Research indicates that fetal cardiac arrhythmias affect 1 % of all pregnancies. Some estimates put the morbidity-causing potential of cardiac arrhythmias at 10 % [1]. Close monitoring of the foetal heart allows obstetricians to make optimal decisions both before and after birth [2]. A fetus's heart activity can be most accurately monitored via magnetocardiography (MCG). But low- and middle-income countries can't afford MCG [3]. Two modern techniques, cardiotocography (CTG) and Doppler ultrasonography, each have their own set of limitations. CTG only offers information on ventricular blood flow and is a valuable technique for fetal heart rate (FHR) determination during the first 28 weeks of pregnancy [4]. Hospitals' widespread use of CTG monitoring has led to a rise in the number of cesarean section births [5]. The results of electrocardiography (ECG) are superior to those of CTG diagnosis even when the pregnancy is 20 weeks along. A foetal ECG can be obtained by invasively placing an electrode on the foetal scalp [6]. However, this treatment is only feasible after cervical dilation, which carries a danger for the mother and the unborn child. Electrodes can also be positioned in the mother's abdomen to record the FECG signal [7].

Fetuses and adults have slightly different heart structures [8]. After birth, the right ventricle is helpful for pumping blood to the lungs to provide oxygen, while the left ventricle is responsible for blood circulation all over the body. On the other hand, the placenta provides oxygen to the developing baby, so there's no longer any need to pump blood into the lungs of the unborn. On the other hand, the pulmonary system relies on the coordinated efforts of the two ventricles [9]. The structural electrocardiogram patterns of both the developing baby and the adult are the same; however, the relative amplitudes of the fetal complexes change during the course of the pregnancy and the postpartum period [10]. For instance, although fetuses and infants experience very weak T-waves, this is where the most significant change takes place.

Although adult ECG and FECG are quite similar, the processing of adult ECG has advanced considerably in modern medicine [11,12]. In contrast, FECG processing remains a major concern. These days, fetal monitoring is done solely by heart rate [13]. Both tachyarrhythmias (heartbeats faster than 160 bpm) and bradyarrhythmias (heartbeats slower than 120 bpm) are common terms used to describe the irregular fetal heartbeat. The features of the FECG waveform, which are the foundation for cardiac evaluation in both adults and infants, are overlooked when the aberrant fetal heartbeat rhythm is the only focus. The main reason for the lack of usage of FECG in clinical settings is the unavailability of the equipment required for accurate assessment [14]. Therefore, a broad association between ECG features and FECG outcomes has not been identified in various studies. The absence of gold-standard datasets for FECG abnormality detection is another reason why further study is necessary for the diagnosis of FECG arrhythmias. It can be explained in part by the lack of thorough clinical data on fetal cardiac activity and in part by the FECG's lower signal-to-noise ratio when compared to the mother's ECG [15].

The proposed approach extracts the fetal ECG from the input ECG signal using a filter. The clean FECG signals are transformed into time-frequency images using a transform so that the next step can be carried out [35,36]. Using the Non-Invasive Fetal ECG Arrhythmia dataset as an input, a novel hybrid deep learning technique is used to identify the presence of arrhythmia in fetal ECG. It is still difficult for physicians to identify fetal abnormal heartbeats during pregnancy. The current techniques for monitoring the fetal heart during pregnancy increase computing time and yield erroneous detection results. The creation of the proposed framework is encouraged by the fact that the current approach lacks enough precision and increases processing time.

The goal of this research is to provide a framework for the efficient detection of fetal arrhythmia, which will help obstetricians determine whether the fetuses' heart rhythms are aberrant. The majority of the approaches either did not take into account discriminability at many scales or used non-domain specific designs. Long-range dependencies have also not been taken into account. In order to do this, the suggested framework presents AConvVTCapNet, which guarantees minimal parametric design and robust modeling.

One of the primary issues with the fetal ECG is that human monitoring is labor-intensive and prone to mistakes. In order to reduce the error rate, effective steps should be taken to identify the infestation. Fetal ECG analysis techniques are widely used in the medical field to give people the best possible treatment. The previous techniques for detecting arrhythmia in fetal ECG require a great deal of prior knowledge. Recently, methods based on deep learning are thought to be developing and have been suggested for detecting arrhythmia in fetal ECG. Deep learning architectures have the ability to learn on their own and concentrate on intricate input properties, in contrast to machine learning models. This work is motivated by the concept of a deep learning-based method for prenatal ECG arrhythmia diagnosis, taking all of these variables into account. The major objectives of the proposed work are,▪

To extract the fetal ECG and to improve classification accuracy, a novel Non-Causal Adaptive Filter with multiple error estimation is implemented in the first stage.

▪

In order to efficiently learn the information behind the inputs, the clean FECG signals are transformed into time-frequency signals through a fractional Stockwell transform.

▪

To detect the available arrhythmia disease from the inputs, a novel Atrous Convolutions aided Vision Transformer based Capsule Network model is proposed.

▪

To promote the ability of the proposed network model and mitigate the overfitting issues, the hyperparameters are tuned in an optimal manner using an opposition based Fire Hawks optimization approach.

The remaining work of this paper is explained as follows: Section II presents the related work based on sentiment analysis. Section III presents the proposed techniques used to detect sentiments. Section IV contains the evaluation metrics and dataset used in this paper, and Section V presents the conclusion of this study.