Introduction
Non-communicable diseases represent the leading cause of death worldwide. Due to its global rise in incidence and prevalence, type 2 diabetes (T2D) is considered among the top deadliest diseases, accounting for 1.6 million deaths annually [
1]. According to the last report of the Global Burden of Diseases, Injuries, and Risk Factors Study, high plasma glucose belongs to the top three risk factors with the largest increase in the world during the last decade [
2]. T2D itself had been considered the greatest pandemic in human history [
3].
According to the International Diabetes Federation, the latest global diabetes prevalence (2019) is estimated to be 9.3%, accounting for 463 million people [
4]. Nevertheless, it has been argued that such figures underestimate the real number of diabetes prevalence, by at least 25% [
3]. Importantly, the underdiagnosis of T2D in low- and middle-income countries, where the resources to perform a T2D screening are limited, could be as high as 46% [
5]. In order to cope with this public health problem, non-invasive techniques to determine the quantity of glucose and glycated hemoglobin (HbA1c) have been proposed recently, such as NIR spectroscopy, Raman spectroscopy, surface-enhanced Raman spectroscopy, and mid-infrared spectroscopy, among others [
6‐
10].
Raman spectroscopy is an optical technique commonly used [
9,
11‐
18]; its instrumentality relies on the interaction of electromagnetic radiation with matter, including biomolecules such as keratin, lipids, myoglobin hemoglobin, or glucose [
19,
20]. As a result of this interaction, part of the incident light is scattered, and in most cases, the wavelength of the scattered photons remains constant; this is called Rayleigh scattering. However, a small part of the light is scattered at a different wavelengths concerning the incident wavelength due to the gained or lost energy after the interaction, which is called Raman scattering [
21,
22]. As a result, a specific spectral signature of the analyzed molecule is obtained [
23].
Raman spectral signature potentially identifies metabolites of clinical importance for T2D diagnostic [
24,
25]. The gold standard is the HbA1c test, in addition to constant measurements of glucose in diagnosed patients with T2D [
26]. Nevertheless, Raman spectroscopy has several types of signal noise, such as shot noise, fluorescence, readout noise, external source noises, and instrumentation-derived noise [
27‐
30]. In order to reduce these signal noises, many techniques have been developed, for instance, Savitzky-Golay filter [
31], wavelet transformation [
32,
33], polynomial curve fit [
34], baseline correction [
35], empirical mode decomposition [
36], the Vancouver Raman algorithm [
36], and the Zernike polynomial fitting [
37], among others.
Furthermore, artificial neural networks (ANN) have been proposed as suitable techniques for Raman spectra analysis [
17,
38,
39]. There are different ANN architectures; for instance, the feed-forward neural network (FFNN) comprehends different neuron layers, on which the output from the neuron in the level
k is connected to the input neuron in the level
k + 1. The output of the network corresponds to the values of the neurons in the output layers [
40]. The FFNN has been used both as a classification method and as a function approximation on Raman spectra analyses [
41].
In vivo studies have explored the potential use of Raman spectroscopy for the quantification of T2D diagnosis biomarkers. For instance, the classification of 86 individuals as free from T2D, controlled T2D, and non-controlled T2D has been reported. In those analyses, the information obtained by Raman spectroscopy was analyzed using principal component analysis and support vector machines (SVM), showing > 90% of specificity and sensibility [
16]. Furthermore, the use of ANN and SVM to discriminate between normoglycemia and hyperglycemia through the Raman spectra has been reported in a different population (eleven individuals), achieving 88.9 to 90.9% of specificity and sensibility [
17].
In vivo quantification of circulating plasma glucose concentrations using Raman spectroscopy directly over the skin of the individuals, using the fingertip, has shown promising results. The calculated concentrations using linear regression were reported to be highly correlated with capillary glucose measurement, getting a correlation coefficient of 0.80 (
p < 0.0001) in 49 individuals [
42]. Raman spectroscopy readings from the forearms analyzed with partial least-squares (PLS) regression have also shown promising results (mean absolute error (MAE) 7.8%, (
N = 17)) [
43]. Furthermore, PLS and Raman spectroscopy were used to predict glucose concentration in the forearm of 111 individuals obtaining a correlation coefficient of 0.83 in independent Raman predictions for the full cohort [
44]. In addition, critical-depth Raman spectroscopy and PLS were used to quantify circulating glucose in 35 individuals for a period of 60 days, obtaining a mean average relative difference (MARD) of 25.8% with 93% of predictions in the areas A and B of the Clarke error grid, in the independent validations [
45]. Recently, it has been reported an improvement of the in vivo quantification of glucose with Raman spectroscopy, in which linear regression and PLS were applied to analyze the Raman spectra of pigs’ ears. The Raman readings showed almost perfect agreement with the gold standard, 0.94 correlation coefficient in intra-subject analyses [
46].
