Empirical Analysis of Apnea Syndrome Using an Artificial Intelligence-Based Granger Panel Model Approach.

Sleep apnea is a serious sleep disorder that occurs when a person's breathing is interrupted during sleep. People with untreated sleep apnea stop breathing repeatedly during their sleep. This study provides an empirical analysis of apnea syndrome using the AI-based Granger panel model approach. Data were collected from the MIT-BIH polysomnographic database (SLPDB). The panel is composed of eighteen patients, while the implementation was done using MATLAB software. The results show that, for the eighteen patients with sleep apnea, there was a significant relationship between ECG-blood pressure (BP), ECG-EEG, and EEG-blood pressure (BP). The study concludes that the long-term interaction between physiological signals can help the physician to understand the risks associated with these interactions. The study would assist physicians to understand the mechanisms underlying obstructive sleep apnea early and also to select the right treatment for the patients by leveraging the potential of artificial intelligence. The researchers were motivated by the need to reduce the morbidity and mortality arising from sleep apnea using AI-enabled technology.

1. Introduction

Biomedical signal analysis is becoming increasingly important in the development of medical therapeutic strategies [1-4]. With the development of information technology and numerical calculation, it has become interesting to integrate a diagnostic aid approach into an automatic calculation process. The choice of mathematical models derived from the signals for the characterization of a given pathology becomes crucial.

By considering risk factors in a comprehensive way, it is possible to determine the probabilities of a patient's suffering or death from a given disease. However, such an assessment does not constitute a guarantee. For example, a person at high risk may live for a very long time, while a person at low risk may have a heart attack. However, people with diabetes, rheumatoid arthritis, or cardiovascular disease are at high risk [1, 3, 4].

When assessing risk factors, the doctor can determine what actions the patient should take, such as quitting smoking, losing weight, eating healthy, and getting more exercise, to reduce risk. In this work, we will expand the analysis of physiological signals based on another mathematical model, belonging to the same Granger family, which is the study of panel data cointegration and short- and long-term quantification if it exists (fully modified OLS (FMOLS) and dynamic OLS (DOLS) estimators).

In this context, we propose an empirical contribution focusing on a phenomenon that has been very popular in recent years, which is the study of sleep apnea.

Patients with sleep disorders should have a clear description of their sleep behavior. It has been shown that the interaction between cardiac autonomic activity and sleep has been studied to explain this increased incidence [5, 6]. The problem related to sleep is called sleep apnea.

Panel data models have been very enthusiastic over the past twenty years. This enthusiasm has resulted in a real explosion in the number of academic studies based on panel data models. The aspects of the transposition of issues from time series to panels are detailed in the following. Also known as the two-dimensional structure, the panel study provides more information than that available in time series [7]. Indeed, it is a particularly valuable statistical source for the analysis of dynamic behaviors.

Certainly, panel data models have many advantages, but they no longer seem sufficient to study all phenomena, especially since our case lies in the study of physiological signals. We must therefore consider the latest developments in panel data in terms of multivariety and nonstationarity in order to correctly estimate our results. There are a number of nonlinear models for panel data, including

Pooled models

Fixed effect models

✓ Estimation of fixed-effect models

✓ Tests for the existence of fixed effects

Random effects models

✓ Estimation of random-effects models

✓ Hausmann tests

Probit and Logit

Tobit I and II

Cointegration on panel data

In the remainder of this article, we will look at this last method to evaluate our contributions. We will limit our study to panel data cointegration models, FM-OLS and DOLS estimators, and panel Granger causality. To achieve the above set objective, we used a multimethodological strategy, an analysis using the panel data method that allowed us to exploit the individual and temporal dimensions.

The analysis approach used follows these steps.

Unit root tests

Panel cointegration

FM-OLS and DOLS

Granger's cause on a panel

The base classifier used is one of the most effective classification algorithms. We chose to use it because of the need for speedy and accurate analysis and predictions. The efficiency and precision are high as seen in this study.

The paper is organized as follows: introduction, cointegration approaches, recapitulation of the analysis procedure, obstructive sleep apnea syndrome, results, and then followed by the conclusion section.

2. Cointegration Approaches

There are a number of tests for panel cointegration. These include Kao [8], Bai and Ng [9], Mackoskey and Kao [10], Westerlund [11-14], Westerlund and Edgerton [15], Hank [16, 17], Gengenbach et al. [18], Gutierrez [19], and the Pedroni tests [20, 21]. In our study, and given the length of the important time dimension of the data, we chose to apply Pedroni's approaches.

