Artificial intelligence-enhanced electrocardiogram analysis for identifying cardiac autonomic neuropathy in patients with diabetes

doi:10.21203/rs.3.rs-3735738/v1

Download PDF

Research Article

Artificial intelligence-enhanced electrocardiogram analysis for identifying cardiac autonomic neuropathy in patients with diabetes

https://doi.org/10.21203/rs.3.rs-3735738/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 11 Apr, 2024

Read the published version in Diabetes, Obesity and Metabolism →

Version 1

posted

You are reading this latest preprint version

Background

Cardiac autonomic neuropathy (CAN) is an important yet often overlooked complication of diabetes, which significantly increases the risk of cardiovascular (CV) events and mortality. Traditional diagnostic methods like CV autonomic function tests (CARTs) are laborious and rarely evaluated in clinical practice. This study aimed to develop and employ machine learning (ML) algorithms to analyze electrocardiogram (ECG) for the diagnosis of CAN.

Methods

We utilized motif and discord extraction techniques alongside Long Short-Term Memory (LSTM) networks to analyze 12-lead, 10 seconds ECG tracings to detect CAN in patients with diabetes. The performance of these methods with the Support Vector Machine (SVM) classification model was evaluated using Ten-Cross Validation (TCV) with the following metrics accuracy, precision, recall, F1 score, and area under the ROC Curve (AUC).

Results

Among 205 patients (mean age 54 ± 17; 54% female), 100 were diagnosed with CAN, including 38 with definite or severe CAN (dsCAN) and 62 with early CAN (eCAN). The best model performance for dsCAN classification was achieved using both motifs and discords, with an accuracy of 0.92, an F1 score of 0.92, a recall at 0.94, a precision of 0.91, and an excellent AUC of 0.93 (95%CI 0.91-0.94). For the detection of any stage of CAN, the approach combining motifs and discords yielded best results with an accuracy of 0.65, F1 score of 0.68, a recall of 0.75, a precision of 0.68, and an AUC of 0.68 (95%CI 0.54-0.81).

Conclusion

Our study highlights the potential of using ML techniques, particularly motifs and discords, to effectively detect dsCAN in patients with diabetes. This approach could be applied in large-scale screening of CAN, particularly to identify definite/severe CAN where CV risk factor modification may be initiated.

Cardiac Autonomic Neuropathy

Diabetes Complications

Machine Learning

Deep Learning

Electrocardiogram

Motif

Discord

Cardiac Autonomic Neuropathy (CAN) is an under-recognized complication of diabetes mellitus (DM), characterized by the impaired autonomic regulation of the cardiovascular (CV) system [1]. It is highly prevalent in type 1 DM (up to 35%) [2] and in type 2 DM (up to 73%) [3, 4].

CAN is associated with both increased mortality and morbidity. A comprehensive meta-analysis encompassing 26 longitudinal studies with over 15,000 patients, followed for a duration of 2.75 to 10.1 years, revealed that a diagnosis of CAN was associated with more than a three-fold increased risk of both all-cause mortality and CV events [5]. The ACCORD trial, which prospectively followed up over 10,000 individuals with type 2 DM found that individuals with CAN had a 1.55 to 2.14-fold higher risk of mortality than those without CAN [6]. Furthermore, prolongation of the corrected QT interval, a marker of CAN, has been identified as an independent predictor of cardiovascular and all-cause mortality in both type 1 and type 2 DM [7–9].

Conventional diagnostic modalities for CAN such as Cardiovascular Autonomic Function Tests (CARTs), though effective, are cumbersome for large-scale screening and are thus rarely utilized in clinical practice; perpetuating underdiagnosis and lack of inclusion in screening guidelines. Recent research has shown the reliability of heart rate variability (HRV) analysis from ultra-short electrocardiogram (ECG) recordings, even as brief as 10 seconds, in distinguishing between individuals with definite CAN and healthy controls. This finding, as reported by Wehler et al. [10], demonstrates the potential of short ECG tracings to capture significant diagnostic information. In support of this approach, the notable success reported in Kulkarni et al. [11] in the detection of DM and pre-diabetes from 10 second ECG tracings using artificial intelligence (AI) algorithms, including the Long-Short Term Memory (LSTM) algorithm, exemplified AI’s capacity to identify complex patterns within ECG data, potentially attributable to the autonomic dysfunction in DM.

This achievement suggests the feasibility of adapting such technology for CAN identification. Furthermore, this approach leverages the widespread availability and accessibility of short ECG tracings. These tracings, usually obtained in diverse clinical environments, enhance the practicability of utilizing AI for CAN detection. Effective use of AI for ECG analysis could open new avenues for early and accurate CAN diagnosis in low-resource settings necessitating advanced diagnostic methods.

We hypothesized that employing machine learning (ML) algorithms in the analysis of short ECG tracings will allow for the detection of CAN. This study aims to develop and assess the ability of ML-driven ECG analysis in detecting CAN in patients with DM.

