Machine Learning-Based Detection of Fetal Respiratory Patterns: A CNN-LSTM Approach for Enhanced Perinatal Monitoring

Monitoring fetal respiratory patterns is essential for evaluating fetal health and detecting potential distress or abnormalities during pregnancy [1]. These patterns, reflected in the rhythmic contractions of the fetal diaphragm, are key indicators of respiratory system [2] maturity and can signify developmental issues if abnormalities are detected [3]. Traditional methods in fetal monitoring, such as cardiotocography (CTG) and ultrasound, provide limited precision in distinguishing subtle respiratory changes, especially in complex signals where baseline drift and noise from maternal movements can obscure vital information [4]. Traditional methods for fetal respiratory monitoring, such as basic signal processing techniques or CTG-based monitoring often fall short in precision, especially when detecting subtle respiratory changes. These methods struggle with high baseline drift, artifacts, and limited feature extraction capabilities [5]. Machine learning, particularly CNN-LSTM architectures, addresses these challenges by capturing complex patterns and dependencies within the data, enabling more accurate and sensitive detection [6].

This paper presents a CNN-LSTM-based framework specifically tailored to overcome these challenges by leveraging both spatial and temporal analysis capabilities [6]. CNNs are employed to capture spatial features and subtle nuances in fetal breathing patterns, while LSTM networks are integrated to detect temporal dependencies across respiratory episodes. This dual-architecture approach allows for more accurate differentiation between normal and abnormal respiratory patterns, addressing signal inconsistencies that often hinder conventional approaches [7].

001.jpg

Figure 1. Fetal breathing movement [2]

Figure 1 shows the fetal movement [2]. Moreover, the proposed methodology includes a comprehensive filtering process, using wavelet transforms to remove baseline drift and power-line interference, which enhances the clarity and reliability of the data [7]. These improvements in preprocessing and feature extraction result in a robust machine learning model that significantly enhances the detection and classification of fetal breathing behaviour, aiming to improve prenatal monitoring practices and outcomes [8].

The proposed methodology delineates a structured approach towards advancing prenatal monitoring through machine learning. Database curation serves as the initial cornerstone, and the repository of fetal acoustic signals is compiled, encompassing diverse respiratory patterns and fetal conditions. The working principle of the proposed methodology for fetal breathing movement analysis involves four key steps:

(1) Database Curation

(2) Preprocessing

(3) Feature Extraction

(4) Classification

•A comprehensive fetal acoustic signal database is curated, containing diverse respiratory patterns across different gestational ages and health conditions. This database serves as the foundation for developing machine learning models.

•Raw acoustic signals are processed to improve quality by reducing noise and normalizing the images. This ensures consistent, clean input data for analysis.

•Deep learning techniques, such as image processing and pattern recognition algorithms, are used to extract important features from the fetal breathing signals, capturing subtle respiratory patterns for further evaluation.

•Using machine learning algorithms, including CNNs and LSTM networks, the extracted features are classified into clinically relevant categories such as normal or abnormal breathing patterns. This aids in identifying fetal distress and improving prenatal monitoring.

This methodology outlines a systematic approach for automating fetal breathing movement detection, ultimately aiming to enhance perinatal care outcomes.

3.1 Block Diagram of the proposed system

The proposed methodology involves utilizing a comprehensive database of fetal acoustic signals, preprocessing them for quality enhancement, analysing fetal breathing movements, and classifying patterns using machine learning algorithms for improved perinatal care. The block diagram of the proposed system is shown in Figure 2.

002.jpg

Figure 2. Block diagram of the proposed system

3.2 Dataset curation

The dataset used in this paper was obtained from a publicly available fetal monitoring database, comprising six-hour continuous recordings of fetal breathing movements collected from six labouring women at gestational ages between 38 and 41 weeks. These recordings were segmented into 250 episodes based on variations in tracheal pressure, with each segment classified as either accentuated (tracheal pressure > 3.5 mm Hg) or non-accentuated (tracheal pressure ≤ 3.5 mm Hg) to capture a wide range of fetal respiratory behaviors. To ensure signal quality and consistency, the dataset underwent additional preprocessing steps, including noise reduction and the removal of baseline drift and 50Hz power-line interference. The dataset was selected due to its structured annotations and high-resolution signal quality, supporting reliable model development and benchmarking. It is important to note that this work relies solely on publicly available datasets for model development and evaluation. No proprietary or clinical datasets were used. A sample signal from the dataset is shown in Figure 3.

