Numerous research works have been conducted to detect infant crying [17,18,19] and to identify the reason behind this crying and if this is related to a pathological case. Most of the current research works have focused on classifying pathological from healthy infants, using crying cues [20]. Other works go into more specifics to diagnose certain pathologies such as hypoacoustic [21], asphyxia [22,23,24], hypothyroidism [25], septic [18], RDS [26], and autism spectrum disorder (ASD) [27]. Such research studies and systems mainly involved two main stages, the feature computation and extraction stage, using the CAS and based on different audio domains, including the cepstral domain features, prosodic domain features, image domain features, time domain features, and wavelet domain [14]. The computed features are fed into the next part of the ML model which could be traditional machine learning models or DL models since researchers have recently begun to explore the use of DL algorithms for analyzing infant crying. DL approaches have shown effective results in automatically extracting useful features from audio signals and in classifying sounds into different categories such as healthy and sick infants [19,22,24,28,29,30,31,32].

Most researchers have adopted the cepstral domain features in the feature extraction from audio signals such as Mel frequency cepstral coefficients (MFCC) [33,34,35,36], linear frequency cepstral coefficients (LFCC) [37], short-time cepstral coefficients (STCC) [37], and Bark frequency cepstral coefficients (BFCC) [38], combined with both DL and traditional ML models. MFCCs were the most used in identifying infant pathologies. For instance, in [33], the authors’ system was used to classify the causes of the infants’ crying into eight reasons, including belly pain, discomfort, hungry, sleepy, and tired. The MFCC coefficients have been used to train three ML algorithms, including the K-nearest neighbors rule (KNN), SVM, and naïve Bayes classifier (NBC). The KNN had the highest accuracy of 76%. In [34], they used a dataset of CAS for healthy and pathological infants including 34 pathologies. As a first step, feature extraction was performed using a different set of techniques including the extraction of MFCC and amplitude modulation features. These features were fed into two machine learning algorithms, probabilistic neural networks, and an SVM algorithm with an accuracy of 72.80% and 78.70%, respectively.

Moreover, the MFCC was adopted for feature extraction from audio signals [28] to be used in the training of set machine learning models, including artificial neural network (ANN), CNN, and long short-term memory (LSTM). These ML models were trained to achieve two purposes, identify sick and healthy babies, and then determine the baby’s needs such as hunger/thirst, need for a diaper change, and emotional needs. On the first goal, CNN was able to achieve an accuracy of 95% and an accuracy of 60% was achieved for the second classification purpose. A similar feature extraction was also used along with KNN in [35] and achieved an accuracy of 71.42% in determining the reason for crying, including hunger, belly pain, need for burping, discomfort, and tiredness. In [36], MFCC was used with the CNN model with multiple variants to test and multistage a heterogeneous stacking ensemble model, which consists of four levels of algorithms, Nu-support vector classification, random forest (RF), XGBoost, and AdaBoost. The classification results of the CNN model outperformed the other ML algorithms, reaching an accuracy of 93.7%.

The prosodic domain features were also employed in the analysis and diagnosis of infants’ crying signals. This domain includes much valuable information, such as variations in intensity, fundamental frequency (F0), formants, harmonicity, and duration, which contribute a lot to infant crying signals analysis. This has been followed by a lot of research regarding whether stand-alone or being combined with the cepstral features improves performance. For instance, in [39], they based the proposed model on mean, median, standard deviation, and minimum and maximum of F0 and F123 to distinguish between full-term and preterm infant cries. In contrast, in [22], they used a combined model of weighted prosodic features and MFCC features, thus feeding them into a DL model which was able to achieve a 96.74% accuracy. The obtained results emphasized the importance of using both domains in extracting and modeling a more efficient feature set.

