is also included as reference.

2.2.

Data Collection

A total of 90 patients at the pathology section for liver function test in the Calcutta Medical Research Institute (CMRI) hospital, Kolkata, were recruited in our study. Data were collected in two stages: first, for calibration of the device; second, for measuring the precision of the software-driven device in contrast to the standard biochemical method. Soon after the blood sample collection, the volunteers were taken for the bilirubin assessment using our setup with a 5-min time window. Due to the noninvasive and noncontact nature of the test, there is no need for disinfecting the measuring probe. Approval of the local medical ethical committee (Ref: IEC/07/2014/APRV/23) and informed consent from the patients’ legally authorized representatives were obtained. Blood samples were taken only for clinical reasons and were obtained by professional technicians from CMRI hospital. A wide variety of age group of the recruited patients with a mean age of 45 years [standard deviation (SD) 14 years] with different skin tones were the subjects of the present study.

2.3.

Software Design

The optomechanical components have been connected to a laptop computer using a USB interface. The spectrometer (STS-VIS), which is the active detector in this setup, has been programmed on LabVIEW platform and can be modified for user defined data acquisition. The online display of the acquired data has been used to analyze the data quality and assess the medical condition of the patients. Finally, the bilirubin level of the patient is displayed with suggested medical attention on the monitor of the DAQ laptop computer. The software for automatic data acquisition has been designed in LabVIEW platform. Figure 2 shows the sequential program flow or the algorithm of the developed software. The instrument is first reinitialized to its power on status to remove any previous custom settings. The software then sets the proper integration time for data acquisition to build up the right signal-to-noise (S/N) ratio of the acquired data. This can either be set manually or automatically as decided by the software using an iterative algorithm. For a particular distance between the probe and the reference surface, the software adjusts the integration time using the mentioned iterative algorithm until the peak count reaches the maximum allowed value (here 14,000). The information is acquired through the raw socket of the USB port and the size of the array is determined, thus the wavelength array is calculated on the basis of instrument specifications. The “dark spectrum” and “reference spectrum,” which can either be preacquired or can be determined in-situ, are then loaded for spectrum processing. The software now acquires data, produces the processed spectrum, generates an online graph, and displays the appropriate bilirubin value. The bilirubin value is calculated using the calibration equation (see Sec. 3.1). The data safety level of the patient is determined by the differential absorption values of wavelengths 460 and 600 nm. The online display also suggests the condition of the patient being within or above the safety limits. The information is stored and a comprehensive medical report is generated for offline use for medical practitioners and patients. Complete care has been taken for the software to be simple on the front panel for ease of operation even with nonscientific personnel having no or minimal medical or instrumentation knowledge.

Fig. 2

The flow chart of the software designed in LABVIEW platform for noncontact online monitoring of bilirubin level in humans (see text for details).

3.1.

Calibration and Statistical Analysis

The stored data (stage I, $n=60$) were then processed to find the correlation between the TSB levels of the volunteers with the spectral information obtained from the conjunctiva of the eye. It has already been reported that the spectral contribution near the 460-nm wavelength is due to bilirubin, the yellow pigment.^18–20 Different characteristic wavelengths over the conjunctival spectrum were selected for assessment, but it was found that the differential absorbance of 460 nm ($a_1$) to 600 nm ($a_4$) and ratiometric values of 470 nm ($a_2$) to 576 nm ($a_3$) were more consistent with the TSB level. The differential absorbance of 460 to 600 nm ($a_1$ to $a_4$) was chosen as the index value ($x_i$) to calibrate the setup with the TSB level. The dependency of the index value ($x_i$) with TSB level is represented in Fig. 3(a). The correlation coefficient ($r$) is found to be 0.84; $P<0.0001$, which shows a significant relationship between the two methods (TSB and $x_i$). Further calibration was done in order to achieve a nearly perfect relationship. The $x_i$ value is found to follow a second order polynomial equation $y_i=74.67x_i^2-7.686x_i+0.748$ (calibration equation), where $y_i$ is the individual TSB level. We used this calibration equation to calculate the bilirubin level from the spectral information ($x_i$) obtained by our device. This modification greatly improved the correlation to almost perfect (correlation coefficient, $r=0.96$; $P<0.0001$). The corresponding linear regression curve is represented in Fig. 3(b) with Pearson correlation coefficient, $r=0.96$; $P<0.0001$ and $F=627.1$; slope 0.932; $y$ intercept 0.118.