Study Design and Participants

This study was an observational, single-centre study conducted as a part of the Silesia Diabetes-Heart Project (ClinicalTrials.gov Identifier: NCT05626413). The study was performed in the accordance with the Declaration of Helsinki principles and received approval from the Ethics Committee of the Medical University of Silesia (KNW/0022/KB1/10/17). It involved consecutive patients with DM admitted to the Department of Internal Medicine and Diabetology at Clinic Hospital Number 1 in Zabrze, Poland, and outpatients from various Diabetology Clinics in the Silesia Region who fulfilled inclusion/ exclusion criteria. The study was carried out from July 29, 2022, to July 9, 2023. Inclusion criteria were as follows: adults aged 18 or older, diagnosis of DM type 1 for at least 5 years or DM type 2 of any duration. Exclusion criteria included presence of the following conditions: proliferative retinopathy, any severe and acute illness, disabled and bedridden patients, solid organ transplant, previously diagnosed causes of neuropathy other than DM, pregnancy, alcohol use disorder, and severe hypoglycemia in the past 24 hours. Informed consent prior to any assessment related to the study was mandatory.

Data Collection and study procedures

On the day of the visit planned for participation in this study data on demographics, comorbid conditions and pharmacotherapy were gathered and body mass and height were measured, based on which body mass index (BMI) was calculated as body mass in kilograms squared divided by height in meters. Moreover, fasting venous blood was collected for creatinine, total cholesterol, low density lipoprotein (LDL) cholesterol, high density lipoprotein (HDL) cholesterol and triglycerides, furthermore, urine spot was collected. CARTs and ECG were performed as described in detail in the paragraph below. The presence of chronic kidney disease was determined using the following criteria: estimated glomerular filtration rate (eGFR) below 60 ml/min per 1.73m² or a urine albumin to creatinine ratio (UACR) of 30 mg/g or more, sustained for at least three months. Diabetic retinopathy grading into mild, moderate, severe, or proliferative was made independently by two experienced ophthalmologists, who relied on fundus images taken at the same visit as CAN testing and ECG exam.

CAN Assessment Protocol

Diagnosis of CAN was performed adhering to the Toronto consensus panel criteria, through CARTs – the acknowledged standard for CAN detection [12]. The DiCAN system (Diabetic Cardiac Autonomic Neuropathy, Medicore, Seoul, South Korea) was used to conduct these tests. Testing sessions were conducted in a controlled environment, ensuring consistent lighting and minimal noise levels in the room. The evaluation included several key tests: the heart rate response to a deep breathing test (calculating the expiration/inspiration ratio), a lying-to-standing test (measuring the 30:15 ratio from supine to upright position), and the Valsalva manoeuvre. Orthostatic hypotension was also assessed by comparing blood pressure measurements before and after the patient stood up. Incomplete heart rate test data led to exclusion of the patient from the study. The staging of CAN was determined by the count of abnormal CART results: early CAN (eCAN) with one abnormal test result; definite CAN with two or more abnormal results and severe CAN characterized by two or more abnormal tests along with orthostatic hypotension. For subsequent analyses, patients diagnosed with definite or severe CAN were categorized collectively as dsCAN. Details about the specific cut offs applied for the CARTs can be found in Supplementary Table S1.

ECG Examination

All 12-lead ECG exams were conducted using an Opus 1 ECG machine (String Polska Ltd, Poland). The patients were positioned comfortably, and electrodes were attached to predefined chest, arm, and leg locations. Each ECG recording, conducted at a 500 Hz sampling rate, lasted for up to 30 seconds. The obtained raw ECG data were then stored digitally for subsequent analysis. In this dataset a small number of patient ECGs (fewer than 20) exhibited artifacts, likely due to patient movement. These artifacts were manually identified and removed.

Proposed Model

ECG classification using AI generally includes three basic stages: pre-processing, feature extraction, and classification model generation. In the pre-processing stage, noise is removed from the ECG signal data, and if there is a need to fix the length, this is also conducted in this step (many machine and deep leaners that work with time series require the time series to be of equal length). The next stage is to extract feature. Feature extraction is a well-established process in AI to: (i) reduce the complexity of machine and deep learning problems and (ii) enhance the performance of the AI models [13]. In this paper, two feature extraction methods are used: motifs and discords, and LSTM feature vector. Once feature extraction was complete the extracted features were used to train an ML model for classification in the last stage. Once generated the model can be used to classify previously unseen data. In the context of the work presented in this paper a support vector machine (SVM) classification model generator was used. Figure 1 illustrates the process. Note that in the figure we have two groups: Group 1 and Group 2: Group 1 include two sets (No CAN/eCAN and dsCAN) and Group 2 include two sets (No CAN and eCAN/dsCAN). Each of the basic stages outlined above is explained in further detail in the following three sub-sections.

Pre-processing

To train the two models, motif/discord-based and LSTM based, the sequence length of each input needed to be fixed. For this reason, first we found out all the different lengths in the dataset (3000 to 15000), then according to the average the (5000) was chosen. Consequently, all ECG signal samples in the dataset were standardized to a length of 5000 samples, corresponding to a duration of 10 seconds. To ensure all samples were of the same sequence length, sample processing was performed as follows. If the number of sample points exceeded 5000, only the first 5000 sample points were retained. If the number of sample points was less than 5000, the sample was padded with zero values until its sequence length reached 5000 (there were only 4 patients, with ECG lengths of 3000, where this was required). This process resulted in a consistent data matrix format of 12x5000 for each ECG, which was used for the subsequent analysis.

