Movements of the human body look very simple but consist of complex coordination systems, subsystems, and monitoring pathways. Any disorder in the coordination system like progressive neuromuscular disorders, injuries to the brain, and genetic disorders can create problems in movement and posture. For example, cerebral palsy (CP) describes a group of disorders of lifelong physical disability caused by a non-progressive brain injury or lesion acquired during the antenatal, perinatal, or early postnatal period [1]. The severity, patterns of motor involvement, and associated impairments, such as communication, intellectual ability, and epilepsy, vary widely and persist across the life course [2]. In addition, neonatal mortality has decreased in preterm infants in the past decade, extremely preterm infants (born at <27 gestational weeks) remain at the highest risk for neonatal morbidity and the occurrence of CP [3]. Therefore, the prevalence of CP has remained stable over the last forty years at 2–3 per 1000 live births in countries with a developed health care system.

       At present, there are no uniform clinical procedures for the prediction of motor impairments like CP in high-risk infants and the recognition of those at the highest risk generally requires the combination of clinical history, various clinical assessments and expertise of the observer [4]. Some studies, e.g., [5,6,7], have exposed the fact that early recognition of motor impairment leads to early interventions that might reduce the severity of the motor impairment and the restrictions in daily activities [8].

       Prechtl presented the General Movements Assessment (GMA) as a valuable tool for the prediction of cerebral palsy in high-risk infants [9,10]. General movements (GMs) are spontaneous movements of infants up to five months post-term involving the whole body. The movements vary in sequence, speed, and amplitude. Depending on the infant’s age, one distinguishes between the general movements (GMs) (preterm general movements (∼28–36/38 gestational weeks) or term/writhing movements (36/38–46/52 gestational weeks)), and the fidgety movements (FMs) (46/50–55/60 gestational weeks) [9]. Next to normal GMs and normal FMs (F+ or F++), one distinguishes between poor repertoire GMs (PR) with a monotonous sequence of movements and reduced variance in speed and amplitude of movements, cramped synchronized GMs (CS) which appear stiff with bilateral contraction and relaxation of the legs and the abdominal wall, and chaotic GMs (Ch) which appear jerky, rowing, fast, and have a large amplitude. The non-normal FMs comprise abnormal FMs (AF) with large amplitude, fast and jerky movements, as well as the absence of FMs (F−). Showing cramped synchronized or chaotic GMs around term or the absence of fidgety movements (F−) at 3 to 5 months post-term have an excellent predictive value for cerebral palsy [11,12]. However, the assessment is based on videos of the infant that are rated by trained professionals, therefore, the method is time-consuming and expensive. 

       As a result of the nominal use of GMA in neonatal follow-up programs, several studies have tried to automate this method. These studies are based on either indirect sensing using visual sensors (2D or 3D video) [7,13,14,15,16,17,18,19,20,21,22,23,24,24], direct sensing using motion sensors [25,26,27,28,29,30,31], or both [32,33,34]. They have shown excellent results, however, they lack full automation and also have several fundamental limitations. First, all the studies are either based on a small number of subjects or a fewer number of data samples with respect to CP [7,18,19,20,25,26,27,32,34]. It is also not clear if the prediction model in these studies has external validity for high-risk infants. Second, the research work in some studies is based on convenience samples that do not reflect the usual clinical cohorts. Third, the movement features used in previous studies lack generality due to less number of subjects and examples. Lastly, all the reviews, except [17,20,23,24,35,36], are not using state-of-the-art Deep Learning (DL) algorithms to automate the GMA process. The DL algorithms are popular approaches of Artificial Intelligence (AI) which not only provide a generalized solution but also perform well for accurate detection of the classes in visual and time-series data. Therefore, an end-to-end system is needed to analyze the infant’s movements in the early infancy.

       There are some related review articles for monitoring body movements of infants using sensor technology. Chen et al. [37] outlines the wearable sensor systems for monitoring body movements of neonates apart from visual sensors and state-of-the-art AI algorithms for the development of an automated end-to-end system. Zhu et al. [38] present a broad overview of wearable sensors intending to measure various types of physiological signals of infants. The authors in [39] discuss state-of-the-art movement recognition technology for assessing spontaneous general movements in high-risk infants, however, they do not focus on the design and development of the system. They discuss the wearable and visual sensors averagely. Zhang [40] review machine learning methods in cerebral palsy research and evaluates algorithms in movement assessment for CP prediction.

       The primary objective of this article is to systematically analyze and discuss the main design features of all existing technological approaches trying to classify the general movements of infants and explain the methodological reasons for their limited practical performance and classification rates. The main contributions of this paper can be summarized as follows:

• We present a structured review of the current technological approaches that detect general movements and/or fidgety movements, and categorize them according to the AI techniques they use. We slice up these approaches into three vital categories: visual sensor-based, motion sensor-based, and multimodal (fusion of visual and motion sensory data).
• We categorize and present a summary of the sensor technology and classification algorithms used in the existing GMA approaches.
• We also present a comparative analysis of reviewed AI-based GMA approaches with respect to input-sample size, type of features, and classification rate.

       Prior to continue, it is worth noting that the correct classification of GMs is a difficult task and relies on clinical expertise. While some previous (machine learning) studies evaluated the ground truth of their data by introducing trained GMA experts, some recognized ambiguous, arbitrary, or incorrect classification or did not present detailed information about the realized process. In order to provide an objective overview, we nevertheless indicate the classes and terms specified in the papers and highlight if the classification was not carried out properly. Moreover, this article does not talk about preprocessing operations, for example (image enhancement, noise attenuation, finding the region of interest, etc.), since they fall outside from the scope of this article. In addition, we duly note that understanding this paper requires knowledge of machine learning concepts and performance evaluation techniques of classifiers. An extensive but straightforward explanation of these concepts can be found in [41,42].

This entry is adapted from the peer-reviewed paper 10.3390/s20185321