A decision support system for type 1 diabetes mellitus diagnostics based on dual channel analysis of red blood cell membrane fluidity
Introduction
Type 1 diabetes mellitus (T1DM) is a chronic disease in which the pancreas produces little or no insulin, a hormone needed to allow glucose to enter cells to produce energy. Despite active research, type 1 diabetes has no cure, but it can be managed through daily insulin administrations [44]. Untreated, or even treated by excessive insulin, T1DM can cause many long term complications, including heart disease, stroke, kidney failure, foot ulcers, nerve failure and damage to the eyes [1], [7]. The glycosylated hemoglobin (HbA1c) is currently endorsed in many countries as a diagnostic test for diabetes as well as for disease monitoring. Since the fraction of glycated hemoglobin increases with the average amount of plasma glucose, HbA1c is a marker for average blood glucose level over the three months before the analysis. Although widely used, HbA1c determination cause over- or underestimation of the amount of glycosylated hemoglobin, it does not adequately discriminate between T1DM and healthy populations in the absence of overt hyperglycemia, and cannot be used with enough accuracy for early detection of diabetes complications [15], [31].
To improve the discriminative power between healthy subjects, T1DM patients, and T1DM patients with complications, in the last two decades decision support systems (DSSs) have been developed [13]. Indeed, the wide availability of analysis has offered plenty of data useful for diagnostics [6]. For instance, in [39] the authors designed a DSS that, using clinical features, detects if a patient shows signs of diabetes according to WHO criteria by means of the principal component analysis. Still using clinical attributes, other authors applied the AdaBoost algorithm to the same aim [42], whereas there are also attempts leveraging on a fuzzy approach [6]. Whilst most of the efforts have been directed towards DSSs employing clinical features and detecting signs of diabetes, the recent availability of fluorescence imaging-based diagnostic strategies has opened the chance to develop DSSs for T1DM diagnostics, trying to overcome the low sensitivity of traditional methods. Besides common immunohistochemical and morphological classifications, metabolic functional imaging method can detect changes in metabolism, regional chemical composition, and phase state of cells [25], [27], [28]. For T1DM diagnostics, the measurement of fluidity alterations can furnish a more sensitive index of disease progression than other detection methods [26]. Membrane fluidity is a quantitative measurement of the order of the lipid packing within the membranes. Uncontrolled glycemic fluctuations and oxidative stress commonly found in T1DM patients can impair membrane fluidity of blood cells, especially red blood cells (RBCs), and contribute to the development of T1DM complications [37]. RBC membrane fluidity can be investigated through Laurdan two-photon fluorescence microscopy of a liquid biopsy [35]: Laurdan is a lipophilic fluorescent probe whose emission spectrum undergoes to a shift towards longer wavelengths upon membrane fluidification. Through a dual channel acquisition of fluorescence a measure of fluidity can be obtained [29].
On these grounds, in this manuscript we present a DSS recognizing healthy subjects and T1DM patients with and without complications based on dual channel analysis of red blood cell membrane fluidity. Although preliminary results were reported in [9], this paper extends previous contribution in some respects. First, in [9] we used images of RBCs and peripheral mononuclear cells, whereas here we leverage only on RBC images to reduce the acquisition burden and to lower costs. Second, previous work studied how a classifier-based approach can discriminate healthy subjects and T1DM patients; conversely, here we consider also T1DM patients with complications. Third, the DSS presented here was tested on a dataset larger than the one used in [9]. Fourth, the performance of the DSS are compared here also against those attained by the HbA1c tests.
The rest of the manuscript is organized as follows: next section presents the image dataset, Section 3 introduces the classification approach, whereas Section 4 describes the experimental setup. Sections 5 and 6 present and discuss the results, respectively. Finally, Section 7 provides concluding remarks.