Feature Extraction

The extracted features from ECG can be: (i) Global characteristics, typically time intervals between key events and the amplitudes associated with the P wave, QRS complex, T wave and RR interval referred to earlier [14, 15], (ii) Discriminative time series sub-sequences such as motif (or shapelets) and/or discords [16, 17], or (iii) feature vectors extracted using LSTM. The last is the most sophisticated because of the deep neural networks (DNNs) ability to learn complex patterns, dependencies and relations in time series [18]. As noted above, two feature extraction methods were considered: (i) motif/discord extraction and (ii) LSTM feature vector extraction. Each is discussed in turn in this sub-section.

Motif and Discord Extraction

For the work presented in this paper, the motif/discord extraction method proposed in [19] was adopted. The basic idea was to extract all-time series sub-sequences of length w from a "parent" time series, and then select the most frequently and infrequently occurring. To count the frequency of occurrence a similarity metric was used, because there was no requirement for sub-sequences to match exactly for them to be said to be the same. In [19] the Correct Matrix Profile (CMP) technique was used which is data structure that features two vectors for each time series: a Distance Profile (DP) and Profile Index (PI). After all sub-sequences of length w were extracted (in this paper w = 250, which was 1/20 of the original length = 5000, was adopted), the distance between each sub-sequence and its nearest neighbour were identified. The identified pairs formed DP and PI vectors. These two vectors then facilitate the extraction of motifs and discords. Figure 2 shows the motif/discord discovery process. For each group, the input to the motif discovery algorithm was divided according to class (given that we were interested in a binary classification). The next step was to process these two sets to extract motifs/discords. The outcome was two sets of motifs/discords, one per class. Figure 3 shows an example of the output for one lead, from top to bottom the figure presents: (a) the one lead original ECG time series is given at the top of the Figure, (b) the CMP calculated from the ECG, (c) the three identified motifs, and (d), the three extracted discords.

The next stage was then the motif/discords selection. The process whereby the motifs/discords that appeared in the majority of patients based on the similarity and count threshold given in [19]. The selected motifs/discords were then stored for the next step. The last stage was to compare the selected motifs/discords from each class and remove the motifs/discords that were attached to both classes and store the remaining set of motifs/discords. This final set was used to extract the count of each motif/discord in each patient to be the input for the classifier generation step.

LSTM Feature Vector Extraction

An LSTM is a type of Recurrent Neural Network, developed by Schmidhuber and Hochreiter in 1997 [20]. It has sequences of memory cells that are capable of learning long-term patterns. In the context of using an LSTM as a feature extractor for ECG, the main objective is to leverage the network's ability to capture and learn relevant temporal patterns in the ECG data. The LSTM network is designed with input, output, and memory cell components (see Fig. 4). The architecture used with respect to the work reported in this paper had four LSTM layers, each with a number of memory cells. ECG data are naturally sequential, the sequences are fed into the LSTM network sequentially. At each time step, the network processes the input data and updates its internal memory state. The memory cells in the LSTM have incorporate forget gates, input gates and output gates, to control the flow of information. These gates enable the network to selectively retain or discard information, allowing it to capture long-term dependencies in the ECG signal.

An LSTM automatically extracts features related to the temporal patterns present in the ECG signal and the hidden states capture the relevant information as extracted features. The final hidden state of the LSTM, obtained after processing the entire ECG sequence, serves as a feature vector. This vector encodes the learned representations and can be used as input to downstream tasks. The process of using the extracted features is explained in Fig. 5. To obtain a good result from our LSTM network, the LSTM first was trained using labelled data, where the ECG sequences are associated with specific classes. The dataset was divided 80/20 for training and validation. After 20 epochs, the LSTM network converged, and the classification accuracy was stable at 94%. During the training, the network adjusts its internal parameters through backpropagation to minimize the difference between predicted and actual outcomes. Then we used the trained model to extract the features from the ECG dataset and used those for features as input for the classification step.

Classification

The last stage in the proposed process was to use the extracted feature to create a feature vector for the classification stage. During this stage the identified set of motifs/discords, or LSTM features, were used. Both were encapsulated as features vector so that, if desired, additional features, such as clinical data, could be added at a later stage. This feature vector idea was proposed by [21] to combine different data sources. For the motif/discord features, each vector represented an ECG time series, and each value in the vector the numeric occurrence counts of a motif or discord in the time series. For the LSTM method, each vector represents an ECG time series, and each value the numeric value from the extracted set of features from this ECG.

Implementation Details

Using both feature extraction mechanisms two SVM classifiers were generated. To evaluate the two generated classifiers four experiments were conducted, or each group and each classifier. Grid search technique was used for determining the optimal parameters for SVM generation. The grid search technique involves various combinations of parameter values (C, gamma, and kernel) for the SVM within a predetermined range. For each experiment Ten-fold Cross Validation (TCV) was used whereby the dataset was divided into ten equally sized subsets. The experiments were run ten times each time using a different tenth as the test set and the remaining nine tenths as the training and validation sets.

Outcome and Performance Metrics

The primary outcome of this study was to evaluate the effectiveness of ML in detecting CAN among patients with DM. This was assessed using standard 10-second, 12-lead ECG recordings. The performance of the generated models in identifying CAN was quantitatively using the area under the receiver operating curve (AUC). Additionally, accuracy, precision, recall, and F1were calculated.