3. Recapitulation of the Analysis Procedure

A method of presentation must be chosen after carefully weighing the advantages and disadvantages of different methods of presentation. In this section, we will present an organizational chart (Figure 1), summarizing the methods used. An organizational chart (often called an organization chart, org chart, organigram (me), or organogram) is a diagram that shows the structure of an organization and the relationships and relative ranks of its parts. These techniques are the most effective for simplifying the analysis.

Figure 1

Flowchart describing the analytical method.

4. Obstructive Sleep Apnea Syndrome

It is a disorder characterized by an interruption of breathing that disturbs sleep. It is one of the most common diseases among adults, affecting up to 5% of women and 15% of men between the ages of 30 and 60. During an apnea attack, the oxygen concentration in the body can decrease significantly and the carbon dioxide concentration increase. The heart must therefore work harder to compensate for this imbalance. Thus, sleep apnea is considered a fatal disease. Whenever there is an interruption in breathing, the brain sends a signal to wake the person and they start breathing again. These people never manage to get a good night's sleep, and this causes them excessive drowsiness, hypoxemia, and Olten's hypercapnia [22].

The causality between sleep Apnea syndrome and cardiovascular morbidity has remained controversial for many years [23, 24]. Factors contributing to sleep apnea include male sex, age (over 60 years of age), overweight (BMI greater than 27 kg/m2), alcohol or taking sleeping pills before bedtime, smoking, and respiratory diseases. Certain anatomical characteristics of the bones of the head can also promote sleep apnea such as jaws and palates that are too narrow, too deep, a nasal cavity that is too small, a chin that is placed too far back, etc. In addition, type 2 diabetes, hypertension, or hypothyroidism are more common in people with sleep apnea, cardiovascular disease (stroke), and neurodegenerative diseases. ECG/EEG monitoring during sleep can be promising.

We examined the panel data causality of a number of physiological signals derived from the database described above. The approach was then applied to ECG, EEG (C4-A1), BP (blood pressure), and RESP (respiratory impedance). These signals are ideal for understanding causality. The symptoms of mainly chronic diseases are silent for a long time; this is why we propose this type of test to predict a patient's future state based on the interactions that exist between his physiological signals. Then, we devote a complete study to this problem, where we evaluated several causality tests.

5. Results

In the empirical part of this paper, we conducted an investigation using a mathematical model to determine whether there is a long-term relationship between cardiological and neurological aspects and pulmonary hemodynamic signals during sleep. This panel is composed of eighteen patients taken from the MIT-BIH polysomnographic database (SLPDB) [25]. Figure 2 shows several physiological signals over a period of 10 s from the SLPDB database [25]. These signals are as follows:

Figure 2

Physiological signals presentation for 10 s from the SLPDB database [25].

ECG: this is a graphical representation of the electrical activity of the heart.

BP (blood pressure): this is the pressure of blood in the arteries. It is also defined as the force exerted by blood on the walls of arteries.

EEG (C4-A1): this is a brain exploration method that measures the electrical activity of the brain through electrodes placed on the scalp. The results are often represented in the form of a trace called an electroencephalogram. There are a number of encephalographic derivatives and the choice of C4-A1 is considered to analyze sleep disorders, and in addition, it scans all other regions (from the right central lobe to the left atrial point) [26].

RESP (respiration): breathing refers to both the gaseous exchanges resulting from the inspiration and exhalation of air.

To study the stationarity of our series, we used unit root tests based on panel data (Levin Lin and Chu, IM Pesaran and Shin, Breitung, Maddala, and Wu, as well as the Hadri and the heteroscedasticity test).

The stationarity tests applied to our physiological series allow us to determine the stationarity or nonstationarity state of our panel. To be able to quantify this, there are a number of tests based on the values of the corresponding probability for each signal. Noting that during our analyses, we will rely much more on the two tests LLC and Hadri because of their performance compared to the others. Thus, we can talk about the panel's nonstationarity if one of these two tests verifies the condition (probability <0.01 in level, probability >0.01 in first difference) for LLC and Hadri, because it stimulates the opposite hypotheses. We start the second part relating to panel cointegration.

