Design of a robot system for improved stress classification using time–frequency domain feature extraction based on electrocardiogram

2.1 Existing techniques used to determine the stress level

In literature, several techniques are commonly used to determine an individual’s stress level, drawing from physiological, psychological, and behavioral indicators. These techniques provide insights into the person’s overall stress response and aid in assessing their well-being as given below:

3.1 ECG data acquisition

The dataset for the proposed methodology was collected from the PhysioNet; the dataset is free to access to motivate the practitioners for study. The investigational study includes the complex physiological signal analyzed in Boston’s Beth Israel Hospital and further evaluated by the Massachusetts Institute of Technology (MIT). The collective dataset is called the “MIT-BIH Arrhythmia Database” which contains 1,000 subjects nearly 48 half-hour ECG recordings [22,23]. The acquired ECG data is then passed to the subsequent modules for processing.

3.2 Pre-processing

The ECG signal collected from the dataset is pre-processed to shrink the noise level. The raw ECG signals are corrupted and diminished by external noise. Moreover, inappropriate placement of the subjects and skin preparation disrupt the obtained signal. In the literature, various pre-processing techniques have been used to process the signal such as filters, empirical models, and wavelet transform.

In this research, VMD is used to reconstruct the original signal fed at the input. The decomposition of the signal is carried out to convert the original signal into the variational mode functions. It is a robust technique to handle and control the noise level [12]. It decomposes the real values function of an input signal (f) into a discrete number of signals that are called variational mode functions ( μ n ). Each variational mode is oriented and signified toward central frequency ( ω n ), computed during the process of decomposition. Moreover, there is a sparsity property for each particular model that is used to reconstruct the signal. The detailed behavior of the VMD is illustrated in Algorithm 1.

Initialize { μ n 1 ˆ } , { ω n 1 ˆ } , λ 1 , ˆ k → 0

Repeat

k ← k + 1

For n = 1:N do

Update μ n ˆ for all s ω ≥ 0 ; μ n ˆ k + 1 ( ω ) ← ( f ˆ ( ω ) − ∑ i < n μ i ˆ k + 1 ( ω ) − ∑ i > n μ i ˆ k ( ω ) + λ ˆ k ( ω ) 2 1 + 2 ∝ ( ω − ω n k ) 2 )

Update ω na ω n k + 1 ← ( ∫ 0 ∞ ω | μ n ˆ k + 1 ( ω ) | 2 d ω / ∫ 0 ∞ | μ n ˆ k + 1 ( ω ) | 2 d ω

End for

Dual ascent for all ω ≥ 0 λ ˆ ( k + 1 ) ( ω ) ← ( ω ) + τ ( f ˆ ( w ) − ∑ n μ n ˆ k + 1 ( ω ) )

Until convergence ∑ n μ n ˆ k + 1 − μ n ˆ 2 k 2 μ ˆ n k 2 2 < ∈

Each data point is indicated as k i demands k-dimensional space to plot the various data points. The objective of the proposed algorithm is to compute an appropriate hyperplane by discriminating the correct data points.

3.3 Feature extraction

In this paper, TFD features were extracted from the processed ECG signals. The ECG waveform is shown in Figure 4. These features are further used as input to the developed model. The detection accuracy improved by properly representing the characteristics of the extracted signal.

Figure 4

ECG waveform [23].

3.4 Feature classification

The extracted and selected features are fed into a stress classification algorithm. There are various classification algorithms such as random forest, KNN, Random Forest, Naïve Bayes, and SVM used in the literature for various applications such as clustering, stress level detection, and others. [8]. These algorithms learn patterns from the features to differentiate between different stress levels. The past study shows that SVM shows better performance in comparison to other machine learning approaches. We can say that it is a unique technique to break the information in data mining and then classify the fragmented part. For instance, it classifies the emotions corresponding to low, medium, and high levels of stress with minor chances of error. SVM generally shows results upon the hyperplane, which is useful in sorting out the new illustrations. However, the 2D plane shows a line that cuts one part into two sections, and each class is located on either side.

In this study, the classification process is undergoes through multi-kernel SVM. The use of SVM is advantageous at this stage as it avoids the multidimensionality of advanced machine learning techniques. Moreover, the programming strategies following the SVM could aid the best hyperplane to categorize the input data. The data trained during the training process is stored in the form of a training dataset trained data computed as follows:-

where, trained data signifies the signal that is optimized. Training data were obtained for three categories low, medium, and high-level ECG signals. Kernel (f) signifies the multi-kernel employed to predict the current output of the ongoing work. The trained data was further used to estimate the unknown ECG signal level. The given Algorithm describes the SVM classification process:

If Dat a tested ( d ) = = ′ low ′ //check if test signal data corresponds to low-level signal

Y N = Signa l normal //category of low-level signal test data

Else

Low S = Signa l stress //category of stress signal test data that is normal;

1. End if

2. If Dat a tested ( d ) = = ′ Medium ′ //check if test signal data corresponds to medium level signal

   Y N = Signa l normal //category of medium-level signal test data

   Else

   Medium S = S igna l stress //category of stress signal test data that is abnormal;

3. End if

4. If Dat a tested ( d ) = = ′ High ′ //check if test signal data corresponds to High-level signal

5. Y N = Signa l abnormal //category of high-level signal test data

6. Else

7. Hig h S = Signa l stress //category of stress signal test data that is not normal;

8. End if

9. END IF

10. END FOR

Return the different levels of signal Y N and Hig h S for abnormal stress

In the given algorithm, the ECG signal is fed as input to compute the SVM parameters. The data is tested to acquire the low, medium, and high levels of stress. This can be done by classifying the properties of the signal. If the tested data falls in the low-level range then data is termed as low-level stress which is normal. Consequently, if the data falls in the medium level range then a signal is of medium level. Further, if the signal crosses the threshold level then the stress level, then it is very high which needs to be proper attention. Then, attached robot was used to provide breathing guide orally using robot so that stress can be reduced.

