Heart Rate Variability measured during rest and after orthostatic challenge to detect autonomic dysfunction in Type 2 Diabetes Mellitus using the Classification and Regression Tree model

Abstract

BACKGROUND:

According to the World Health Organization, one in ten adults will have Type 2 Diabetes Mellitus (T2DM) in the next few years. Autonomic dysfunction is one of the significant complications of T2DM. Autonomic dysfunction is usually assessed by standard Ewing’s test and resting Heart Rate Variability (HRV) indices.

OBJECTIVE:

Resting HRV has limited use in screening due to its large intra and inter-individual variations. Therefore, a combined approach of resting and orthostatic challenge HRV measurement with a machine learning technique was used in the present study.

METHODS:

A total of 213 subjects of both genders between 20 to 70 years of age participated in this study from March 2018 to December 2019 at Smt. Kashibai Navale Medical College and General Hospital (SKNMCGH) in Pune, India. The volunteers were categorized according to their glycemic status as control ( $n=$ 51 Euglycemic) and T2DM ( $n=$ 162). The short-term ECG signal in the resting and after an orthostatic challenge was recorded. The HRV indices were extracted from the ECG signal as per HRV-Taskforce guidelines.

RESULTS:

We observed a significant difference in time, frequency, and non-linear resting HRV indices between the control and T2DM groups. A blunted autonomic response to an orthostatic challenge quantified by percentage difference was observed in T2DM compared to the control group. HRV patterns during rest and the orthostatic challenge were extracted by various machine learning algorithms. The Classification and Regression Tree (CART) model has shown better performance among all the machine learning algorithms. It has shown an accuracy of 84.04%, the sensitivity of 89.51%, a specificity of 66.67%, with an Area Under Receiver Operating Characteristic Curve (AUC) of 0.78 compared to resting HRV alone with 75.12% accuracy, 86.42% sensitivity, 39.22% specificity, with an AUC of 0.63 for differentiating autonomic dysfunction in non-diabetic control and T2DM.

CONCLUSION:

It was possible to develop a Classification and Regression Tree (CART) model to detect autonomic dysfunction. The technique of percentage difference between resting and orthostatic challenge HRV indicates the blunted autonomic response. The developed CART model can differentiate the autonomic dysfunction using both resting and orthostatic challenge HRV data compared to only resting HRV data in T2DM. Thus, monitoring HRV parameters using the CART model during rest and after orthostatic challenge may be a better alternative to detect autonomic dysfunction in T2DM as against only resting HRV.

Keywords

Type 2 diabetes mellitus (T2DM)Heart Rate Variability (HRV)autonomic dysfunction orthostatic challenge machine learning classification and regression tree (CART)

1. Introduction

According to the International Diabetes Federation report, around 642 million people will be affected with Type 2 Diabetes Mellitus (T2DM) in 2040 [1]. T2DM is a syndrome characterized by a high glucose level in the blood. The hyperglycemia in adults is usually due to increased insulin resistance, followed by reduced insulin secretion from beta cells of the pancreas. Uncontrolled T2DM over a period of time leads to various micro and macrovascular complications like retinopathy, nephropathy, and coronary artery disease [2, 3]. These complications are suggested to be due to concurrent autonomic dysfunction, which may precede clinical symptoms of complication. Autonomic dysfunction is estimated to be present in 60% of T2DM patients having the disease of more than five years. It is therefore crucial to detect autonomic dysfunction at an early stage in diabetes patients so that timely intervention for preventing the complications can be employed. Autonomic dysfunction is usually detected by Ewing’s battery test (CART-cardiac autonomic reflect test) [4] and resting HRV per the guidelines of the American diabetic association [5]. However, large intra and inter-individual variations in HRV and inherent nonlinearity have limited its widespread utility for detecting autonomic dysfunction. Therefore in the presented study, we employed HRV measurement both during supine resting and orthostatic challenge and used the CART machine learning algorithm to capture its non-linear characteristic.

Various researchers have studied the diagnosis of autonomic dysfunction using resting HRV parameters [6, 7, 8, 9]. In the literature survey, many authors adopted the exercise-based investigation of autonomic dysfunction diagnosis in T2DM [10]. The main focus of such investigation is to study the recovery performance of heart rate after exercise [10]. The performance of heart rate recovery areas such as atrial and ventricular repolarization after exercise was investigated by Banthia et al. [11]. Michal et al. stated that exercise could be a useful diagnostic and monitoring tool for studying the progression of the impaired autonomic nervous system (ANS), which could be a sign of cardiovascular complications [12]. Even though exercise-based analysis provides more diagnostic information, many researchers believe that HRV analysis in resting position is still a more viable technique because patients are less involved in the entire process.

Researchers are now looking into the immediate effect of a stimulus or orthostatic challenge on T2DM patients in order to detect autonomic dysfunction. As a result, we investigated resting HRV parameters and orthostatic challenge HRV parameters in the current study. However, to the best of our knowledge, this is the first study that uses the resting and orthostatic challenge HRV parameters to detect autonomic dysfunction in T2DM subjects using a machine learning approach. In the literature, only the gold standard method Ewing’s battery test and resting HRV parameters are used to detect autonomic dysfunction [4].

In the present study, a few new approaches were introduced.

1.
The new approach to find blunted response in T2DM patients in terms of percentage difference was proposed, and the direction of changes in HRV parameters was observed.
2.
Recording of HRV during rest and after orthostatic challenge to quantify the percentage difference between HRV indices in T2DM as compared to normoglycemic control.
3.
Use of machine learning to capture non-linear difference of HRV patterns measured both during rest and after orthostatic challenge in both groups.
4.
Machine learning-based low-cost, non-invasive tool to detect autonomic dysfunction in T2DM patients.
5.
The present study suggests that resting and orthostatic challenge HRV parameters are a superior diagnostic indicator of autonomic dysfunction than resting HRV alone in T2DM patients.