Section snippets
Dataset
Blood samples were obtained after consent from 8 healthy subjects, 10 T1DM patients and 9 T1DM patients with complications. In the following, these three classes are referred to as Healthy, T1DM and T1DMwC, respectively. Upon collection, the RBCs were isolated from blood by density gradient centrifugation, counted, seeded in an uncoated two well dish (RPMI-1640 5% FCS) and labeled with Laurdan, 2-dimethylamino (6-lauroyl) naphthalene (Laurdan, Molecular Probes, Inc., Eugene, OR, USA), a
Methods
We are interested in discovering if the information contained in the two channels of RBC membrane images allows distinguishing between T1DM patients, T1DM patients with complications and healthy subjects. To this aim, we adopted a dual channel approach, as depicted in Fig. 2 A). Given a set of patient images, the image classification block classifies each image in the stack for each channel, assigning also a reliability measure to each decision. Next, for each channel, the labels and the
Experimental setup
The experiments were performed on the dataset presented hereinbefore using a leave-one-person-out cross-validation procedure. Hence, for each cross-validation run our approach is trained on all the data except for all the images belonging to one patient. In the test phase a classification is then made for that patient. This procedure guarantees that for each iteration the images of the same patient are not in the training and test set, that is, the separation of training and test data is done
Results
The results of the experiments are reported in Table 1, where the performance are measured in terms of accuracy, precision, recall and F1 score per each of the three classes Healthy, T1DM and T1DMwC (Section 2). This table shows the results attained by our proposal and by the three other approaches described in previous section.
To analyze how much the PCA and the multispectral approach affect the results, Table 2 reports the performance with and without the PCA step achieved by: the proposed
Discussion
Based on the results, in Table 1 we first observed that the proposed approach, referred to as MS-DSS, provides a perfect classification of the samples: nevertheless, the same performance are not attained by the state-of-the art technique for early detection of complications in T1DM. Indeed, thresholding the values of HbA1c on the cohort of patients contained in our database provides the performance reported in the second column of Table 1. Turning our attention to the results attained by the
Conclusion
In this paper we have presented a DSS leveraging on dual channel metabolic functional imaging data computed from the physical state of RBC membranes, a novel biomarker that allows to monitor T1DM progression. It uses features based on first- and second-order statistical measures as well as on rotation invariant co-occurrence local binary patterns. The dual channel approach also introduces a certain degree of redundancy that lowers the effect of image misclassifications. The recognition approach
Acknowledgement
The authors thank the anonymous reviewers that provided useful suggestions to improve the first version of the manuscript.
References (46)
- et al.
Two photon fluorescence microscopy of coexisting lipid domains in giant unilamellar vesicles of binary phospholipid mixtures
Biophys. J.
(2000) - et al.
Subclass discriminant analysis of morphological and textural features for HEp-2 staining pattern classification
Pattern Recognit.
(2014) - et al.
Mitotic cells recognition in hep-2 images
Pattern Recognit. Lett.
(2014) - et al.
On the use of classification reliability for improving performance of the one-per-class decomposition method
Data Knowl. Eng.
(2009) - et al.
Investigation of the spatial distribution of glutathione redox-balance in live cells by using fluorescence ratio imaging microscopy
Biosens. Bioelectron.
(2009) - et al.
Time evolution of noise induced oxidation in outer hair cells: role of nad (p) h and plasma membrane fluidity
Biochimica et Biophysica Acta (BBA) General Subjects
(2014) - et al.
A comparative study of texture measures with classification based on feature distribution
Pattern Recognit.
(1996) - et al.
Phase fluctuation in phospholipid membranes revealed by laurdan fluorescence.
Biophys. J.
(1990) - et al.
An expert system approach based on principal component analysis and adaptive neuro-fuzzy inference system to diagnosis of diabetes disease
Digit. Signal Process.
(2007) - et al.
A decision support system for crithidia luciliae image classification
Artif. Intell. Med.
(2011)
Er stress induced by the och1 mutation triggers changes in lipid homeostasis in kluyveromyces lactis
Res. Microbiol.
International Diabetes Federation
IDF Diabetes Atlas
Parameter tuning or default values? an empirical investigation in search-based software engineering
Empir. Softw. Eng.
