In comparison to other methods used for long-term monitoring of fetal health, fetal phonocardiography has the potential to be more convenient and affordable due to its non-invasive nature and the possibility of implementation on omnipresent devices such as smartphones. Fetal phonocardiography signals can oftentimes be misinterpreted due to various sources of sound in the womb. Therefore, the question remained whether a machine learning model trained for fetal heartbeat detection containing a conventional set of audio features could be improved by introducing features taken from signal representations processed by two methods: empirical mode decomposition (EMD) and pitch shifting. Furthermore, features based on psychoacoustics were proposed as an additional input to the model. In other words, the main goal of this research was to employ EMD and pitch shifting as preprocessing steps, as well as psychoacoustics descriptors such as perceptual linear prediction coefficients, in order to enable the utilization of additional characteristics from the phonocardiography signal. Features extracted in this fashion were assessed through the analysis of their relevance, usefulness and significance in relation to the performance metrics, such as accuracy or precision. Two raw datasets of audio data were employed as input, one custom recorded and collected through fetal heartbeat acquisition obtained from 8 pregnant women, while the other was taken as an available dataset of simulated fetal heartbeat sounds with different noise levels. Two scenarios with different inputs were introduced in this research. In Scenario A, the custom dataset was utilized to train a machine learning model from features originating from raw, filtered and EMD-processed versions of the audio signal. The results consistently indicated high ranking of features based on EMD and their ability to improve general detection accuracy once they were introduced to the set of audio features. Namely, the selected subset of combined audio and EMD-based features in comparison to all audio features, improved the detection accuracy by up to 10.28%. Scenario B contained 6 different cases, incorporating 3 extracted datasets generated from custom raw data and variable segmentation window lengths and 3 extracted datasets acquired through taking the simulated raw dataset with 3 different signal-to-noise ratio values for the fetal heartbeat sound signal. The results of feature selection and ranking methods indicated consistently high relevance of psychoacoustic features, especially in the case where frequency shifting was used as a preprocessing step. In addition to the random forest models trained with the selected feature subsets, two new classifiers (support vector machine with cubic kernel and bagged trees) were introduced to assess the impact of the entire sets of characteristics added through the proposed preprocessing and feature extraction methods. The analysis showed that the included dimensionality gained through the pitch shifting and EMD in the preprocessing steps with audio and psychoacoustic feature extraction raised the detection accuracy to a higher level, reducing misclassification rate up to nearly 3 times in some instances. As a final result of this research, the impact of the proposed preprocessing and feature extraction methods in the automatic FPCG classification was shown to be substantial. In addition, the showcased approaches have been demonstrated to be feasible for the implementation in an algorithm for real time usage, further highlighting the possible benefits of the system and its components in the biomedical industry.