Our study has improved the detection of autonomic dysfunction compared to the standard available screening test.
2. Material and methods

2.1 Patients demographics

A total of 216 volunteers (52 control (Euglycemic) and 164 patients with T2DM between 20 to 70 years of both genders) participated in the study conducted at OPD of Smt. Kashibai Navale Medical College and General Hospital (SKNMGCH), Pune, India. The data of three participants were not used in the final analysis. The study was approved by the ethical committee of Smt. Kashibai Navale Medical College and General Hospital and informed consent was obtained from all participants.

The exclusion criteria were: 1) Data that falls outside of mean $\pm$ 3 standard deviations, 2) Data falls above the third quartile, below the 1 ${}^{\textit{st}}$ quartile, and outside of 1.5 times the interquartile range, and 3) Frequent ectopy in their ECG. Patients with factors affecting HRV such as thyroid disorder, drugs affecting HRV, congenital heart disease, arrhythmias, and previous myocardial infarction were not included in this study.

2.2 Data collection protocol

The criteria to carry out the experimental research were as per the guidelines of the Task Force of the European Society for Cardiology and the North American Pacing and Electrophysiology Society [13], and Standards for Reporting Diagnostic Accuracy Studies (STARD) [14].

2.3 ECG recording

ECG sampled at 1000 Hz was recorded using Chronovisor HRV-Dx 1.0.1 ambulatory monitor of all participants at the Central research lab of SKNMCGH. A standard ECG lead II procedure was used. Patients were informed of the test protocol prior to the procedure. They were advised to avoid tea, coffee, alcohol, smoking, or any other stimulant form that could affect the test for at least 24 hours before the ECG recording. The blood pressure (BP), height, and weight was recorded using standard procedure. The patients were advised to take a rest for 5 minutes before recording. After five minutes of rest, the electrode was placed as per the American Heart Association protocol [15]. Twenty-one minutes is the average time needed for the entire recording process. The ECG was recorded for fifteen minutes in a resting position, five minutes after giving an orthostatic challenge to the resting body, and a one-minute deep breathing test. The breathing rate (BR) was measured from ECG post hoc.

For the study of the HRV, the five minutes of ECG segments in resting and after orthostatic challenges were used for analysis. A tilt test at 70 degree using a motorized tilt table was used for the orthostatic challenge. The following figure shows the ECG recording in resting and after giving orthostatic stimuli.

Figure 1.

ECG recording – resting position.

Figure 2.

ECG recording – orthostatic challenge.

2.4 HRV parameters

The HRV was analyzed using Chronovisor HRV analysis software suite version 1.1.487 as per the HRV task force guideline [13]. The recorded ECG were manually tested for removal of artifacts, ectopic beats, and interpolated. The RR interval between two successive beats was evaluated in offline mode using the Chronovisor software. The HRV signal was extracted from the ECG signal and analyzed in the time domain, frequency domain, and non-linear dynamics. The time-domain analysis of HRV consists of the mean of RR interval (Mean RR), which represents the average value of RR intervals that is in milliseconds, the standard deviation of all RR interval (SDNN) represents the variance, and the root mean square differences between adjacent RR intervals (RMSSD) represents the parasympathetic modulation of the ANS [13, 16].

In the frequency domain analysis of HRV, RR time series intervals were decomposed into a frequency spectrum using the Autoregressive method (AR method) and Fast Fourier Transform method (FFT). We used the FFT method for frequency spectrum analysis. The frequency spectrum power was defined as total power (TP) in absolute (ms ${}^{2}$ ) between 0.003–0.5 Hz, reflecting the overall variations of the RR intervals. In areas of low frequency (LF) between 0.05–0.15 Hz, and high frequency (HF) between 0.15–4 Hz bands, the power spectral density of the frequency spectrum was measured in absolute (ms ${}^{2}$ ) and normalized unit (nu). LF power is a measure of sympathetic activity/baroreflex activity/sympathovagal activity. The effect of parasympathetic activity is indicated by HF power. We obtained frequency domain parameters such as TP (ms ${}^{2}$ ), LF (ms ${}^{2}$ ), HF (ms ${}^{2}$ ), LF, and HF in the normalized unit, and the LF to HF ratio through this spectral analysis [13, 16].

The non-linear dynamic approach of HRV analysis offers valuable knowledge for physiological understanding and risk assessment. A technique taken from non-linear dynamics is the Poincare plot, a graph of the current RR interval against the previous RR interval [13, 16]. Two parameters have been extracted using this method, namely standard deviation 1(SD1) representing the width. Standard deviation 2(SD2) represents the length of the Poincare geometry. SD1 referred to the short-term beat to beat variability, while SD2 specifies long-term beat to beat variability [13]. Detrended fluctuation analysis (DFA), approximate entropy (AppEn), and sample entropy (SampEn) are other major parameters in non-linear dynamics. DFA $\alpha$ 1 and DFA $\alpha$ 2 are the scaling exponents in the RR interval time series signal that reflect the short-term and long-term self-similar properties. Approximate and sample entropy measure the regularity and complexity of signal in a physiological time series. The values of entropy close to 1 are deemed to be healthy [13].

A deep breathing test consisting of deep inhalation and exhalation for six consecutive cycles was performed to obtain an Expiration/Inspiration (E: I) ratio to detect autonomic dysfunction [17, 18].

