The study and methods used were approved by the Moscow City Ethics Committee on the basis of the Regulations on the City Ethics Committee and in accordance with the standards of operating procedures (Protocol No. 68 dated July 4, 2022). The methods were carried out in accordance with relevant guidelines and regulations. All patients gave their informed consent to participate in the study.
Cohort used for the training and evaluation of models predicting low Hb level
To train and evaluate models predicting the risk of low haemoglobin, we collected diffuse reflectance spectra of the forearm skin and fingernails and RGB-imaging data of fingernails for 240 volunteers undergoing examination in the emergency department of the 67th City Clinical Hospital (Moscow, Russia). The investigated cohort sample was balanced by sex – there were 117 female and 123 male patients. The average age was 56 ± 20 y.o. spanning from 18 to 95 y.o.
In parallel with optical measurements, invasive determination of the Hb level was carried out according to standard clinical protocols for all patients [34,35]. Venous blood collected from the antecubital vein using the Vacutainer was characterized using the CBC analysis in a certified (ISO 15189:2012) clinical diagnostic laboratory. The measurements were performed 10-15 minutes after collecting optical data and further were used as a ground truth for low Hb prediction.
The invasively determined Hb level for female volunteers was 120 ± 21 g/L, while the average level for male patients was 137 ± 27 g/L. Patients were split into the group with low Hb (< 120 g/l) with 70 volunteers among whom 52 were female, respectively, and the group with high Hb (≥ 120 g/l) with 170 volunteers, 65 of whom were female.
Independent cohort used to evaluate the risk of low Hb level
For additional verification of models trained using the acquired optical response, we also collected a database of DRS data from forearm skin among students and employees of the Lomonosov Moscow State University without additional blood sampling – only non-invasive measurements were made using the optical method. In total, statistics were collected on 274 patients (161 female, 113 male) with a mean age of 31 ± 10 y.o. spanning from 18 to 70 y.o.
Diffuse reflectance spectra acquisition and preprocessing
Measurements using DRS were made on two areas of the tissue – the inner side of the forearm and the nail plate (Fig. 1A). For DRS measurements the custom-built fiber probe was used. The fiber probe consisted of transmitting and receiving fused silica fibers (core diameter of 600 µm), located at a distance of 2.5 mm at the measuring ends. The transmitting fiber was connected to the broadband illumination source (halogen lamp with continuous spectrum in the 450–2500 nm range, 20 W), while the collecting fiber was connected to the Ocean Optics Maya 2000 spectrometer (Ocean Optics, USA) detecting light response in the 200–1100 nm range with a spectral resolution of ~3 nm. Distance between the measuring ends of the dual-fiber probe was chosen so that the intensity of diffusely reflected light in the spectral region corresponding to the Hb absorption (500-600 nm) was sufficiently high, while the signal was collected from a depth which approximately corresponds to the skin dermis, where small and large blood vessels are located.
Before measuring diffuse reflectance spectra from the skin, a calibration procedure was carried out. Upon calibration the background signal for each wavelength acquired by the spectrometer in the absence of any external illumination was measured, after that the intensity of the diffuse reflectance signal from Spectralon (Labsphere, USA) white standard was measured with the probe located at a distance of 5 mm from the standard. Then the diffuse reflection intensity from the skin was measured and used to calculated diffuse reflectance coefficient at given wavelength using the formula (1):
Then, the diffuse reflectance spectrum was used to calculate the effective optical density spectrum according to the equation (2):
For a more robust assessment of diffuse reflectance spectra, a series of 10-50 diffuse reflectance spectra were measured for each patient from each tissue site, after that the median value of the OD at each wavelength was calculated using spectra of raw spectral density. The acquired median spectra of optical density were then used to calculate optical descriptors. A detailed description of the descriptors calculation is presented below in Results section.
Monte-Carlo modelling of light transport in tissues
To estimate the light probing depth by the used dual-fiber probe, we simulated the propagation of light in a medium imitating the optical properties of skin at a wavelength of 570 nm, using two fibers with a numerical aperture of 0.22 and a refractive index 1.47, located at a distance of 2.5 mm, as a source and detector.
The skin was modelled as a 3 mm thick three-layer environment consisting of epidermis, dermis and hypodermis with a length of 6 mm. The thickness of the epidermis was 100 μm, the dermis was 2 mm, and the hypodermis modelled the remaining part of the environment. The absorption and scattering coefficients of the media were taken from the Jacques et al. [36] to match the optical properties of epidermis and dermis of the I-II skin type according to Fitzpatrick. Namely, it was assumed that the absorption of the epidermis is determined by the presence of melanin in it (volume concentration 2.5%) so that the absorption coefficient in the epidermis was 10.5 cm-1, the scattering coefficient was 540 cm-1, the absorption coefficient in the dermis was determined by the presence of blood in the dermis with volume content of 0.2% and saturation of 0.67, so that the absorption coefficient was 0.5 cm-1 and the scattering coefficient was 375 cm-1. The absorption and scattering coefficients of the hypodermis were 0.05 cm-1 400 cm-1, correspondingly. Scattering was modelled with the Henyey-Greenstein scattering function with anisotropy factor of g = 0.9 for all three layers. The simulation was carried out in using PyXOpto python package [37], the number of photon packages used to estimate probing depth was equal to 5·106.