003.jpg

Figure 3. A sample of signal in the dataset

3.2.1 Dataset curation and protocol

The fetal breathing movement (FBM) data used in this work was sourced from a publicly available fetal monitoring database, originally collected using the Philips Avalon FM50 cardiotocography (CTG) monitor. The recordings include high-sensitivity acoustic signals captured at gestational ages between 38 and 41 weeks. Signals were sampled at 1,000 Hz with 16-bit resolution and were band-limited to 20–30 Hz to isolate fetal breathing components. The data was pre-processed and annotated with tracheal pressure-based classifications (accentuated >3.5 mm Hg; non-accentuated ≤3.5 mm Hg). The dataset is structured in European Data Format (EDF) to preserve signal integrity and facilitate standardized biomedical analysis [23-25].

3.2.2 Statistical analysis of the dataset

The statistical analysis of the data set allowed the estimation of its characteristics and limitations. Table 3 has been given in summarization of normal and abnormal episodes, average durations, and tracheal pressure variances obtained across subjects. Although the size of the dataset is limited, so limiting its generalizability, its richness with segmented data and high-resolution features help overcome the above considerations. Future efforts will be to expand the dataset in collaboration with healthcare institutions that can include more subjects with varied gestational conditions.

Table 3. Detailed breakdown of dataset

3.3 Preprocessing

Pre-processing fetal breathing signals using Wavelet Transform (Eq. (1)) is a sophisticated technique aimed at enhancing signal quality by removing noise, artifacts, and unwanted components such as maternal movements like burb, heart rate and fetal movements kick.

$w(a, b)=\iint_{\infty} \frac{s(t)}{\sqrt{a}} \psi \frac{t-b}{a} d t$            (1)

where,

w(a,b) is the wavelet coefficient.

s(t) is the input signal.

$\mathrm{\psi(\mathrm{t})}$ is the mother wavelet.

a is the scale parameter (dilation).

b is the translation parameter (shift).

Wavelet transforms are highly effective for analyzing non-stationary signals like fetal breathing behavior. These signals contain varying frequency components that standard Fourier transforms cannot adequately address due to their assumption of stationarity. The wavelet transform however, allows for signal decomposition across both time and frequency domains, providing a multi-resolution analysis.

This is crucial in fetal monitoring, where transient events such as fetal respiratory movements and irregular breathing episodes need precise localization in both time and frequency. By applying WT during preprocessing, respiratory signals can be isolated from noise and artifacts, which improves the accuracy of subsequent CNN-based feature extraction.

3.4 Feature extraction and classification

In this paper, feature extraction from the preprocessed fetal breathing signal is performed using Enhanced Convolutional Neural Networks (CNNs), a robust deep learning architecture well-regarded for its effectiveness in analyzing sequential data, including time-series signals. Long Short-Term Memory (LSTM) architecture is crucial for classifying fetal breathing signal data, as it excels in capturing temporal dependencies within sequential data. The enhanced CNN-LSTM architecture is shown in Figure 4.

004.jpg

Figure 4. Enhanced CNN-LSTM architecture

3.4.1 Dataset split and over fitting prevention techniques

To ensure model robustness and avoid overfitting, given the complexity of fetal respiratory signals and the relatively small dataset, several strategies were implemented: data split, cross-validation, regularization and Dropout, Early Stopping.

(1) Data split

The dataset was split into 70% for training, 15% for validation, and 15% for testing, a common approach that allows sufficient training while reserving part of the data to monitor generalization. However, given the limited dataset size, this split alone may not prevent over fitting effectively.

(2) Cross-validation

Applied k-fold cross-validation during the model training phase, where the dataset is divided into k subsets. In each iteration, one subset serves as the validation set while the others are used for training. This rotation continues until each subset has been used for validation, ensuring that each data point is used in training and validation phases, thus enhancing generalization. k=5 is used to balance computational efficiency with robust validation, which helped in detecting any tendency toward overfitting.