2.5 Statistical analysis

1.
A descriptive statistics and nonparametric Mann-Whitney U test was used to compare resting HRV parameters between control and T2DM using Epi. Info. 7 statistical software tool. In the present study, a $p$ -value $<$ 0.05 was considered as statistically significant. The data was presented in the form of mean $\pm$ standard deviation.
2.
The behavior of the resting system after the orthostatic challenge in both the group was calculated using the percentage difference formula. The percentage difference is given in absolute value, and a negative sign indicates the direction of changes.

$\displaystyle\textit{\% Difference in HRV parameter}=(-)\frac{\left|{\textit{% Resting}-\textit{Orthostatic Challenge}}\right|}{\frac{\left({\textit{Resting}% -\textit{Orthostatic Challenge}}\right)}{2}}\times 100$

2.6 Classification and Regression Tree (CART) model development

A decision tree is a supervised machine learning algorithm that recursively separates the data into groups or classes. The algorithms used in the decision tree are- ID3, C4.5, C5.0, CHAID, CRUISE, QUEST, and CART [19, 20, 21, 22]. A decision tree consists of a base node called a root node. The nodes connected with the root node edges are called the internal node, and the rest of the node is called a leaf node. The splitting of nodes is decided by the algorithm Information Gain, Chi-square test, and Gini Index. In this study, we used the CART decision tree algorithm which can handle both categorical values (Classification) and continuous values (Regression) [22].

Figure 3.

A typical decision tree.

The CART was first developed by Breiman in 1984 [22] and consists of binary trees. One of the advantages of CART is that it handles missing values and outliers. A binary tree is constructed by splitting a node recursively into two internal nodes, resulting in a leaf node for each split. Figure 3 shows a typical binary tree. It is supposed that dataset $D=(D_{1},D_{2}\ldots D_{n})$ is the vector of independent parameters, and $T_{i}$ and $L_{i}$ indicates the value of threshold and leaf, respectively. The tree grows from top to bottom approach until the stopping criteria or pre-defined limit is satisfied. Creating a CART model involves selecting input features, splitting criteria, stopping criteria, and pruning the tree.

2.6.1 Splitting criteria

The towing splitting criteria is used to find the best split such that the leaf node has maximum homogeneity or purity. The tree repeats the splitting process until it cannot find another best splitting node that increases the purity of the leaf node. The index used for checking the best splitting node is the Gini Index. This index indicates leaf node purity and is illustrated as:

$\displaystyle G=1-\sum_{i=1}^{n}P_{i}^{2}$ (1)

where $G$ indicates the Gini Index for all classes, and $P_{i}$ is the probability of instances with the class $i$ . For a binary problem with classification, it can be written as:

$\displaystyle G=1-\left({P_{1}^{2}-P_{2}^{2}}\right)$ (2)

The Gini Index for each node by the total number of instances in the root node is calculated as follows-

$\displaystyle G=P_{0}\left({1-P_{0}}\right)+P_{1}\left({1-P_{1}}\right)$ (3)

where the probability of instances in classes 0 and 1 are $P_{0}$ and $P_{1}$ .

2.6.2 Stopping criteria

The stopping criteria strongly influence the performance of the model or tree. The recursive binary splitting tree needs to know when to stop; otherwise, it would recursively split the instances. The most common procedure to stop splitting is to assign a pre-defined limit to each leaf node. The number of instances in the node is less than the pre-defined limit, then that node is taken as the final node. Another stopping criterion is, if the depth and purity of the nodes are greater than the pre-defined limit, then stop the splitting of the tree.

2.6.3 Pruning the tree

One of the significant difficulties in creating a decision tree model is overfitting. Overfitting occurs due to training errors. The pruning technique is used in the decision tree to prevent overfitting and reduce the size of the decision tree. The pre-pruning approach does not allow the tree to classify the training set perfectly, while the post-prune method enables the tree to classify the training set perfectly. In CART, post-pruning/cost complexity pruning can be used. The pseudocode for Classification and Regression Tree is given below (Algorithm 1).

Algorithm 1. Pseudo code of CART model
Input: Dataset $D$ , Input feature vector $D=(D_{1},D_{2}\ldots D_{n})$ , Class_label $=$ C
Start
1.	Initialize the root node $N$
2.	Find the best split that minimizes the Gini Index or Gini Impurity
3.	If pre-defined stopping criteria reach to limit, exit
	Otherwise, continue
4.	Prune the tree with post-pruning
End

In order to develop a CART model for our application to classify autonomic dysfunction in control and T2DM, the python with jupyter notebook was used. The programming is performed on a PC with an Interl ${}^{\@setsize{\scriptsize}{9.5pt}{\viiipt}{\@viiipt}\textregistered}$ Core™ i7-4790 CPU @ 3.60 GHz (8 CPUs) and 8 GB RAM. Before the training, the best hyperparameters of the CART algorithm were found out using an inner 5-fold GridSearch cross-validation approach. The optimized hyperparameters of the CART model are – Criterion $=$ ‘gini’, Splitter $=$ ‘best’, max_depth $=$ 5, min_samples_split $=$ 3, min_sample_leaf $=$ 5.

The CART machine learning algorithm was employed using HRV indices measured in both during supine rest and orthostatic challenge in both the groups (labeled as control and T2DM). All the resting and resting $+$ orthostatic challenge HRV parameters of control and T2DM subjects were given as input to the CART machine learning model.

2.7 Performance evaluation metrics

In this study, we have developed a CART machine learning model to detect autonomic dysfunction. The training and testing phase was carried out with 10-fold cross-validation. The data chosen for training is 80 percent and 20 percent for testing for every ten iterations. The performance metrics such as accuracy (ACC), sensitivity (SEN), specificity (SPC), and area under the receiver operating characteristic curve (AUC) were used to test the performance of the CART algorithm. The performance metrics were calculated using the confusion matrix, and the formulas are given in Table 1.