Even though Raman spectroscopy and machine learning methods have been used for T2D-related biomarkers’ quantification, the in vivo quantification of glucose needs to be improved. Moreover, the in vivo quantification without blood extraction of HbA1c remains unexplored; being this the gold standard for diabetes detection, finding new methods for HbA1c quantification is relevant due to that HbA1c is not only a biomarker to evaluate the glucose control, but also a diagnostic one [
47]. Therefore, the present study investigates whether Raman spectroscopy coupled with feature selection methods and FFNN is suitable for the non-invasive quantification of HbA1c and glucose in people with and without T2D diagnosis.
Discussion
Currently, the global burden associated with T2D is estimated to be 67.9 million disability-adjusted life-years (DALYs); the latest projections point towards an increment of 11.4, resulting in 79.3 million by 2025 [
80]. Furthermore, underdiagnosis of T2D remains as a key problem in low- and middle-income countries [
5]. It is proposed that the development of low-cost and non-invasive methods could alleviate this problem. The development of non-invasive methods could potentially alleviate the increasing environmental footprint of health care associated to diagnostic methods that requires many contaminants, and single-use tests [
81,
82]. Here, we showed that the implementation of FFNN for the analysis of the non-invasive quantification of HbA1c and glucose by means of Raman spectroscopy enhances its ability to identify subjects with T2D.
The Raman spectra in vivo measurements have been used for the determination of T2D diagnosis biomarkers in order to discriminate among T2D and healthy patients [
16,
17]. Also, non-invasive measurements to quantify glucose have used different regression methods [
42,
46]. In this work, the Raman spectra from three different parts of the body in 46 individuals were acquired, from which a spectral analysis was made to identify the representative peaks of glucose and HbA1c, since they have been reported in blood samples [
15,
83], lyophilized HbA1c, and different concentrations [
79], even in vivo measurements [
16,
17].
Moreover, FFNN was implemented to quantify HbA1c and glucose concentrations using the Raman spectra. However, feature selection methods and SOM networks were required to improve the results due to the low intensity and noisy signal in the Raman spectra (supplementary material, Tables
S1–
S4). Remarkably, the Hba1c and glucose Raman spectra obtained in the wrist and forearm performed better than the fingertip. This could be explained by the collagen-enriched tissue in the fingertips. Given the fact that the collagen has its own Raman spectra [
84], these spectra may interfere in the quantification of glucose and Hba1c. In addition, the epidermis layer is thicker than the forearm and wrist and it might have a high variation among individuals [
85,
86].
In order to overcome the above-mentioned technical limitations of the spectroscopy, we proposed the use of FFNN. It is important to notice that the FFNN can approximate any fitting of a data set representing a relationship [
87]. Considering this, FFNN was implemented to Raman spectra without neither data selection nor enhancing data; however, the in vivo Raman spectra are noisy data, and the prediction error was high (1.03 ± 0.09%, and 60.32 ± 5.27 mg/dL for HbA1c y glucose, respectively). In order to reduce the error in the predictions, an early selection of valuable features should be conducted. Therefore, we decided to use feature selection and extraction methods, such as the SOM network and RReliefF, which have not been implemented in Raman spectra.
SOM is an unsupervised classification algorithm [
57] that allows us to know the distribution of the data in function of intrinsic features of the data and generate prototypes based on the HbA1c and glucose concentrations, which helps us to have a better representation of the signal (reducing the RMSE-CV). Despite that, this representation has the same number of features as the input signal has. It should be noted that the fewer features the problem has, the shorter the FFNN execution time; this is suitable for an immediate result and a future embedded application. For this reason, we use feature selection methods, as with RReliefF, among others (see supplementary material), and the number of features that obtained the best result per each part of the body is presented in Table
2.
