Analyzing infant cry to detect birth asphyxia using a hybrid CNN and feature extraction approach

Asphyxia, a critical respiratory condition, poses significant risks to newborns and can lead to catastrophic outcomes. Early detection of asphyxia is crucial for reducing infant mortality rates. Traditional medical diagnosis methods can be time-consuming, whereas early detection through artificial intelligence (AI) can expedite the process and improve survival rates. Despite the importance of early asphyxia detection, existing methods are often delayed and not always effective. This research addresses the need for a faster, more accurate approach to detecting infant asphyxia using machine learning (ML) and deep learning (DL) techniques. This study aims to develop a robust AI-driven system to detect asphyxia in newborns using ML and DL models, focusing on improving accuracy and efficiency over traditional diagnostic methods. This study explores feature extraction using Mel-Frequency Cepstral Coefficients (MFCCs), where the features are categorized into time and frequency domains. Data preprocessing techniques, such as noise removal, handling missing values, outliers, and label encoding, are applied to ensure clean data. To address class imbalance, the Random Oversampling (ROS) technique is employed. Hyperparameter optimization is performed using GridSearchCV for various machine-learning models. Deep learning models, including custom artificial neural networks (ANN1) and convolutional neural networks (CNN1, CNN2), are introduced with hidden layers for improved performance. The performance of different ML and DL models is evaluated, with Logistic Regression (LR) achieving an accuracy of 99.16% and a 0.008% error rate. In comparison, ANN1 outperforms other DL models with an accuracy of 98.20% and a 0.018% error rate. The results demonstrate that both ML and DL techniques can significantly enhance early asphyxia detection in newborns. The Logistic Regression model offers the highest accuracy in machine learning, while ANN1 performs optimally in deep learning, suggesting their potential for deployment in clinical settings to improve neonatal care.