AUC is another important technique for evaluating algorithm performance. The AUC analysis measured the area under the receiver operating characteristic curve (ROC). The AUC near 1 indicates the perfect prediction, while the AUC near to 0.5 indicates the worst prediction. The appropriate AUC range is between 0.70 and 1 [23].

3. Results

3.1 Comparison of resting HRV between control and T2DM

3.1.1 Time-domain parameters of control and T2DM

The parameters analyzed in time domain measures included Mean RR, SDNN, RMSSD, and BR. In the resting position, Mean RR and SDNN significantly decreased in T2DM compared to control, and BR was significantly high, as shown in Table 2.

Table 1
Performance metrics

Performance metrics	Description
ACC	(TP $+$ TN)/(TP $+$ FP $+$ TN $+$ FN)
SEN	(TP)/(TP $+$ FN)
SPC	(TN)/(TN $+$ FP)
AUC	The area under the ROC curve

TP: True Positive, TN: True Negative, FP: False Positive, FN: False Negative.

Table 2

Time-domain parameters between control and T2DM during lying resting position

HRV measures	Control resting ( $n=$ 51)	T2DM resting ( $n=$ 162)	$p$ -value inference
Mean RR	838.79 $\pm$ 110.06	769.02 $\pm$ 115.3	$<$ 0.001
SDNN	70.63 $\pm$ 32.21	31.91 $\pm$ 14.42	$<$ 0.001
RMSSD	73.2 $\pm$ 50.46	68.25 $\pm$ 37.23	0.870
BR	12.91 $\pm$ 2.97	15.26 $\pm$ 3.21	$<$ 0.001

3.1.2 Frequency-domain parameters of control and T2DM

The parameters analyzed in frequency domain measures included TP (ms ${}^{2}$ ), power of LF, HF in ms ${}^{2}$ , and normalized unit (nu), and the ratio of LF to HF. In the resting position, TP, LF, and HF power, LF: HF ratio were significantly reduced, as shown in Table 3.

Table 3
Frequency domain parameters between control and T2DM during lying resting position

HRV measures	Control resting ( $n=$ 51)	T2DM resting ( $n=$ 162)	$p$ -value inference
TP (ms ${}^{2}$ )	3183.28 $\pm$ 2688.49	755.04 $\pm$ 836.24	$<$ 0.001
LF (ms ${}^{2}$ )	783.36 $\pm$ 632.34	175.58 $\pm$ 247.75	$<$ 0.001
HF (ms ${}^{2}$ )	884.46 $\pm$ 948.89	221.85 $\pm$ 357.95	$<$ 0.001
LF (nu)	52.17 $\pm$ 17.68	46.33 $\pm$ 15.78	$<$ 0.001
HF (nu)	47.83 $\pm$ 17.68	57.2 $\pm$ 54.76	$<$ 0.001
LF/HF	1.55 $\pm$ 1.66	1.12 $\pm$ 1.01	0.041

3.1.3 Non-linear parameters of control and T2DM

Poincare plot parameters SD1 and SD2, DFA- $\alpha$ 1 and DFA- $\alpha$ 2, AppEn, and SampEn were analyzed in control and T2DM subjects. The Poincare plot measures SD1, and SD2 was significantly low in the resting position, as shown in Table 4. DFA- $\alpha$ 1 and DFA- $\alpha$ 2, AppEn, and SampEn were not statistically different between the control and T2DM subjects.

Table 4
Non-linear parameters between control and T2DM during lying resting position

HRV measures	Control resting ( $n=$ 51)	T2DM resting ( $n=$ 162)	$p$ -value inference
SD1	40.92 $\pm$ 25.45	21.94 $\pm$ 12.29	$<$ 0.001
SD2	89.83 $\pm$ 40.19	38.92 $\pm$ 18.19	$<$ 0.001
DFA $\alpha$ 1	0.91 $\pm$ 0.22	0.84 $\pm$ 0.23	0.059
DFA $\alpha$ 2	0.93 $\pm$ 0.24	0.90 $\pm$ 0.18	0.273
AppEn	1.13 $\pm$ 0.18	1.14 $\pm$ 0.21	0.288
SampEn	1.43 $\pm$ 0.32	1.36 $\pm$ 0.34	0.219

3.2 Orthostatic challenge (stimuli) HRV parameters comparison between control and T2DM

As most of the studies of HRV in diabetes were done in resting conditions only, we have also given an orthostatic challenge/stimulus and noted the response of ANS to that orthostatic challenge measured by HRV [11]. Orthostatic challenges usually lead to increased heart rate, decreased variance, and increased blood pressure, which is the normal physiological response. The results of orthostatic challenges are shown in Table 5. The percentage difference in direction between control and T2DM after the orthostatic challenge is shown in Fig. 4.