This approach ensures the model does not continue training past the optimal point, thereby enhancing generalization. Context vector is passed through fully connected layers for dimensionality reduction, leading to the final output classification as:

$H_i=\sigma\left(\sum X \times K_{i j}+b_i\right)$          (2)

where, K_ijrepresents the convolution kernel, $b i$ the bias, and $\sigma$ an activation function like ReLU. A residual connection adds the original input back to the output of the convolutional block, represented as:

$Y=F(X, W)+X$               (3)

The residual connection facilitates direct information flow and preventing gradient vanishing. The LSTM layer captures temporal dependencies in respiratory patterns, computing various gates and states, with the hidden state given by:

$h_i=O_t \times \tanh \left(C_t\right)$                    (4)

The attention mechanism focuses on crucial segments of the sequence, calculating attention scores and weights to create a context vector that emphasizes important features. The context vector is passed through fully connected layers for dimensionality reduction, leading to the final output classification as:

$Y=\sigma\left(W_{ {OUT}} \cdot \mathcal{c}+b_{ {out}}\right)$               (5)

The final output classification determines whether the respiratory patterns are normal or abnormal.

Attention Mechanism: This layer emphasizes crucial temporal segments, enhancing the model’s ability to detect subtle respiratory pattern abnormalities that is critical in fetal monitoring.

3.4.2 Ablation study for architectural justification

To rigorously evaluate the individual contributions of each architectural component within the proposed CNN-LSTM framework, a systematic ablation study was performed. This analysis involved training and testing four progressively structured model variants under identical experimental conditions to ensure comparability. The first variant consisted of a CNN-only architecture, designed to extract spatial features from the input signal. The second employed a standalone LSTM model, focused solely on capturing temporal dependencies. The third was a CNN-LSTM combination, forming the baseline without any residual connections or attention mechanisms. The final configuration was the proposed full model—integrating CNN and LSTM with residual connections and an attention mechanism. All models were trained using the same dataset split, with 70% allocated for training, 15% for validation, and 15% for testing. Table 4 shows the comparison of performance for each configuration. The models are optimized using the Adam optimizer with a learning rate of 0.001. The performance of each model was evaluated using standard metrics including accuracy, F1 score, and area under the ROC curve (AUC), providing a quantitative basis for assessing the effectiveness of each architectural enhancement.

The integration of residual connections yielded an approximate 3.8% improvement in accuracy, primarily by mitigating vanishing gradient issues and preserving low-level signal characteristics across layers. The inclusion of an attention mechanism further enhances the performance by focus the model on temporally significant features, increasing both F1 score and AUC.

Table 4. Comparison of performance metrics for each configuration

This validates the novelty of the proposed architectural design and underscores the additive benefit of each component in the full pipeline.

3.4.3 Class imbalance mitigation and performance justification

Despite the inherent imbalance in the dataset, with a higher number of normal respiratory episodes (approximately 95) compared to abnormal cases (approximately35) a few specific countermeasures were implemented to mitigate potential bias in model training. First, during preprocessing, stratified sampling was applied to ensure that all training, validation, and test splits maintained a representative distribution of both classes. Additionally, a weighted categorical cross-entropy loss function was employed, assigning higher weights to the minority class (abnormal episodes) to penalize misclassifications more heavily. This approach helps the model learn balanced decision boundaries despite skewed class proportions. It also been experimented with Synthetic Minority Over-sampling Technique (SMOTE) on feature representations during training to generate synthetic abnormal samples, which improved minority class recall. The effectiveness of these strategies is evident from the high sensitivity (94.8%) and specificity (96.1%), indicating that the model did not favor the majority class and could reliably identify abnormal cases. These mitigation strategies ensure that the model remains robust, fair, and clinically reliable in real-world scenarios where data imbalance is common. Table 5 shows the performance outcomes measured in 5-fold cross validation. The 'Support (N)' row in Table 5 indicates the number of actual samples for each class used during the evaluation.

Table 5. Performance outcomes measured in 5-fold cross validation

3.5 Novelty of the proposed method

This enhanced architecture combines both spatial and temporal analysis with an emphasis on key features, providing a robust, novel approach for fetal respiratory pattern detection. Residual connections, while standard in image recognition, represent a novel approach in fetal respiratory monitoring due to their limited use in medical signal processing, particularly for fetal health. Medical signals, including fetal respiratory patterns, are inherently noisy and complex, often leading to challenges in capturing subtle yet clinically significant variations. By embedding residual connections within CNN layers, this architecture maintains essential signal fidelity across layers, addressing vanishing gradient issues and enhancing feature extraction depth. This method allows the model to differentiate fine-grained respiratory changes indicative of potential distress or abnormalities, which is critical for accurate, real-time detection. The novel integration of residual learning in this domain also improves robustness, enabling better adaptation to varied signal patterns across patients, offering a pioneering contribution to automated fetal monitoring applications where conventional approaches struggle to maintain signal integrity and interpretability.