In this study, 205 patients (mean age 54, SD 17; 54% female) were included in the analysis. Exclusions were due to various factors: refusal to consent for ECG and CAN examination, absence of ECG examination, poor ECG quality, and missing or incomplete CAN testing. These details are depicted in the study flowchart (Fig. 6). Among the analyzed patients, 100 (49%) had CAN, including 38 (17%) diagnosed with dsCAN and 62 (30%) with eCAN. Demographic and clinical characteristics of these participants are summarized in Table 1. Patients with CAN were significantly older, predominantly had type 2 DM, and a higher number of comorbidities. Those with advanced CAN stages had a longer DM duration and higher prevalence of hypertension, coronary artery disease, and chronic kidney disease, alongside lower estimated glomerular filtration rates.

Table 1

Baseline characteristics of study population
Characteristic	No CAN N = 105¹	eCAN N = 62¹	dsCAN N = 38¹	p-value²
Male	53 (50%)	26 (42%)	15 (39%)	0.4
Age	49 (32, 61)	62 (52, 69)	66 (61, 70)	< 0.001
BMI	28.2 (24.6, 32.4)	30.0 (26.4, 32.3)	30.3 (27.5, 34.0)	0.2
Type of diabetes				0.001
1	39 (37%)	14 (23%)	3 (7.9%)
2	66 (63%)	48 (77%)	35 (92%)
Diabetes duration	10.0 (4.0, 14.0)	10.0 (5.0, 20.0)	15.0 (9.3, 20.0)	0.008
Diabetic retinopathy				0.3
mild	12 (14%)	5 (10%)	2 (6.9%)
moderate	7 (8.0%)	8 (17%)	5 (17%)
severe	4 (4.6%)	4 (8.3%)	0 (0%)
proliferative	0 (0%)	0 (0%)	1 (3.4%)
Chronic kidney disease	27 (26%)	20 (32%)	19 (50%)	0.023
Hypertension	59 (56%)	41 (66%)	32 (84%)	0.008
Coronary artery disease	3 (2.9%)	5 (8.1%)	12 (32%)	< 0.001
History of stroke	5 (4.8%)	3 (4.8%)	0 (0%)	0.4
History of myocardial infarction	5 (4.8%)	4 (6.5%)	8 (21%)	0.013
Atrial fibrillation	6 (5.7%)	2 (3.2%)	6 (16%)	0.053
HBA1C [%]	7.6 (6.6, 8.6)	8.0 (6.4, 10.4)	7.8 (6.9, 9.3)	0.15
UACR	8.1 (4.5, 12.4)	9.9 (6.8, 22.4)	9.3 (4.9, 18.8)	0.12
eGFR	93.0 (79.5, 104.7)	90.0 (73.2, 100.8)	74.9 (53.2, 85.1)	< 0.001
Total cholesterol	4.7 (4.1, 5.5)	4.7 (3.8, 5.5)	4.6 (3.6, 5.2)	0.5
LDL	2.5 (2.2, 3.2)	2.2 (1.5, 3.2)	2.2 (1.5, 3.2)	0.12
Triglycerides	1.4 (0.8, 2.2)	1.4 (1.1, 2.0)	1.9 (1.4, 2.3)	0.037
Antihypertensive drugs use	49 (47%)	37 (60%)	28 (74%)	0.012
ACEi	25 (24%)	26 (42%)	14 (37%)	0.039
ARB	11 (10%)	7 (11%)	6 (16%)	0.7
Beta blocker	32 (30%)	28 (45%)	19 (50%)	0.046
CARTs
Deep breathing test	0 (0%)	27 (44%)	22 (58%)	< 0.001
Valsalva test	0 (0%)	8 (14%)	11 (30%)	< 0.001
30:15 test	0 (0%)	13 (24%)	5 (14%)	< 0.001
Orthostatic hypotension test	20 (19%)	17 (27%)	9 (24%)	0.5
¹ n (%); Median (IQR) ² Pearson’s Chi-squared test; Kruskal-Wallis rank sum test; Fisher’s exact test Abbreviations: ACEi, angiotensin converting enzyme inhibitor; ARB, angiotensin receptor blocker; BMI, body mass index; CAN, cardiac autonomic neuropathy; CARTs, cardiovascular autonomic reflex tests; eCAN, early CAN; dsCAN, definite or severe CAN; eGFR, estimated glomerular filtration rate; LDL, low density lipoprotein; UACR, urine albumin creatinine ratio.

Classification of dsCAN

The results obtained with respect to the classification of dsCAN are given Table 2. The model utilizing both motifs and discords demonstrated the best overall performance, accurately classifying 94% of patients with dsCAN, and achieving an accuracy of 0.92 and an AUC of 0.93 (95%CI 0.91–0.94).

Table 2

Performance Metrics for ML Models in the Classification of dsCAN
Model	Accuracy	F1 Score	Recall	Precision	AUC
Motifs Only + SVM	0.91	0.91	0.91	0.91	0.91 [0.9, 0.93]*
Discords Only + SVM	0.86	0.86	0.86	0.88	0.87 [0.86, 0.88]*
Motifs + Discords + SVM	0.92	0.92	0.94	0.91	0.93 [0.91, 0.94]*
LSTM + SVM	0.88	0.88	0.91	0.87	0.88 [0.85, 0.92]*
Abbreviations: AUC, Area Under the ROC Curve; LSTM, Long Short-Term Memory; SVM, Support Vector Machine. *95% CI

In comparison, the models employing only motifs or only discords yielded lower performance metrics. The use of LSTM for feature extraction combined with a SVM resulted in an accuracy of 0.88 and an AUC of 0.88 (95%CI 0.85–0.92), lower than those of motifs + discords + SVM.