Table 5
Percentage difference in resting and orthostatic challenge HRV parameters of control and T2DM subjects

HRV

measures

Control resting (

n=

51)

Control orthostatic challenge (

n=

51)

Control %

difference in

resting and

Orthostatic

challenge

Diabetic resting (

n=

162)

Diabetic orthostatic challenge (

n=

162)

T2DM %

difference in

resting and

orthostatic

challenge

Mean

Mean RR

838.79

110.06

698.92

88.67

-

18.19

-

21.53

769.02

115.3

688.72

112.01

-

11.02

-

2.89

SDNN

70.63

32.21

55.17

19.77

-

24.58

-

47.86

31.91

14.42

32.15

13.46

0.75

-

6.89

RMSSD

73.20

55.49

38.73

-

27.52

-

61.59

68.25

37.23

71.55

38.78

4.72

4.08

12.91

2.97

12.82

-

0.70

1.01

15.26

3.21

14.91

3.26

-

2.32

1.55

TP (ms ${}^{2}$ )

3183.3

2688.5

1871.1

1641.2

-

51.92

-

48.38

755.04

836.2

617.18

681.71

-

20.09

-

20.36

LF (ms

{}^{2}

)

783.36

632.34

716.68

691.24

-

8.89

8.90

175.58

247.8

149.66

189.34

-

15.94

-

26.73

HF (ms ${}^{2}$ )

884.46

948.89

302.68

423.13

-

98.01

-

76.64

221.85

358

120.81

204.69

-

58.97

-

54.48

LF (nu)

52.17

17.68

73.13

15.06

33.46

-

16.00

46.33

15.78

53.84

17.53

14.99

10.51

HF (nu)

47.83

17.68

26.87

15.06

-

56.12

-

16.00

57.20

54.76

46.22

17.52

-

21.23

-

103.04

LF: HF

1.55

1.66

4.12

3.23

90.65

64.21

1.12

1.01

1.94

2.73

53.59

91.98

SD1

40.92

25.45

23.58

13.81

-

53.77

-

59.30

21.94

12.29

19.07

11.12

-

14.00

-

10.00

SD2

89.83

40.19

73.66

25.80

-

19.78

-

43.61

38.92

18.19

39.99

17.64

2.71

-

3.07

DFA

\alpha

0.91

0.22

1.23

0.20

29.91

-

9.52

0.84

0.23

1.36

5.28

47.27

183.30

DFA

\alpha

0.93

0.24

0.89

0.16

-

4.40

-

40.00

0.90

0.18

0.97

0.21

7.49

15.38

AppEN

1.13

0.18

0.97

0.19

-

15.24

5.41

1.14

0.21

0.93

0.26

-

20.29

21.28

SampEN

1.43

0.32

1.13

0.27

-

23.44

-

16.95

1.36

0.34

1.04

0.34

-

26.67

0.00

Figure 4.

Direction of changes in HRV parameters.

4. Discussion

4.1 Findings

In this study, we initially studied the ANS response using resting HRV parameters. We found that the resting time (Mean RR interval, SDNN), frequency-domain HRV parameters (absolute power of TP, LF, HF), and non-linear parameters SD1 and SD2 of T2DM are significantly reduced than the resting HRV in the control group. No significant difference in resting non-linear features DFA $\alpha$ 1 and $\alpha$ 2, AppEn, and SampEn was observed between the two groups. Clinically, the use of resting HRV for the detection of autonomic dysfunction is limited due to large intra and inter-individual variations in resting HRV [24, 25, 26, 27]. Therefore we made an additional exploration in the present study. We used the orthostatic challenge as a stimulus and studied the ANS reactivity of the system using HRV in both groups. A blunted response of ANS reactivity was observed in T2DM. We quantified the blunted response by measuring the percentage difference between resting and orthostatic challenge HRV parameters, as shown in Table 5.

The percentage difference and direction of changes in control and T2DM subjects after orthostatic challenge leads to the following findings:

1.
SDNN, TP (ms ${}^{2}$ ), and SD2 (long-term variability) reflect the overall variance. The percentage difference between resting and orthostatic challenge in SDNN is $-$ 24.52% in control compared to 0.75% in T2DM. The percentage difference in TP (ms ${}^{2}$ ) is $-$ 51.92% in control compared to $-$ 20.09% in T2DM. The percentage change in SD2 is $-$ 19.78% in control compared to 2.71% in T2DM. It indicates less reduction in the variance in T2DM exposed to orthostatic challenge indicating blunted autonomic response in T2DM.
2.
In control, resting HF (ms ${}^{2}$ ) 884.46 (ms ${}^{2}$ ) is reduced to 302.68 (ms ${}^{2}$ ) when challenged with orthostatic stimuli. It implies the withdrawal of the vagal during an orthostatic challenge, which is a normal physiological response. However, the percent difference in HF (ms ${}^{2}$ ) is only $-$ 58.97% in diabetes compared to $-$ 98.01% in the control group, indicating that the optimum vagal withdrawal required during an orthostatic challenge is not occurring in T2DM, reflecting an altered response. The loss of effective vagal withdrawal may lead to dizziness and hypotension with an orthostatic challenge, as seen clinically in patients with autonomic dysfunction in T2DM.
3.
Similarly, the percentage difference in LF: HF ratio to the orthostatic challenge is 90.65% in the Control group compared to 53.59% in T2DM. The percent difference in LF to HF ratio can occur because of either increase in LF (ms ${}^{2}$ ) or a decrease in HF (ms ${}^{2}$ ) power, but we have not found much change in LF (ms ${}^{2}$ ). In the present study, the change in LF: HF ratio could be mediated by the HF component of HRV.

As the vagal withdrawal is seen to be less in T2DM, the percentage difference in the LF: HF ratio is blunted.

In the present study, we employed a Classification and Regression Tree machine learning algorithm to explore the inherent non-linear characteristics of HRV, which is difficult to capture by routine linear statistics. The alogirthm will also take care of the intra and inter-individual variations that are observed in resting HRV. The HRV indices were measured both in resting and during the orthostatic challenge and subjected to machine learning algorithms. The machine was trained with resting HRV indices alone and resting HRV plus HRV indices measured during the orthostatic challenge. These non-linear characteristics may differ in disease and non-disease states.
4.2 Implementation of classification and regression tree (CART) model

We took resting and resting $+$ orthostatic challenge HRV parameters with their label (Control and T2DM), and gave them as input to the CART model. We observed that HRV parameters of resting alone have an accuracy of 75.12%, a sensitivity of 86.42%, and specificity of 39.22%, and AUC is 0.63 (Table 6). The decision tree, feature importance graph (i.e. he importance of each parameter in the development of the decision tree-based model), and AUC curve using resting HRV data are depicted in Figs 5–7, respectively.