3.5.1 Hyper parameter optimization and justification

To ensure transparency, reproducibility, and optimal model performance, a comprehensive hyper parameter tuning process was conducted using grid search, with all evaluations performed under 5-fold subject-wise cross-validation. Each model variant was assessed using macro F1-score and AUC as selection criteria to ensure balanced performance across all classes, especially the minority (pathologic) class. For the LSTM module, hidden unit sizes of 64, 128, and 256 were tested, with 128 units providing the best balance between learning temporal dependencies and avoiding overfitting. The attention mechanism was evaluated with 1, 2, and 4 attention heads; 2 heads consistently yielded the highest performance, providing sufficient focus on relevant temporal patterns without model instability. CNN filter sizes were optimized in an increasing-decreasing pattern (32–64–64–32), which was found to enhance spatial feature extraction while reducing redundancy. Dropout values of 0.3, 0.5, and 0.7 were explored, with 0.5 demonstrating the most consistent regularization without performance degradation. Learning rates of 1e-3, 5e-4, and 1e-4 were tested, where 5e-4 provided the fastest and most stable convergence, as observed through early stopping on validation loss. Batch sizes of 16, 32, and 64 were evaluated and 32 were selected as it offered the best trade-off between convergence stability and computational efficiency on the available GPU hardware. Each combination was trained for 50 epochs using the Adam optimizer and categorical cross-entropy loss, with early stopping (patience = 8) to prevent overfitting.

This optimization framework ensured that the final model architecture is not only high-performing but also fully reproducible.

Figure 5 shows the power spectral density of a fetal breathing signal after wavelet-based denoising. Figure 6 shows frequency components of segment 25 showing dominant respiratory bands. Figure 7 shows filtered time-domain signal of segment 25 post wavelet preprocessing. Figure 8 shows the detected peak group 13 in segment 25 indicating breathing activity.

005.jpg

Figure 5. Power spectral density of a fetal breathing signal after wavelet-based denoising

006.jpg

Figure 6. Frequency components of segment 25 showing dominant respiratory bands

007.jpg

Figure 7. Filtered time-domain signal of segment 25 post wavelet preprocessing

008.jpg

Figure 8. Detected peak group 13 in segment 25 indicating breathing activity

This work presents a novel CNN-LSTM architecture incorporating residual connections and attention mechanisms—an innovative and rarely explored approach in fetal respiratory analysis. This combination enhances feature retention and focuses learning on critical signal segments, significantly improving pattern recognition, especially in noisy or low-quality biomedical data. Despite a limited dataset, the model demonstrates high accuracy, sensitivity, and specificity, offering a non-invasive and automated solution for fetal respiratory monitoring. Looking ahead, future efforts will focus on expanding the dataset through collaborations with healthcare institutions to improve generalizability across diverse populations and gestational stages. Optimization for real-time deployment on portable devices will be pursued using model pruning and quantization, enabling low-latency monitoring in both clinical and remote settings. Signal quality assessment (SQA) and adaptive filtering will be developed to handle signal variability, ensuring robust performance in real-world use. Integration of multimodal data such as maternal heart rate and uterine contractions will support a more holistic assessment of fetal well-being. Personalization through transfer learning will tailor predictions to individual patients, benefiting high-risk pregnancies that require frequent monitoring. Clinical trials and regulatory evaluation will be essential for validating safety, usability, and clinical effectiveness, ensuring readiness for adoption in modern prenatal care practices. This research lays the foundation for scalable, intelligent fetal monitoring systems, advancing toward accessible, real-time prenatal care that can enhance outcomes for mothers and infants across diverse healthcare environments.

The authors would like to thank Dr. Ilayaraja Venkatachalam (Pixel Scans, Trichirappalli) and Dr. Rajkumar (Government Hospital, Ramnathapuram) for providing clinical insights during the model evaluation phase.