Detection of Any Stage of CAN

The results obtained in the context of CAN classification are given in Table 3. From Table 3 it can be seen that the Motifs Only and the Discords Only models showed poor performance, with accuracy, F1 score, recall, precision, and AUC all not exceeding 0.55. However, the combination of Motifs with Discords marked a significant improvement, yielding an accuracy of 0.65, F1 score of 0.68, recall of 0.75, precision of 0.68, and an AUC of 0.68 (95%CI 0.54–0.81). The model with LSTM and SVM also showed modest performance, with accuracy of 0.54 and AUC of 0.55 (95%CI 0.49–0.62).

Table 3

Performance Metrics for ML Models in the detection of CAN
Method	Accuracy	F1 Score	Recall	Precision	AUC
Motifs Only + SVM	0.53	0.52	0.53	0.54	0.53 [0.51, 0.55] *
Discords Only + SVM	0.54	0.53	0.54	0.54	0.54 [0.51, 0.57] *
Motifs + Discords + SVM	0.65	0.68	0.75	0.68	0.68 [0.54, 0.81] *
LSTM + SVM	0.54	0.56	0.59	0.56	0.55 [0.49, 0.62] *
Abbreviations: AUC, Area Under the ROC Curve; LSTM, Long Short-Term Memory; SVM, Support Vector Machine.
Abbreviations: CAN, Cardiac autonomic neuropathy; eCAN, early CAN; ECG, electrocardiogram; dsCAN, definite or severe CAN

In this study, we have successfully shown the feasibility of detecting CAN using short, 10-second 12-lead ECG recordings with the more cumbersome gold-standard CARTs as a reference. We observed robust performance, particularly in identifying dsCAN, when motifs and discords were used with an SVM classifier. This approach yielded high accuracy and precision. However, in the broader context of detecting any stage of CAN, our models’ performance was more modest.

These results highlight the potential of ML in detecting subtle physiological variations characteristic of CAN in 12-lead ECGs, demonstrating the usefulness and clinical relevance of this ML-driven approach in CAN diagnostics. Recently, a study of ML of infant shows that ECGs contain important information about autism spectrum disorder familial likelihood through HRV and sympathetic and parasympathetic activities [22]. Thus, demonstrating the possible wide-ranging utility of ML for ECG analysis beyond cardiac rhythm disorders.

With regard to classifying patients with eCAN, we were able to achieve good recall, but with the trade-off in accuracy, reflected in an AUC of 0.68. Jelinek et al. [23] attempted to detect eCAN utilizing HRV features with fast Fourier transform and Lomb-Scargle periodogram based on ECG lasting from 10 seconds to 5 minutes. Jelinek et al. achieved similar success with AUC values ranging from 0.6 to 0.75 depending on the ECG duration and employed method of analysis. This highlights the inherent challenges in ECG-based eCAN detection.

Numerous studies have leveraged ECG as a diagnostic tool for CAN, primarily utilizing HRV analysis. Notably, Tang et al. [24] explored the diagnostic utility of short-term HRV in CAN detection using Bayesian estimation, challenging the conventional assumption of CARTs as the gold-standard. Their findings indicate that short-term HRV analysis was non-inferior to CARTs for CAN diagnosis. Additionally, specific HRV features like the Root Mean Square of Successive Differences (RMSSD) and the standard deviation of the averages of RR intervals (SDNN) were identified as crucial for CAN detection when employing ensemble classification with random forest [25]. These HRV features are known to be diminished in CAN, attributed to parasympathetic withdrawal leading to sympathetic dominance [26, 27]. However, relying solely on SDNN for CAN detection demonstrated limited efficacy, yielding an AUC of 0.73 [26]. Our study was not limited to the extraction of only HRV features from ECG. The ML algorithms we employed for feature extraction were capable of discerning a broader range of differences in ECG between groups.

Most of the existing ECG-based biometric recognition methods in the literature are based on the identification of key points, referenced using the notation P, QRS, and T. These are specific points of interest within an ECG signal. The accuracy of these methods depends on the precise detection of these points, which is considered challenging in real-world biometric applications [28]. To address this limitation, we used two alternative feature extraction methods that do not require any specific form of point detection before extraction. The first was founded on motif and discord extraction. A motif is time series sub-sequence that repeats frequently, while a discord is a time series sub-sequence that occurs infrequently. Both, it is argued [19], are good discriminators of class. The second was founded on the LSTM algorithm, a deep learning algorithm which is increasingly used in the field of health care [29, 30]. LSTMs frequently adopted in with respect to recent work on ECG data analysis, because of their reported success [18, 31].

This study builds upon existing research, and enhances ECG-based CAN diagnosis by incorporating ML algorithms. While prior studies have achieved promising results in detecting CAN using ECG-derived features such as QT or RR intervals [32] and HRV parameters [33], these often suffered from methodological shortcomings. Common limitations include small sample sizes, the use of non-gold-standard CAN criteria or failure to clearly define these criteria, and the lack of a robust validation process.

Recent studies predicted CAN based on biochemical, demographic data and patients’ history with the use of ML techniques [34]. Nedergaard et al. [25] developed ensembled classification using established ECG-derived features along with clinical and biochemical data. Our study suggests the potential of enhancing such models by integrating novel ECG features identified through motifs, discords, or LSTM along with other types of features.