RGB-imaging of fingernails
Along with the measurements of the diffuse reflectance spectra, we collected RGB-images of fingernails of participated volunteers under standard illumination conditions. The setup consisted of a ~40×40×20 cm box made of aluminium composite material, inside which a Logitech C615 USB-camera and white LED illuminator providing color temperature of ~6000 K were mounted. The box had a rectangular slot with the size of ~15x10 cm, so the patient could put his hand. After that, the image was taken under standard conditions with the exposure time fixed for all patients.
In the obtained images, the regions corresponding to the fingernails and skin of the index, middle and ring fingers were segmented. To account for possible inhomogeneities, we used only the inner region of the segmented area, selecting the sub image of [0.25w,0.75w]× [0.25h,0.75h], where w and h are the width and height of the initial segmented region. In these selected regions we calculated 5-, 15-, 25-, 50-, 75-, 85-, and 95th percentiles of the intensity distribution for each color channel of the image. Then, the average percentile values were calculated for all fingernails and skin areas for each patient and used as features to build a predictive model.
Evaluated classification models
We assessed quality of two relatively simple models – the logistic regression and the random forest – in the task of classification of patients with low and high Hb level in blood using optical descriptors obtained via DRS and RGB-imaging.
As the first model, a logistic regression (LR) with L2-regularization was used to assess the possibility of classifying low and high Hb patients with a linear classifier. Before training a LR using certain subset of features, the features were normalized by subtracting the median value of the feature and dividing by the interquartile range estimated using the training set. We additionally tuned the hyperparameter responsible for the contribution of the regularization term to the LR-loss function using classification score on the validation splits. The specified parameter varied in the 10-4–104 range with a logarithmic step.
As the second model, the random forest was used [38]. This model, on the one hand, makes it possible to take into account the nonlinearity of the separating surface in the feature space, on the other hand, it usually demonstrates good performance over a wide range of selected hyperparameters. In our experiments we used a random forest with the number of decision trees equal to 150, with additional tuning of the maximum depth tree on the small subset of values – {3,5,10,15}. The input features were not normalized before being fed into the model.
Predictive models evaluation
Hyperparameters tuning and model evaluation was carried out using the random splitting method: the sample was divided into a training part, on which training was carried out, and a validation part, on which the classification quality was calculated. The validation part was chosen to be balanced, so that the split included the same number of objects of the positive class (objects with low Hb) and the negative class (high Hb). The size of the validation sample was 15% of the total sample size. The ROC-AUC was chosen as a metric assessing the quality of classification.
Input features
We evaluated which sets of descriptors best classify people with low and high blood Hb. For the DRS data we explored the use of two sets of descriptors – features manually engineered from the spectra of the effective optical density as described in Results, “Features engineering from the experimental data” Section , and features automatically extracted from the absorbance spectra using the principal component analysis. In the latter approach, the selection of principal components was carried out using only the training set, while the calculation of amplitudes for the validation set was based on the decomposition obtained from the training set. The number of components was equal to 10 to approximately coincide with the number of features extracted manually for the DRS data. For the RGB-imaging method, we used the percentiles of image intensities obtained for individual color channels as features, extracted as described in “RGB-imaging of fingernails” Section.
The input features were calculated separately for the fingernails’ and skin areas. This allowed us to compare the quality of classification of patients with low and high Hb according to optical indicators obtained for different tissue sites. To do such comparison, we trained independent models using features computed for the skin, for the fingernails, and combined descriptors computed from skin and nail-plate regions. The combination of features was carried out by simply concatenating features of a certain modality into one feature vector.
Construction of a meta-model combining RGB-imaging and diffuse reflection data
To build a meta-model that combines predictions obtained from RGB-imaging data and diffuse reflectance data, we proceeded using following strategy. We selected the best models trained using DRS and RGB-imaging data: in the case of the “DRS” model, this model was a random forest trained using diffuse reflection data from areas of the skin of the forearm, and in the case of “RGB-imaging” model, it was a logistic regression, which simultaneously uses data from the skin and fingernails areas. Using each of the classifiers, we predicted the probability of each object belonging to the low Hb group using leave-one-out cross-validation: i.e. for each object in the sample, we trained a separate model, and then for the object not participating in training, we predicted the corresponding probability. After this, we trained a simple logistic regression that took as features the probabilities predicted from both “DRS” and “RGB-imaging” models, and evaluated its classification score using standard evaluation procedure described in “Predictive models evaluation” Section.
Assessing the risk of low Hb in an independent cohort
Using diffuse reflectance spectra of the forearm skin collected from volunteers of the independent cohort described in “Independent cohort used to evaluate the risk of low Hb level” Section, which did not participate in the training and evaluation of model quality, we predicted the risk of low Hb level as the probability of belonging to the low Hb class according to the trained model.
Statistical analysis
All data preprocessing, visualization of results, training and validation of models were carried out using custom-written Python 3.9 scripts using the Pandas, NumPy, Matplotlib, Scikit-learn, Scikit-image, Seaborn libraries.