Table 6
Resting HRV parameter performance to detect autonomic dysfunction

HRV parameters (only resting)	TN	FP	FN	TP	ACC	SEN	SPC	AUC
All HRV Parameter	20	31	22	140	75.12%	86.42%	39.22%	0.63

Figure 5.

Decision tree using resting HRV parameters.

Table 7

Resting and orthostatic challenge HRV parameter performance to detect autonomic dysfunction

HRV parameters (Resting $+$ Orthostatic challenge)	TN	FP	FN	TP	ACC	SEN	SPC	AUC
All HRV Parameter	34	17	17	145	84.04%	89.51%	66.67%	0.78

Figure 6.

Important resting HRV parameters.

Figure 7.

ROC using resting HRV parameters.

When HRV parameters were taken together during rest and orthostatic challenges, we observed accuracy of 84.04%, sensitivity of 89.51%, and the specificity of 66.67% with AUC 0.78 (Table 7). The decision tree, feature importance graph, and AUC curve using resting and orthostatic challenge HRV data can be visualized in Figs 8–10, respectively.

Table 8

Resting HRV parameter performance of various machine learning models

Machine learning model	TN	FP	FN	TP	ACC	SEN	SPC	AUC
Logistic regression	21	34	26	131	71.70%	83.44%	38.18%	0.60
Naïve bayes	19	32	28	134	71.83%	82.37%	37.25%	0.59
Support vector machine	18	38	25	132	70.42%	84.08%	32.14%	0.58
Artificial neural network	20	30	29	134	72.30%	82.21%	40%	0.61

Figure 8.

Decision tree using resting $+$ orthostatic challenge HRV parameters.

It was noted that all the performance metrics of the CART model were improved by considering orthostatic challenge HRV data along with the resting HRV data. Precisely the specificity of the model was shifted from 39.22% to 66.67%. Based on our results, we can suggest that the machine learning approach using resting as well as orthostatic challenge HRV indices can segregate diseased v/s control groups more efficiently than resting HRV alone.

4.3 Performance comparison of CART model with other machine learning models

In this study, the performance of the CART model was compared with the various machine learning algorithms like Logistic Regression (LOR) [28], Naïve Bayes (NB) [29], Support Vector Machine (SVM) [30], and Artificial Neural Network (ANN) [31]. The performance of the machine learning models are shown in Tables 8 and 9.

Table 9
Resting and orthostatic challenge HRV parameter performance of various machine learning models

Machine learning model	TN	FP	FN	TP	ACC	SEN	SPC	AUC
Logistic regression	29	21	23	140	79.34%	85.89%	58%	0.71
Naïve bayes	31	20	20	142	81.22%	87.65%	60.78%	0.74
Support vector machine	28	24	23	138	77.93%	85.71%	53.85%	0.69
Artificial neural network	30	21	19	143	81.22%	88.27%	58.82%	0.73

Figure 9.

Important resting $+$ orthostatic challenge HRV parameters.

Figure 10.

ROC using resting $+$ orthostatic challenge HRV parameters.

It was observed that among all the machine learning models, the CART model (Tables 6 and 7) has the highest classification accuracy, sensitivity, and specificity. Therefore we have emphasized more on the CART model.

4.4 Computational complexity analysis

The computational complexities of ML Models are Time Complexity and Space complexity. The time complexity of an algorithm can be defined as how quickly or slowly it performs for a given input size. The input features are often used to determine the time complexity. The amount of additional memory needed to run your algorithm is known as space complexity. It’s expressed in terms of input features, much like time complexity. Time complexity is associated with training time and prediction time. The training time for the algorithm is 9 milliseconds, while the prediction time is 6 milliseconds.The time complexity and space complexity may vary with the computational configuration.

4.5 Overview and comparison with previous relevant studies

CARTs (cardiovascular autonomic reflex tests) are still the gold standard for detecting autonomic dysfunction in T2DM patients [4]. CARTs are dynamic heart rate (HR) and blood pressure (BP) tests that detect parasympathetic and sympathetic dysfunctions in the ANS. HRV is a non-invasive method for studying autonomic function in both healthy and diseased individuals. According to a recent study, HRV at rest can accurately diagnose autonomic dysfunction in T2DM patients. HRV at rest is widely used due to its ease of use; however, due to a lot of intra and inter-variation in HRV parameters, analyzing it during resting conditions may not provide complete information about autonomic dysfunction [32]. Resting HRV only measures static autonomic activity and does not evaluate dynamic autonomic control. Experimental evidence shows that the ANS is more responsive to sudden stimulation. Thus dynamic evaluation methods such as sudden stimulus or orthostatic challenge may precisely indicate autonomic dysfunction [12].