After checking the nonstationarity properties for all panel variables, we investigate the existence of a long-term relationship between these variables by applying Pedroni's cointegration tests, which are based on unit root tests on estimated residues, trying to test cointegration for ECG and EEG signals. Tables 1 and 2 represent, respectively, Pedroni cointegration for ECG and EEG signals.

Table 1

Panel cointegration tests for ECG.

	Within dimension (panel statistics)			Between dimension (individuals statistics)
Methods	Test	Statistics	Prob	Test	Statistics	Prob
Pedroni [21]	Panel v-statistic	33.08050	0.0000	Group ρ-statistic	−145.7978	0.0000
	Panel rho-statistic	−290.7709	0.0000	Group pp-statistic	−58.83045	0.0000
	Panel PP-statistic	−88.35383	0.0000	Group ADF-statistic	−137.3478	0.0000
	Panel ADF-statistic	−97.02370	0.0000
Pedroni [20] weighted statistic	Panel v-statistic	15.88081	0.0000
	Panel rho-statistic	−109.5049	0.0000
	Panel PP-statistic	−52.30408	0.0000
	Panel ADF-statistic	−117.2825	0.0000

Table 2

Panel cointegration tests for EEG.

	Within dimension (panel statistics)			Between dimension (individuals statistics)
Methods	Test	Statistics	Prob	Test	Statistics	Prob
Pedroni [21]	Panel v-statistic	53.95472	0.0000	Group ρ-statistic	−7960.843	0.0000
	Panel rho-statistic	−5387.902	0.0000	Group pp-statistic	−422.3723	0.0000
	Panel PP-statistic	−385.6036	0.0000	Group ADF-statistic	−35.02034	0.0000
	Panel ADF-statistic	−15.23803	0.0000
Pedroni [20] weighted statistic	Panel v-statistic	20.59964	0.0000
	Panel rho-statistic	−9425.909	0.0000
	Panel PP-statistic	−514.7113	0.0000
	Panel ADF-statistic	−21.76301	0.0000

Pedroni proposes two families of tests, one made in 1999 based on seven tests (four based on the intraindividual dimension and three on the interindividual dimension) and another family of tests made in 2004, suggesting four weighted statistical tests. Both categories of tests are based on the null hypothesis of no cointegration. The cointegration of the variables depends on the value of the probability associated with each statistic (probability <0.01). Tables 1 and 2 summarize the results of Pedroni's cointegration statistics. From the results of Pedroni's cointegration tests, we can see that on all statistical tests, all probability values are less than 1% (they are all at 0.0000). As a result, all these tests show the existence of a cointegration relationship. In the following, we will estimate the long-term relationship of cointegration using the most appropriate methods for this type of approach. Long-term relationships with FMOLS and DOLS: the estimation results are reported in Tables 3–6.

Table 3

FMOLS estimates for ECG.

Dependent variable ECG	FMOLS
	Independent variables
	BP	EEG	RESP
Intraindividual	[−0.000258–6.396131 (0.0000)∗	[0.63681214.40809 (0.0000)∗	[−0.000661–0.207017 (0.8360)
Interindividual	[−0.000360–9.034251 (0.0000)∗	[0.94618919.29055 (0.0000)∗	[−0.158511–1.123774 (0.2611)

Table 4

DOLS estimates for ECG.

Dependent variable ECG	DOLS
	Independent variables
	BP	EEG	RESP
Intraindividual	[−0.000267–6.195763 (0.0000)∗	[0.89976118.38183 (0.0000)∗	[−0.0010–0.30556 (0.7599)
Interindividual	[−0.000393–8.779509 (0.0000)∗	[1.68458422.25036 (0.0000)∗	[−0.0694–0.25507 (0.7987)

Table 5

FMOLS estimates for EEG.

Dependent variable EEG	FMOLS
	Independent variables
	BP	ECG	RESP
Intraindividual	[1.80E − 051.530271 (0.1260)	[0.0071106.599660 (0.0000)∗	[0.0026842.877364 (0.0040)<
Interindividual	[2.44E − 051.964065 (0.0495)	[0.0068114.866931 (0.0000)∗	[-0.0006520.006750 (0.9946)

Table 6

DOLS estimates for EEG.