6. glycemic targets
Diabetes Care
LIBSVM: A library for support vector machines
ACM Trans. Intell. Syst. Technol.
A fuzzy expert system for diabetes decision support application
IEEE Trans. Syst. Man Cybern. Part B
Type 1 diabetes through the life span: a position statement of the american diabetes association
Diabetes Care
Reliability parameters to improve combination strategies in multi-expert systems
Pattern Anal. Appl.
Early experiences in using blood cells biomembranes as markers for diabetes diagnosis
2016 IEEE 29th International Symposium on Computer-Based Medical Systems (CBMS)
Analysis of mammographic microcalcifications using gray-level image structure features
IEEE Trans Med Imag.
Solving multiclass learning problems via error-correcting output codes
J. Artif. Intell. Res.
Towards personalization of diabetes therapy using computerized decision support and machine learning: some open problems and challenges
Smart Health
Pattern Classification
Cited by (30)
Artificial intelligence and diabetes technology: A review
2021, Metabolism: Clinical and ExperimentalVideo-based eye tracking performance for computer-assisted diagnostic support of diabetic neuropathy
2021, Artificial Intelligence in MedicineCitation Excerpt :For instance, diagnostic support systems powered by Artificial Intelligence (AI) algorithms assessing the clinical history record of a patient can be used to predict the onset of complications related to diabetes [8–11]. Red blood cell analysis paired with AI methods can also facilitate early diagnosis of diabetes [12], while blood glucose monitoring can be used for continuous monitoring and prediction of diabetes evolution [13]. Externally, complications in the cardiovascular system can be first observed on the patient's skin.
Artificial intelligence applications in medical imaging: A review of the medical physics research in Italy
2021, Physica MedicaCitation Excerpt :DL has been proposed for classification of cells (e.g. Neutrophils, Lymphocytic) in rhino-cytogram microscopic images [94]. Type 1 diabetes mellitus can be diagnosed from Laudan intensity images of red blood cells obtained with confocal microscope, using ML classifier and radiomic features [161]. Clonogenic assays, quantifying the anti-proliferative effect of treatments on cell cultures, are usually performed by error-prone and operator-dependent manual counting of cells.
Auto-Regressive Time Delayed jump neural network for blood glucose levels forecasting
2020, Knowledge-Based SystemsCitation Excerpt :First, most methods exploit heterogeneous features to perform reliable predictions [3,5,6,9,12,14–17]. Indeed, the mechanism underlying the trend of glucose concentration in blood involves several players at different scales, but it is still not fully explained [23]. For example, some features like electrocardiogram signals and heart rate variability can lead to both diabetes diagnosis and management [24].
Erythrocyte membranes in metabolic and neurological diseases—supplementation with fatty acids and membranes remodeling
2020, Dietary Sugar, Salt and Fat in Human HealthRed blood cells membrane micropolarity as a novel diagnostic indicator of type 1 and type 2 diabetes
2019, Analytica Chimica Acta: XCitation Excerpt :The main limit of this work is the small sample size, and the validation of the approach on a larger population is necessary. However, this investigation paves the possibility for the use of the RBC micropolarity assay in diagnostics, in assessing the response to therapies and in fully describing the progression of T1DM and T2DM, complementing HbA1c in defining the quality of long-term management of diabetes [10,49]. Conceptualization, GM.; Methodology, GM, FD, GB, DP; Software, GM, FD; Validation, GM, GB, DP, GER.; Formal Analysis, GB, GM, GC,AR, FD, GER.; Investigation, GM, GB, FD, GER, AF, ED, MD.; Resources, GM, GC, AR, AF, FD, GER, ED, MD.; Data Curation, GB, FD, GER, ED, MD, GC, AR; Writing-Original Draft Preparation, GM.; Writing-Review & Editing, GM,GB, DP, FD,MD; Visualization, GB, FD, GER, ED, MD; Supervision, GM; Project Administration, GM,DP,MD; Funding Acquisition, GM.