3.5 Human–robot interaction and feedback mechanism

A vital role in offering emotional support and relieving stress is played by human–robot contact, which can take many different forms in stress management. Robots, for instance, can be brought into the house to help families manage their stress. They could organize family get-togethers, promote candid conversation, and provide stress-relieving activities that the whole family can take part in. Pet therapy robots could mimic the sensation of engaging with a real pet for people who find comfort in the company of animals [25,26]. Through their interactions and company, these robots could be able to relieve tension and offer affection and companionship. To effectively relieve stress and offer emotional support, it is important to integrate cutting-edge AI and robotics technology with empathy and understanding.

In this study, the robots are linked to wearable devices that continuously record the user’s ECG readings. The user’s heart rate and heart rhythm, which are indicators of stress levels, are disclosed by these signals. From the ECG data, the robot extracts pertinent aspects such as rhythm analysis, peak detection, and HRV. These characteristics shed light on the user’s physiological reactions to stress. The robot analyses the retrieved features using machine learning algorithms, including deep learning and traditional statistical techniques. The robot evaluates the user’s current stress level by analyzing the ECG data. It can identify if the person is relaxed, under a lot of stress, or somewhere in between.

Personalized stress management advice is given by the robot based on past data and the user’s current stress level classification. These suggestions might be as breathing exercises. If the user is under a lot of stress, the robot might suggest guided meditation or deep breathing techniques to help them relax. It adjusts its behavior to offer support, relaxation techniques, or notifications. The robot also sends notifications to the user in the form of text message or healthcare professionals if abnormal patterns are detected. The interface interacts with the user to provide feedback and personalized stress management techniques. The robot’s responses are adaptive, changing based on the user’s stress level. For instance, if the system detects high stress, the robot might initiate breathing exercises, soothing music, or guide the user through mindfulness techniques.

4.1 Performance parameters

The performance metrics used in this study are accuracy, F-measure, recall, and precision, which indicate the bias level and computation of variance of the proposed approach. The use of four performance metrics evaluates the performance of the classifier. True positive indicates the correctly classified signals and false-positive signifies the erroneous recordings. The proposed methodology ensures that positive value is predicted to reflect greater quality. The performance metrics are explained in the given equations (Tables 1 and 2, Graphs 1–4).

• Accuracy: It is the ratio of the sum of true positive (TP) and true negative (TN) to the sum of TN, false positive (FP), TP, and false-negative (FN) as given in equation (1). TP and TN indicate the prediction of correct data points using the classifier and FP and FN indicate the incorrect classified data points.

• Sensitivity: It is known as the TP rate. Generally, it is the ratio of TP to the sum of TP and FN. This also signifies the ability of the system to correctly recognize the disease as given in equation (2).

  (2) Senstivity = TP TP + FN .

• F-measure: It is defined as the ratio of TP to the sum of TP and 0.5 of a sum of FP and FN.

  (3) F - measure = 2 · Precision × Recall Precision + Recall = TP TP + ( 0.5 ) ( FP + FN ) .

• Precision: It is defined as the ratio of TP to the sum of TP and FP.

Graph 1

Comparison of accuracy using VMD + TFD + SVM and using VMD + SVM only.

Graph 2

Comparison of precision using VMD + TFD + SVM and using VMD + SVM only.

Graph 3

Comparison of recall using VMD + TFD + SVM and using VMD + SVM only.

Graph 4

Comparison of F-measure using VMD + TFD + SVM and using VMD + SVM only.

4.2 Comparative analysis

In this section, the proposed approach is compared with the existing techniques to determine the effectiveness and robustness. The existing literature uses the same dataset MIT-BIH and considers 1,000 ECG samples to model the results (Table 3).

Recall values are compared with existing techniques as given in Table 3. The techniques employed in the existing work [13] use VMD for denoising and features were extracted using the time domain and further SVM was used to classify the features. The recall obtained using the existing work was 96.11%. However, the study done in 2021 [21] by the researchers provided 96.83%. The KNNs and Decision Tree [27] achieved recall parameter value of 85.4%. But, the average recall value using the proposed technique is 97.46% which is better than the existing work (Table 4, Graph 5).

Graph 5

Recall comparison of the proposed approach with the existing techniques.

The accuracy of the proposed approach is compared with the existing techniques. The proposed one has better performance than the existing approaches. The study was done using the GA, PSO, and SVM classifier [21] shows 98.93% accuracy while using the PCA [5], which is only 97.78%. Using KNN [28] classifier and Random Forest (RF) model, the accuracy achieved 93.7 and 98.2% respectively. Further, the use of VMD and time domain feature extraction method with SVM [13] shows only 95.74%. Thus, the proposed approach outperforms the other techniques (Graph 6).

Graph 6

Accuracy comparison of the proposed approach with the existing techniques.

Table 5 and Graph 7 depict that the study done using the GA, PSO, and SVM classifier shows 96.83% precision, and using the TFD feature extraction method with SVM shows only 61.7%. The KNNs and Decision Tree [27] achieved a precision parameter value of 88.9%. The proposed one is improved by 0.95% from [21], 36.07% from the existing approaches [10] and 8.88% from 88.9% [27].

Graph 7

Precision comparison of the proposed approach with the existing techniques.