Dependent variable EEG	DOLS
	Independent variables
	BP	ECG	RESP
Intraindividual	[1.91E − 053.374540 (0.0007)∗<	[0.00853112.73472 (0.0000)∗	[0.0027146.271880 (0.0000)∗
Interindividual	[2.22E − 053.709818 (0.0002)∗<	[0.0071348.048417 (0.0000)∗	[−0.01553–0.30825 (0.7579)

Tables 3–6 establish the long-term elasticity between the different variables of the model from the FMOLS and DOLS estimators. The modeling of the within dimension allows us to take into account the heterogeneity of the coefficients in their temporal and/or individual dimensions. The within estimator eliminates individual-specific effects. The results obtained for the ECG signal indicate that a 1% increase in the BP, EEG, and RESP variables increases the ECG by −0.000258%, 0.636812%, and −0.000661%, respectively, according to the FM-OLS model, as well as −0.000267%, 0.899761%, and −0.0010% according to the DOLS model. These results are for the intraindividual. For the interindividual, a 1% increase in the variables BP, EEG, and RESP increases the EEG by 0.000360%, 0.946189%, and −0.158511%, respectively, according to FM-OLS and 0.000393%, 1.684584%, and −0.0694% according to DOLS.

For the EEG signal, a 1% increase in the BP, ECG, and RESP variables increases the EEG by 1.80E − 05%, 0.007110%, and 0.002684%, respectively, according to the FM-OLS model, as well as 1.91E − 05%, 0.008531%, and 0.002714% according to the DOLS model. These results are deducted for the intraindividual. For the interindividual, a 1% increase in the BP, ECG, and RESP variables increases the EEG by 2.44E − 05%, 0.006811%, and −0.000652%, respectively, according to FM-OLS and 2.22E − 05%, 0.007134%, and −0.01553% according to DOLS.

A Granger causality analysis is performed to determine if there is a potential power of predictability from one indicator to another. The test results for all variables are summarized in Table 7. It should be noted that the optimal delay (Lag) was established using the information criteria of Akaike and Schwarz [27-32].

Table 7

Panel Granger causality.

Lag = 46	ECG	BP	EEG	RESP
ECG	—	109.348∗	16.001∗	1.42665
	—	(0.0000)	(1E − 111)	(0.0374)
EEG	45.640∗	5.6421∗	—	0.98476
	(0.0000)	(2.E − 28)	—	(0.4982)

The results show that for the eighteen patients with sleep apnea, there is a bilateral relationship between ECG-BP, ECG-EEG, and EEG-BP. This information about the direction of causality and the rate of causality is essential for monitoring people at risk. Knowledge and quantitative understanding of these interactions allows for better intervention by health professionals. The purpose of this study is to propose the study of the causal directionality between the signals mentioned above and the cointegration test. Our contribution can be summarized in the following points:

Two-dimensional analysis of the heart and brain during sleep

Verifying if there is a long-term relationship between ECG/EEG, BP, and RESP

Modeling and quantifying the convergence rate of this long-term relationship, if it exists

Defining the causal direction between ECG/EEG and hemodynamic respiratory signals during sleep based on the Granger panel causality.

Trying to understand the impact of hemodynamic respiratory signals on ECG/EEG during sleep for our eighteen patients in the long term

The study will help clinical treatment decisions that rely on the prognostic evaluation of a patient's future health outcomes, as earlier stated by Billheimer et al. [33]. The study would support the effectiveness of medical decisions.

6. Conclusion

The panel tests revealed several complications. First, the problem of heterogeneity of a large number of data, resulting in a multivariate study, makes the parameters very difficult to model. The results show a high level of accuracy and efficiency and clearly highlight the impact of hemodynamic respiratory signals on ECG/EEG during sleep for our eighteen patients in the long term. They also introduced a new approach for identifying epileptic seizures using the time-frequency domain features. This paper's physicians and health personnel take all precautions to facilitate the choice of appropriate intervention and the necessary decisions to save people's lives suffering from apnea syndrome. This work has introduced a new approach for identifying epileptic seizures using time-frequency domain features. The datasets used also produced a high level of accuracy. Consequently, accuracy and precision were used as performance metrics for measuring the acceptance of the new approach introduced in this study. This was simply done by measuring the number of correct decisions made by the classifier and then dividing it by the total number of test examples. In the future, we will work to improve the overall classification accuracy of the classifier by generating multiple base classifiers using ensemble modeling.