ObjectivePremature babies have several immature functions and begin their life under high medical supervision. Since the sleep organization differs across postmenstrual age, its analysis may give a good indication of the degree of brain maturation. However, sleep analysis (polysomnography or behavioral observation) is difficult to install, time consuming and cannot systematically be used. In this context, development of new ways to automatically monitor the neonates, using contactless modalities, is necessary. Therefore, this study presents an innovative non-invasive approach to semi-automatize the classification of infant behavioral sleep states.MethodsFirst, three descriptors were extracted from audio and video recordings: vocalizations, motion and eye state of the baby. For this purpose, an original semi-automatic algorithm for the estimation of the eye state was proposed. Secondly, the three descriptors were used in order to obtain an estimation of the behavioral sleep states. Five classifiers (K-nearest neighbors, linear discriminant analysis, support vector machine, random forest and multi-layer perceptron) were compared to an expert annotation.ResultsFirstly, the comparison of the semi-automatic eye state estimation to manual annotations of 10 videos led to a mean accuracy of 99.4%. Secondly, sleep stage classification was performed. Best results were obtained with random forest, for quiet alert and active alert stages, with 93.5% and 99.0% of accuracy respectively.ConclusionThe proposed method provides a high capacity to identify alert sleep stages but the differentiation between quiet sleep and active sleep only by behavioral observations still remains a difficult task to achieve.SignificanceResults presented in this paper are new since no similar approach was proposed in the literature in the context of Neonatal Intensive Care Unit. They augur well for the automatic sleep organization assessment to improve newborn care.