Unsupervised Fetal Behavioral State Classification Using Non-Invasive Electrocardiographic Recordings Amna Samjeed, Maisam Wahbah, Ahsan H. Khandoker, Leontios Hadjileontiadis Khalifa University, Abu Dhabi, United Arab Emirates Abstract Understanding the Fetal Behavioral States (FBSes) is one of the ways to understand the fetal Autonomic Nervous System (ANS) maturation. This preliminary work aims to automatically classify FBSes using the unsupervised k- means clustering technique. Non-invasive electrocardio- gram signals were recorded from 67 healthy fetuses with Gestational Age (GA) range of 20–40 weeks for a duration of 10 min. Features extracted from the original fetal Heart Rate (HR) and detrended HR are used to classify the FB- Ses. Results showed that during the early gestational pe- riod, the prominent state was 1F compared to other states and the least common state was 4F. The decrease in 1F fre- quency and the increase in 2F frequency in late gestation represent the coordination and overall maturation of the fetal ANS. Results showed that the k-means clustering al- gorithm had good overall performance and stronger clas- sification ability with good Cohen’s κ score. Unsupervised classification of FBSes based on electrocardiography data is possible. It is achievable to incorporate this algorithm into future implantable devices for in depth understanding of fetal brain maturation and well-being. 1. Introduction Fetal Behavioral State (FBS) estimation is one of the ways to understand fetal Autonomic Nervous System (ANS) development [1]. Fetus exhibits four different be- havioral states: 1F (quiet sleep), 2F (active sleep), 3F (quiet awake-this state seldom occurs), and 4F (Active awake). The FBSes are usually defined by combining fe- tal Heart Rate Variability (HRV) and fetal movements [2]. These states are repetitive over time and exhibit distinct changes when moving from one state to another. As the gestation progresses, FBSes become more consistent and stable [3]. In clinics, fetal well-being is assessed using car- diotocography (CTG), which provides rudimentary infor- mation on fetal Heart Rate (HR) baseline and its variations. But it does not provide the complete electrophysiological analysis of the heart, which can only be obtained using non-invasive fetal electrocardiography (NI-fECG) or fetal magnetocardiography (fMCG). Compared to fMCG, NI- fECG is easier and cheaper[4]. NI-fECG allows detection of the morphology of signal (P, R, and T waves), and from this R-R series, HRV can be determined. NI-fECG ob- tained by placing electrodes on the mother’s abdomen has many advantages: continuous long-term monitoring, ease of use, and safety. However, NI-fECG is highly sensitive to background noise, maternal ECG, respiration, and motion interference [5]. Supervised learning is a process in which the model it- self is trained with the input data set, and a mapping func- tion is generated from the input data to the target data. While in the case of unsupervised learning, there is no training procedure, and thus structures in data sets have not been discovered [6]. The clustering technique is the most widely used method in unsupervised learning techniques. This tech- nique separates the data set into groups based on the sim- ilarity within the cluster and dissimilarity between other clusters (e.g., distance) [7]. Clustering aims to group het- erogeneous data into different homogeneous groups. The k-means algorithm is a widely used clustering technique in many fields. This algorithm divides n number of data into k number of clusters, where k needs to be defined initially. This will affect the k-means algorithm efficiency and out- come [7]. This preliminary work aims to classify the FBSes into the four states (1F, 2F, 3F and 4F) using unsupervised k- means clustering technique and analyze how efficiently the k-means algorithm could classify it. 2. Methods 2.1. Subjects and ECG Signal Processing Non-invasive Electrocardiogram (ECG) signals were recorded from 67 healthy fetuses with a Gestation Age (GA) range of 20–40 weeks for a duration of 10 min with participants in the supine position. These data set were obtained from Kanagawa Children’s Medical Cen-