The authors in [40] depended on the wavelet domain audio feature by using the discrete wavelet transform (DWT) method to extract the coefficient characteristics. These coefficients have been used in the classification process using a single-layer neural feed-forward (SLNF) network. This system was able to distinguish between five categories of crying: Eh, Eairh, Neh, Heh, and Owh. Each one is related to a specific condition in a baby, where Heh is related to the feeling of discomfort, and Owh is related to feeling sleepy. Neh indicates thirst or hunger, and Eairh is related to the feeling of burping due to congested air in the chest or stomach. The crying signals were passed through discrete wavelet transform for feature extraction where all signals were then extracted for cry classification using five scaling functions of the wavelet transform, namely Haar, Db2, Coif1, Sym2, and Bior3.1, where the output of each function is used as an input for SLNF. The average accuracy of all discrete wavelet functions on the baby language is over 80%.

Furthermore, the image domain features were used in this field of study, where the main feature is the spectrogram, which is an image or a time–frequency representation of audio [14]. For example, the researchers in [32] classified the neonatal cry signals into pain, hunger, and sleepiness, using the short-time Fourier transform (STFT) technique to generate the spectrogram images, which were used as an input for training a deep convolutional neural network (DCNN), where the extracted features from the DCNN were used as an input for the SVM classifier, which was able to reach an accuracy of 88.89% using the radial basis function (RBF) kernel. Similarly, the spectrogram for the feature extraction and SVM classifier obtained an accuracy of 71.68% [41]. Moreover, the researcher in [29] used the spectrogram with the CNN model for classifying the condition of the baby, whether sleepy or in pain, and obtained an accuracy of 78.5%.

Some researchers have gone more deeply into this topic to diagnose a specific disease. For instance, the authors in [42] suggested a machine learning model to diagnose hypoxic ischemic encephalopathy disease in newborns based on CAS analysis. Multiple feature extraction techniques were used, including the MFCC and Gammatone frequency cepstral coefficients (GFCCs). These features were utilized by a basic deep network, achieving an accuracy of 96%. The authors in [37] introduced a classification model between healthy and unhealthy newborn cries. A set of feature extraction techniques were used, including MFCC, LFCC, STCC, and Teager energy cepstral coefficients (TECC). The classification process is based on the Gaussian mixture model (GMM) and SVM algorithms. Both models have been trained using the different features extracted separately and the results justified the superiority of the TECC representations with the GMM classifier, which achieved an accuracy of 99.47%. Furthermore, in [31], the researchers developed a DL approach that can classify healthy and pathological babies based on the infant’s CAS, where the signals were processed using cepstrum analysis to extract the harmonics in the cry records, and the outputted spectrum was fed into three DL models including deep feed-forward neural networks (DFFNN), LSTM, and CNN. The latter DL model outperformed the other algorithms with an accuracy of 95.31%. Similarly, the researchers in [43] adopted the cepstrum to build a model to distinguish between healthy and pathological infants based on the crying signal by evaluating DFFNN, naïve Bayes, SVM, and a probabilistic neural network. The DFFNN achieved a 100% accuracy.

Few researchers have followed a combined features domain similar to the work in [8] where they combined both GFCC and HR features by using simple concatenation to distinguish between RDS and sepsis. Using SVM and MLP, the SVM achieved 95.29% compared to 92.94% for the GFCC alone and 71.03% for the HR. While in [44], they combined images that contain the prosodic feature lines including F0, intensity, and formant spectrogram CNN and waveform CNN, producing a 5% better accuracy. This study [45] explored the use of DL models with hybrid features to classify asphyxia cries in infants. The models used a combination of MFCC, chromagram, Mel-scaled spectrogram, spectral contrast, and Tonnetz features. The results showed that the DNN models performed better with the hybrid features, achieving a 100% accuracy for normal and asphyxia cries, and a 99.96% accuracy for nonasphyxia and asphyxia cries. The CNN model performed better with the MFCC alone. The study demonstrated the effectiveness of using DL models with hybrid features for classifying asphyxia cries in infants.