In contrast to Ewing’s test battery, an earlier study by author Tank et al. found HRV to be a less reliable test for diagnosing autonomic dysfunction. However, they came to this conclusion by using a predetermined cut-off value of the HRV parameters [6]. Few previous studies also found that resting HRV performed similarly to Ewing’s battery test in assessing ANS activity [33, 34, 35]. The authors Virtanen et al. measured HRV parameters at rest, during exercise, and after exercise for detecting autonomic dysfunction. In comparison to HRV parameters measured at rest and during exercise, they found that HRV parameters measured after exercise are a better diagnostic indicator of autonomic dysfunction [36]. Khandoker et al. found that resting non-linear HRV measures like sample entropy and SD1/SD2 ratio are more specific and sensitive indicators of autonomic dysfunction in T2DM than conventional resting linear time and frequency domain measures [9]. The author Vinik et al. concluded that resting HR, BP responses to standing, and time and frequency parameters of HRV in response to deep breathing, standing, and the Valsalva maneuver, are provoked in T2DM, which may give an early indication of autonomic dysfunction [37]. Several studies suggest that abnormal non-linear HRV parameters are associated with diabetes, and that non-linear HRV parameters in diabetics may have prognostic and predictive value for detecting asymptomatic cardiovascular diseases and autonomic dysfunction [38, 39]. Recent research on the Chinese population using short-term HRV under resting conditions found that resting HRV parameters were not inferior to the standard Ewing’s battery test for detecting autonomous dysfunction [8]. In the literature, many studies have shown the link of resting HRV parameters to detect autonomic dysfunction. Although, to our knowledge, no study has been done on T2DM patients to detect autonomic dysfunction by combining the resting HRV and orthostatic challenge HRV parameters.

4.6 Strength and limitations

The present study has a potential strength. Both ECG recordings and HRV analyses were carried out in compliance with the recommendations of the task force. In this study, the HRV time domain, frequency domain, and non-linear method were analyzed in resting as well after orthostatic challenge. The study was performed on real and authentic data. To the best of our knowledge, this is the first study that defines the percentage difference between the resting and orthostatic challenge HRV parameters of the control and T2DM group. Unequal group sizes, with less numbers in the control group, is the main limitation of the study. This may have affected the low specificity that we observed in the present study.

5. Conclusion

The reduced resting HRV, altered autonomic response to orthostatic challenge measured by percentage difference of HRV indices, the difference in resting and orthostatic challenge HRV patterns in non-diabetic controls, and T2DM patients are sensitive markers to screen autonomic dysfunction associated with Type II Diabetes Mellitus. The CART machine learning model was implemented using resting and orthostatic challenge HRV parameters to detect autonomic dysfunction, and results are desirable. The present study suggests that resting and orthostatic challenge HRV parameters are a superior diagnostic indicator of ANS dysfunction than resting HRV alone in T2DM patients. Our research will pave the way for future research, specifically using the machine learning approach for screening autonomic dysfunction in Type II Diabetes Mellitus.

Footnotes

Conflict of interest

The authors declare no conflict of interest.

Funding

No funding reported.

References

Ogurtsova

da Rocha Fernandes

Huang

Linnenkamp

Guariguata

Cho

Cavan

Shaw

Makaroff

. IDF Diabetes Atlas: Global estimates for the prevalence of diabetes for 2015 and 2040. Diabetes Research and Clinical Practice. 2017; 128: 40-50.

Wadhokar

Phadke

Bhat

. Heart rate variability indices in patients with micro-and macrovascular complications of type 2 diabetes: A cross-sectional study. Diabetologia. 2018; 61: 516-516.

Shah

Vajravelu

Bacha

Farrell

Gidding

Katz

LEL

Tryggestad

White

Urbina

. Heart Rate Variability and Cardiac Autonomic Dysfunction: Prevalence, Risk Factors, and Relationship to Arterial Stiffness in the Treatment Options for Type 2 Diabetes in Adolescents and Youth (TODAY) Study. Diabetes Care, 2019; 42(11): 2143-2150.

Ewing

Clarke

. Diagnosis and management of diabetic autonomic neuropathy. British Medical Journal (Clinical Research ed.). 1982; 285: 916-918.

American Diabetes Association. 2. Classification and diagnosis of diabetes. Diabetes Care. 2015; 38(Supplement 1): 8-16.

Tank

Jordan

Neuke

Mölle

Weck

. Spontaneous baroreflex sensitivity and heart rate variability are not superior to classic autonomic testing in older patients with type 2 diabetes. The American Journal of the Medical Sciences. 2001; 322(1): 24-30.

ChuDuc

NguyenPhan

NguyenViet

. A review of heart rate variability and its applications. APCBEE Procedia. 2013; 7: 80-5.

Chen

Yang

Liu

Tang

. Diagnostic performance analysis for diabetic cardiovascular autonomic neuropathy based on short-term heart rate variability using Bayesian methods: preliminary analysis. Diabetology & Metabolic Syndrome. 2015; 7(1): 1-8.

Khandoker

Jelinek

Palaniswami

. Identifying diabetic patients with cardiac autonomic neuropathy by heart rate complexity analysis. Biomedical Engineering Online. 2009; 8(1): 1-2.

10.

Sacre

Jellis

Coombes

Marwick

. Diagnostic accuracy of heart-rate recovery after exercise in the assessment of diabetic cardiac autonomic neuropathy. Diabetic Medicine. 2012; 29(9): e312-20.

11.

Banthia

Bergner

Chicos

Pelchovitz

Subacius

Kadish

Goldberger

. Detection of cardiovascular autonomic neuropathy using exercise testing in patients with type 2 diabetes mellitus. Journal of Diabetes and Its Complications. 2013; 27(1): 64-9.

12.

Michael

Graham

Davis

. Cardiac autonomic responses during exercise and post-exercise recovery using heart rate variability and systolic time intervals – a review. Frontiers in Physiology. 2017; 8: 301.

13.

Electrophysiology

. Heart rate variability: standards of measurement, physiological interpretation, and clinical use. Circulation. 1996; 93(5): 1043-65.

14.