Fig. 3

Stage I, calibration: (a) The dependency of the total serum bilirubin (TSB) value from blood test with the index value from our instrument ($n=60$) has been represented graphically. The correlation between them is found to follow a second order polynomial equation with an $R^2$ value of 0.89. (b) Correlation between the TSB value from blood test and from our instrument [correlation factor $(r)=0.96$] with the 95% confidence limits and the 95% prediction interval have been represented.

In order to find the statistical significance of the noncontact optical device for online assessment of the bilirubin level, correlation and regression analyses were used.^21–23 We have also used the Bland-Altman method for assessing the agreement between the conventional biochemical technique and our noncontact optical device.²⁴ Two crucial factors decide whether a new method can be used interchangeably with an already established method: the amount of agreement between the methods and its clinical evaluation. We compared our proposed noninvasive bilirubin detection method to an established biochemical method using the approach described by Bland and Altman^24,25 in order to assess the statistical agreement. Thirty patients (stage II, $n=30$) of all age groups were included in our study. Linear regression analysis and Bland-Altman plots are shown in Fig. 4. Data obtained from the linear regression analysis [Fig. 4(a)] show that the two methods show strong correlation as the Pearson correlation coefficient, $r=0.99$; $P<0.0001$ and $F=1588$; slope 1.026; $y$ intercept 0.018.

Fig. 4

Stage II, statistical significance: (a) the linear regression plot of the TSB level measured in both the ways and (b) Bland-Altman analysis: difference against mean for the TSB data (see text).

For adequate comparison of the two methods, the difference in measurement of the two methods is plotted against their average [Fig. 4(b)]. The mean difference between the two methods is depicted as a horizontal line and is rated as bias. The other two horizontal lines ($\mathrm{Mean}\pm 2\mathrm{SD}$) represent limits of agreement which explains that 95% of the differences were assumed to lie within these limits. The results exhibit reasonable agreement between our proposed method and the conventional pathological method of bilirubin detection. The difference in the two methods (conventional-proposed) has mean value of $-0.06\mathrm{mg}/\mathrm{dL}$ and SD value of 0.182. The limits of agreement are from $-0.42$ to $0.30\mathrm{mg}/\mathrm{dL}$. Hence, it can be inferred that for 95% of individuals, a measurement by our method would be between 0.42 units less and 0.30 units greater than a measurement by the conventional method. This small difference has no serious clinical significance in the diagnosis of jaundice. The mean value of the differences indicates a small bias of approximately $-0.06\mathrm{mg}/\mathrm{dL}$. The 95% confidence interval (CI) for the bias represented in Fig. 4(b) is $-0.12$ to 0.00. As the CI includes 0.00, the bias is statistically nonsignificant.²⁶ The negative bias along with CI indicates that the predominant tendency of our instrument is to overestimate the bilirubin levels, so dangerous clinical errors are unlikely to occur. In addition, the coefficient of variation (CV) between our method and the conventional biochemical method was found to be 1.81%, which is comparable to the CV range of 0.35% to 1.96% for laboratory chemical analyzers in repeatability studies.²⁷ This clearly states the bias to be nonsignificant in clinical diagnosis.

3.2.

Reproducibility

In order to establish the potential of the device in terms of reproducibility, 20 patients were repetitively examined by our device. We found excellent precision between the bilirubin levels detected from the same subject by two independent observers. Mean and SD were almost the same in both observations and the intraclass correlation values were highly significant ($r=0.98$; $P<0.0001$). Linear regression analysis also illustrates the accuracy of the two measurements ($F=557.8$; slope 1.04; $y$ intercept $-0.06$) [Fig. 5(a)]. Furthermore, the Bland-Altman plot of the two successive measurements by two different observers is represented in Fig. 5(b) (mean $0.01\mathrm{mg}/\mathrm{dL}$ and SD 0.18). The bias should be zero for an ideal instrument.²⁴ However, in our case, the bias is $0.01\mathrm{mg}/\mathrm{dL}$ and the CV in between repetitive measurements is 0.79%, which have insignificant contributions for clinical diagnosis.

Fig. 5

Stage II, reproducibility: (a) the linear regression plot of the TSB level measured successively by two different observers and (b) Bland-Altman analysis: reproducibility in measuring the TSB level by the device (see text).