The potential medical applicability of our findings, particularly in screening and CV disease (CVD) prevention, is noteworthy. Utilizing ML techniques in conjunction with ECG analysis, as demonstrated in this study, provides a novel approach for the early, non-invasive detection of CAN. This condition often precedes more severe cardiovascular complications, making its early identification crucial [5]. The ability of our method to discern subtle ECG alterations indicative of CAN, through the analysis of motifs, discords, and LSTM feature vectors, holds significant promise for identifying individuals at an elevated risk of developing CVD. Timely detection can facilitate earlier intervention, including lifestyle changes and preventive medical therapies, potentially slowing or preventing the progression to overt CVD [35, 36]. Furthermore, the non-invasive nature of ECG screening, combined with the efficiency and accuracy of AI algorithms, makes this approach highly feasible for widespread implementation in clinical practice, potentially enhancing the effectiveness of CVD prevention strategies.

Given that individuals with CAN are at much higher CV risk, newer therapies such as Sodium-glucose cotransporter 2 inhibitors (SGLT2i) and Glucagon-like peptide-1 receptor agonists (GLP-1 ra) may be important in future clinical pathways which define dsCAN. For instance, SGLT2i mediate reduction in CV death and heart failure through mechanism that extends beyond improving glycaemic control [37] and may include reduction in sympathetic tone [38]. However, the development of point-of-care screening for CAN will be essential in risk stratifying those with the highest CV risk.

Limitations

This study enrolled a modestly sized cohort (n = 205). The restricted participant pool from a single centre may not adequately represent the broader, diverse population affected by CAN, potentially limiting the generalizability of our findings. Moreover, the use of neural network models in our analysis, might be hindered by the smaller dataset. Such models typically require extensive data to effectively learn and generalize. Additionally, without external validation across various clinical settings and populations, the robustness and applicability of our ML models remain to be fully established. These factors underscore the necessity for further, expansive research to validate and enhance the reliability of our ML-driven diagnostic approach. Moreover, future studies could focus on the classification based on ECG enhanced by additional clinical or biochemical data.

Our study’s ML approach, particularly using motifs and discords, effectively identified dsCAN in patients with diabetes. The approach showed high accuracy, indicating its potential for enhancing CAN diagnosis, future screening and aiding in timely intervention in this high CV risk population.

Ethics approval

The study was performed in the accordance with the Declaration of Helsinki principles and received approval from the Ethics Committee of the Medical University of Silesia (KNW/0022/KB1/10/17).

Consent for publication

Not applicable

Funding

This study received funding from Polish Ministry of Education and Science (SKN/SP/534630/2022) and Statutory work No PCN-1-085/N/2/K of Medical University of Silesia.

Conflict of interest

The authors declare no conflicts of interest.

Availability of data and materials’ statement

Data are available upon reasonable request.

Authors’ Contributions

KN, HK, GYHL, UA substantially contributed to the conception and design of the work. KI, M Hendel, HK, JP, JK, PI, SB, M Herba and KK oversaw recruitment and collected the data. KI curated the data. HA worked on the proposed algorithms and conducted experimental testing. KI, HA and M Hendel prepared the tables and figures. KI, HA and KN drafted the manuscript. KI, HA, M Hendel, HK, FC, JG, ZY, GYHL, UA and KN reviewed and edited the original draft. KN, FC, JG, YZ, GYHL and UA provided senior oversight and authorized the final manuscript. KN is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. All of the authors have read and approved the submitted version of the manuscript.