Cohen

Korevaar

Altman

Bruns

Gatsonis

Hooft

Irwig

Levine

Reitsma

De Vet

Bossuyt

. STARD 2015 guidelines for reporting diagnostic accuracy studies: explanation and elaboration. BMJ Open. 2016; 6(11): 1-17.

15.

Drew

Califf

Funk

Kaufman

Krucoff

Laks

Macfarlane

Sommargren

Swiryn

Van Hare

. Practice standards for electrocardiographic monitoring in hospital settings: an American Heart Association scientific statement from the Councils on Cardiovascular Nursing, Clinical Cardiology, and Cardiovascular Disease in the Young: endorsed by the International Society of Computerized Electrocardiology and the American Association of Critical-Care Nurses. Circulation. 2004; 110(17): 2721-46.

16.

Acharya

Joseph

Kannathal

Lim

Suri

. Heart rate variability: a review. Medical and Biological Engineering and Computing. 2006; 44(12): 1031-51.

17.

Rivera

Estañol

Fossion

Toledo-Roy

Callejas-Rojas

Gien-López

Delgado-García

Frank

. Loss of breathing modulation of heart rate variability in patients with recent and long standing diabetes mellitus type II. PloS one. 2016; 11(11): e0165904.

18.

McCraty

Atkinson

Dispenza

. One-minute deep breathing assessment and its relationship to 24-h heart rate variability measurements. Heart and Mind. 2018; 2(3): 70.

19.

Quinlan

. Induction of decision trees. Machine Learning. 1986; 1(1): 81-106.

20.

Weber

. Fuzzy, ID3: a class of methods for automatic knowledge acquisition, In proceedings of the 2nd International Conference on Fuzzy Logic and Neural Networks, Iizuka, Japan, 1992, pp. 265-268.

21.

Quinlan

. C4. 5: programs for machine learning. Elsevier. 2014; 1-299.

22.

Breiman

Friedman

Olshen

Stone

. Classification and regression trees. Monterey, CA: Wadsworth & Brooks/Cole Advanced Books & Software, 1984: 358.

23.

Hajian-Tilaki

. Receiver operating characteristic (ROC) curve analysis for medical diagnostic test evaluation. Caspian Journal of Internal Medicine. 2013; 4(2): 627-635.

24.

Abdalrada

Abawajy

Chowdhury

Rajasegarar

Al-Quraishi

Jelinek

. Meta learning ensemble technique for diagnosis of cardiac autonomic neuropathy based on heart rate variability features. International Conference on Computer Applications in Industry and Engineering, CAINE. 2017; 169-175.

25.

Fisher

Tahrani

. Cardiac autonomic neuropathy in patients with diabetes mellitus: current perspectives. Diabetes, Metabolic Syndrome and Obesity: Targets and Therapy. 2017; 10: 419.

26.

Bhati

Moiz

Naqvi

Hussain

. Diagnostic performance of resting and post-exercise heart rate variability for detecting cardiac autonomic neuropathy in type 2 diabetes mellitus. Autonomic Neuroscience. 2019; 219: 53-65.

27.

Sambyal

Saini

Syal

. Microvascular Complications in Type-2 Diabetes: A Review of Statistical Techniques and Machine Learning Models. Wireless Personal Communications. 2020; 12: 1-26.

28.

Hassan

Butt

Baba

. Logistic regression versus neural networks: the best accuracy in prediction of diabetes disease, Asian J Comp Sci Technol. 2017; 6(2): 33-42.

29.

Rish

. An empirical study of the naive Bayes classifier. InIJCAI 2001 Workshop on Empirical Methods in Artificial Intelligence. 2001; 3(22): 41-46.

30.

Ding

Tan

. An overview on theory and algorithm of support vector machines. Journal of University of Electronic Science and Technology of China. 2011; 40(1): 2-10.

31.

Al-Shayea

. Artificial neural networks in medical diagnosis. International Journal of Computer Science Issues. 2011; 8(2): 150-4.

32.

Franca da Silva

Penachini da Costa de Rezende Barbosa

Marques Vanderlei

Destro Christofaro

Marques Vanderlei

. Application of heart rate variability in diagnosis and prognosis of individuals with diabetes mellitus: Systematic review. Annals of Non-Invasive Electrocardiology. 2016; 21(3): 223-35.

33.

Vinik

Ziegler

. Diabetic cardiovascular autonomic neuropathy. Circulation. 2007; 115(3): 387-97.

34.

Neil

Thompson

John

McCarthy

Mann

. Diabetic autonomic neuropathy: the prevalence of impaired heart rate variability in a geographically defined population. Diabetic Medicine. 1989; 6(1): 20-4.

35.

Risk

Bril

Broadbridge

Cohen

. Heart rate variability measurement in diabetic neuropathy: review of methods. Diabetes Technology & Therapeutics. 2001; 3(1): 63-76.

36.

Vinik

Casellini

Parson

Colberg

Nevoret

. Cardiac autonomic neuropathy in diabetes: a predictor of cardiometabolic events. Frontiers in Neuroscience. 2018; 12: 591.

37.

Roy

Ghatak

. Non-linear methods to assess changes in heart rate variability in type 2 diabetic patients. Arquivos Brasileiros de Cardiologia. 2013; 101(4): 317-27.

38.

Silva-e-Oliveira

Amélio

Abranches

Damasceno

Furtado

. Heart rate variability based on risk stratification for type 2 diabetes mellitus. Einstein (Sao Paulo). 2017; 15(2): 141-7.

39.

de Godoy

. Non-linear analysis of heart rate variability: a comprehensive review. Journal of Cardiology and Therapy. 2016; 3(3): 528-33.