Besides, a comparison between FFNN, SVM, and LR was made using the best-case per body region. The used features for HbA1c were SOM network in the spectral interval from 200 to 1800 cm−1 for the forearm, and 512 selected features by RReliefF-SOM for the wrist. It is worth mentioning that these characteristics do not correspond to the spectral region from 600 to 1600 cm−1. And for the index finger, 50 features were selected by RReliefF-SOM. For the glucose case, the used features were 200, 150, and 150 using RReliefF and SOM for the forearm, wrist, and index finger, respectively.
Results shown in Table
3 depict that the error obtained using SVM and LR is higher than that achieved by the FFNN (0.69% ± 0.07%), and comparing SVM and LR, the first one has better performance; this may be since different kernels were used to perform the regression, that means, kernels are not necessarily linear. Thereby, linear regressions may not present a good performance in HbA1c quantification. In the glucose case, the results from the forearm region for LR and SVM were very close to the FFNN results. However, our proposal is still better with RMSE-CV and a standard deviation of 30.12 ± 0.53 mg/dL (see Table
2).
Former studies in which Raman spectroscopy had been used to estimate the glucose values have evaluated the specificity and sensitivity in relation to the binary classification (healthy individuals vs. controlled T2D), reporting both specificity and sensitivity of 100% [
16]. In our case, these parameters were calculated in multiclass classification, and the best results were 94.44% sensitivity and 99.73% specificity in the healthy class. On the other hand, previous studies had used FFNN to classify subjects in two categories (healthy and T2D) [
17]. Their best result was 96% of accuracy and sensitivity and specificity values of 88.9% and 90.9%, respectively. An accuracy of 96.01% was similarly obtained implementing our methodology, despite the difference is not significant (
P > 0.05,
P = 0.31 using the non-parametric method Kruskal Wallis); an improvement is observed in the percentages of sensitivity and specificity; although it should be considered that this study [
17] was obtained through a binary classification, while the present work is a multiclass classification. Hence, our proposed methodology outperforms the already published state-of-the-art methods.
Sensitivity and specificity were calculated from the predicted HbA1c percentages by the FFNN. The results are shown in Table
7, in which the percentage metrics were 100–93.10% for healthy, 60.71–87.97% for prediabetes, and 87.50–80.70% for T2D, which means a decrease for prediabetes and T2D groups; this is due to the error obtained in the regression model. Concerning the works reported in the literature [
76,
88], there have been reported sensitivity and specificity for the A1c commercial test (boronated affinity high-performance liquid chromatography-HPLC) with values for prediabetes in a range of 84 to 95% and 86 to 93%, respectively, and for T2D at around 45 and 99% [
72] and also 44 and 79%, respectively [
88]. That means that even invasive commercial tests are imperfect and may present low sensitivity. Moreover, this is the first approximation to the development of a painless method since no work has been reported nowadays in order to obtain an in vivo quantification of HbA1c by non-invasive techniques such as Raman spectroscopy.
Although several investigations have been made in order to achieve reliable quantification of HbA1c, a combination like the one presented in this work had not been reported so far, which consists of different concentrations in a population (46 individuals) of multiple individuals with 36 different concentrations of HbA1c and 43 concentrations of glucose, as well as the combination of feature selection methods and artificial neural networks. Our proposal obtained a RMSE-CV carried out from the Raman measurements taken on the wrist in a range of 5.2–14% of HbA1c of 0.69%. This is the first work implementing non-invasive measures to quantify HbA1c in humans.
Regarding the glucose measurements, we obtained a RMSE-CV and standard deviation of 30.12 ± 0.53 mg/dL in the forearm region, and in the percentage of success in Clarke error grid for zone A 82.61%, zone B 15.22%, zone D 2.17%, and zones C and E 0%, the glucose values varied from 56 to 400 mg/dL. Concerning the reports in the state-of-art, percentages in Clarke error grid have been reported with the following results: 78.4% of success in zone A using partial least square (PLS) from the forearm of 111 individuals [
44]; 72% in zone A measured to the middle finger of 29 individuals and applied linear regression [
42]; 93% zones A and B by an intra-subject analysis into 35 individuals and implemented critical-depth Raman spectroscopy and PLS [
45]; another work reports no incidences in zone D; however, percentages were not reported, and their study was intra-subject [
43].
Publisher's note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.