Acknowledgements

Not applicable

John APP, Udupa K, Avangapur S, Sujan MU, Inbaraj G, Vasuki PP, et al. Cardiac autonomic dysfunctions in type 2 diabetes mellitus: an investigative study with heart rate variability measures. Am J Cardiovasc Dis. 2022;12:224–32.
Pop-Busui R, Low PA, Waberski BH, Martin CL, Albers JW, Feldman EL, et al. Effects of Prior Intensive Insulin Therapy on Cardiac Autonomic Nervous System Function in Type 1 Diabetes: The Diabetes Control and Complications Trial/Epidemiology of Diabetes Interventions and Complications Study (DCCT/EDIC). Circulation. 2009;119:2886–93. https://doi.org/10.1161/CIRCULATIONAHA.108.837369.
Zoppini G, Cacciatori V, Raimondo D, Gemma M, Trombetta M, Dauriz M, et al. Prevalence of Cardiovascular Autonomic Neuropathy in a Cohort of Patients With Newly Diagnosed Type 2 Diabetes: The Verona Newly Diagnosed Type 2 Diabetes Study (VNDS). Diabetes Care. 2015;38:1487–93. https://doi.org/10.2337/dc15-0081.
Low PA, Benrud-Larson LM, Sletten DM, Opfer-Gehrking TL, Weigand SD, O’Brien PC, et al. Autonomic symptoms and diabetic neuropathy: a population-based study. Diabetes Care. 2004;27:2942–7. https://doi.org/10.2337/diacare.27.12.2942.
Chowdhury M, Nevitt S, Eleftheriadou A, Kanagala P, Esa H, Cuthbertson DJ, et al. Cardiac autonomic neuropathy and risk of cardiovascular disease and mortality in type 1 and type 2 diabetes: a meta-analysis. BMJ Open Diabetes Research and Care. 2021;9:e002480. https://doi.org/10.1136/bmjdrc-2021-002480.
Pop-Busui R, Evans GW, Gerstein HC, Fonseca V, Fleg JL, Hoogwerf BJ, et al. Effects of Cardiac Autonomic Dysfunction on Mortality Risk in the Action to Control Cardiovascular Risk in Diabetes (ACCORD) Trial. Diabetes Care. 2010;33:1578–84. https://doi.org/10.2337/dc10-0125.
Soedamah-Muthu SS, Chaturvedi N, Witte DR, Stevens LK, Porta M, Fuller JH, et al. Relationship Between Risk Factors and Mortality in Type 1 Diabetic Patients in Europe: The EURODIAB Prospective Complications Study (PCS). Diabetes Care. 2008;31:1360–6. https://doi.org/10.2337/dc08-0107.
Cox AJ, Azeem A, Yeboah J, Soliman EZ, Aggarwal SR, Bertoni AG, et al. Heart Rate–Corrected QT Interval Is an Independent Predictor of All-Cause and Cardiovascular Mortality in Individuals With Type 2 Diabetes: The Diabetes Heart Study. Diabetes Care. 2014;37:1454–61. https://doi.org/10.2337/dc13-1257.
Rossing P, Breum L, Major-Pedersen A, Sato A, Winding H, Pietersen A, et al. Prolonged QTc interval predicts mortality in patients with Type 1 diabetes mellitus. Diabet Med. 2001;18:199–205. https://doi.org/10.1046/j.1464-5491.2001.00446.x.
Wehler D, Jelinek HF, Gronau A, Wessel N, Kraemer JF, Krones R, et al. Reliability of heart-rate-variability features derived from ultra-short ECG recordings and their validity in the assessment of cardiac autonomic neuropathy. Biomed Signal Process Control. 2021;68:102651. https://doi.org/10.1016/j.bspc.2021.102651.
Kulkarni AR, Patel AA, Pipal KV, Jaiswal SG, Jaisinghani MT, Thulkar V, et al. Machine-learning algorithm to non-invasively detect diabetes and pre-diabetes from electrocardiogram. BMJ Innovations. 2023;9. https://doi.org/10.1136/bmjinnov-2021-000759.
Spallone V, Ziegler D, Freeman R, Bernardi L, Frontoni S, Pop-Busui R, et al. Cardiovascular autonomic neuropathy in diabetes: clinical impact, assessment, diagnosis, and management. Diab/Metab Res Rev. 2011;27:639–53. https://doi.org/10.1002/dmrr.1239.
Khalid S, Khalil T, Nasreen S. A survey of feature selection and feature extraction techniques in machine learning. 2014 Science and Information Conference, 2014, p. 372–8. https://doi.org/10.1109/SAI.2014.6918213.
Yazdani S, Vesin J-M. Extraction of QRS fiducial points from the ECG using adaptive mathematical morphology. Digit Signal Proc. 2016;56:100–9. https://doi.org/10.1016/j.dsp.2016.06.010.
Sasikala P, Wahidabanu DRSD, Robust R, Peak. QRS detection in Electrocardiogram using Wavelet Transform. Int J Adv Comput Sci Appl (IJACSA). 2012;1. https://doi.org/10.14569/IJACSA.2010.010608.
Elangovan R. A Review on Time Series Motif Discovery Techniques an Application to ECG Signal Classification: ECG Signal Classification Using Time Series Motif Discovery Techniques. IJAIML. 2019;9:39–56. https://doi.org/10.4018/IJAIML.2019070103.
Ji C, Hu Y, Liu S, Pan L, Li B, Zheng X. Fully convolutional networks with shapelet features for time series classification. Inf Sci. 2022;612:835–47. https://doi.org/10.1016/j.ins.2022.09.009.
Liu X, Wang H, Li Z, Qin L. Deep learning in ECG diagnosis: A review. Knowl Based Syst. 2021;227:107187. https://doi.org/10.1016/j.knosys.2021.107187.
Aldosari H, Coenen F, Lip GYH, Zheng Y. Motif Based Feature Vectors: Towards a Homogeneous Data Representation for Cardiovascular Diseases Classification. In: Golfarelli M, Wrembel R, Kotsis G, Tjoa AM, Khalil I, editors. Big Data Analytics and Knowledge Discovery. Cham: Springer International Publishing; 2021. pp. 235–41. https://doi.org/10.1007/978-3-030-86534-4_22.
Hochreiter S, Schmidhuber J. Long Short-Term Memory. Neural Comput. 1997;9:1735–80. https://doi.org/10.1162/neco.1997.9.8.1735.
Aldosari H, Coenen F, Lip GYH, Zheng Y. Addressing the Challenge of Data Heterogeneity Using a Homogeneous Feature Vector Representation: A Study Using Time Series and Cardiovascular Disease Classification. In: Bramer M, Ellis R, editors. Artificial Intelligence XXXVIII. Cham: Springer International Publishing; 2021. pp. 254–66. https://doi.org/10.1007/978-3-030-91100-3_21.
Tilwani D, Bradshaw J, Sheth A, O’Reilly C. ECG Recordings as Predictors of Very Early Autism Likelihood: A Machine Learning Approach. Bioengineering. 2023;10:827. https://doi.org/10.3390/bioengineering10070827.
Jelinek HF, Adam MTP, Krones R, Cornforth DJ. Diagnostic Accuracy of Random ECG in Primary Care for Early, Asymptomatic Cardiac Autonomic Neuropathy. J Diabetes Sci Technol. 2017;11:1165–73. https://doi.org/10.1177/1932296817703670.
Tang Z-H, Wang L, Zeng F, Li Z, Yu X, Zhang K, et al. Bayesian estimation of cardiovascular autonomic neuropathy diagnostic test based on short-term heart rate variability without a gold standard. BMJ Open. 2014;4:e005096. https://doi.org/10.1136/bmjopen-2014-005096.
Nedergaard RB, Scott M, Wegeberg A-M, Okdahl T, Størling J, Brock B, et al. Features characterising cardiac autonomic neuropathy in diabetes using ensembled classification. Clin Neurophysiol. 2023;154:200–8. https://doi.org/10.1016/j.clinph.2023.06.011.
Pop-Busui R, Backlund J-YC, Bebu I, Braffett BH, Lorenzi G, White NH, et al. Utility of using electrocardiogram measures of heart rate variability as a measure of cardiovascular autonomic neuropathy in type 1 diabetes patients. J Diabetes Invest. 2022;13:125–33. https://doi.org/10.1111/jdi.13635.
Metelka R. Heart rate variability - current diagnosis of the cardiac autonomic neuropathy. A review. Biomed Pap Med Fac Univ Palacky Olomouc Czech Repub. 2014;158:327–38. https://doi.org/10.5507/bp.2014.025.
Ribeiro Pinto J, Cardoso JS, Lourenço A, Evolution. Current Challenges, and Future Possibilities in ECG Biometrics. IEEE Access. 2018;6:34746–76. https://doi.org/10.1109/ACCESS.2018.2849870.
Abdel-Jaber H, Devassy D, Al Salam A, Hidaytallah L, EL-Amir M. A Review of Deep Learning Algorithms and Their Applications in Healthcare. Algorithms. 2022;15:71. https://doi.org/10.3390/a15020071.
Shamshirband S, Fathi M, Dehzangi A, Chronopoulos AT, Alinejad-Rokny H. A review on deep learning approaches in healthcare systems: Taxonomies, challenges, and open issues. J Biomed Inform. 2021;113:103627. https://doi.org/10.1016/j.jbi.2020.103627.
Ebrahimi Z, Loni M, Daneshtalab M, Gharehbaghi A. A review on deep learning methods for ECG arrhythmia classification. Expert Syst Applications: X. 2020;7:100033. https://doi.org/10.1016/j.eswax.2020.100033.
Sharanya S, Sridhar PA. A model for early diagnosis of Cardiac Autonomic Neuropathy (CAN). J Phys: Conf Ser. 2021;2089:012053. https://doi.org/10.1088/1742-6596/2089/1/012053.
Mehta B, Dave V. Detection of Cardiovascular Autonomic Neuropathy using Machine Learning Algorithms. J Integr Sci Technol. 2023;11:562–2. https://scholar.google.com/scholar?q=urn:nbn:sciencein.jist.2023.v11.546.
Abdalrada AS, Abawajy J, Al-Quraishi T, Islam SMS. Prediction of cardiac autonomic neuropathy using a machine learning model in patients with diabetes. Therapeutic Adv Endocrinol. 2022;13:20420188221086693. https://doi.org/10.1177/20420188221086693.
Tang Y, Shah H, Bueno Junior CR, Sun X, Mitri J, Sambataro M, et al. Intensive Risk Factor Management and Cardiovascular Autonomic Neuropathy in Type 2 Diabetes: The ACCORD Trial. Diabetes Care. 2020;44:164–73. https://doi.org/10.2337/dc20-1842.
Serhiyenko VA, Serhiyenko AA. Cardiac autonomic neuropathy: Risk factors, diagnosis and treatment. World J Diabetes. 2018;9:1–24. https://doi.org/10.4239/wjd.v9.i1.1.
McMurray JJV, Solomon SD, Inzucchi SE, Køber L, Kosiborod MN, Martinez FA, et al. Dapagliflozin in Patients with Heart Failure and Reduced Ejection Fraction. N Engl J Med. 2019;381:1995–2008. https://doi.org/10.1056/NEJMoa1911303.
Zinman B, Wanner C, Lachin JM, Fitchett D, Bluhmki E, Hantel S, et al. Empagliflozin, Cardiovascular Outcomes, and Mortality in Type 2 Diabetes. N Engl J Med. 2015;373:2117–28. https://doi.org/10.1056/NEJMoa1504720.

No competing interests reported.

Download PDF

Journal Publication

published 11 Apr, 2024

Read the published version in Diabetes, Obesity and Metabolism →

Version 1

posted

You are reading this latest preprint version

Artificial intelligence-enhanced electrocardiogram analysis for identifying cardiac autonomic neuropathy in patients with diabetes

Status:

Journal Publication

Version 1

Abstract

Figures

Introduction

Methods

Study Design and Participants

Data Collection and study procedures

CAN Assessment Protocol

ECG Examination

Proposed Model

Pre-processing

Feature Extraction

Motif and Discord Extraction

LSTM Feature Vector Extraction

Classification

Implementation Details

Outcome and Performance Metrics

Results

Classification of dsCAN

Detection of Any Stage of CAN

Discussion

Limitations

Conclusions

Declarations

References

Additional Declarations

Status:

Journal Publication

Version 1