There is an urgent need to develop out-of-hospital methods for monitoring health. For instance, over 10% of infants develop with a risk of lifelong neurodevelopmental compromise, however an optimal treatment would require very early detection of abnormal motility. The greatest bottlenecks in achieving this goal are caused by poor scalability and high subjectivity in the current diagnostic methods. Machine learning based solutions hold great promise in alleviating these problems, but the application of state-of-the-art machine learning methodology to novel clinical data domains involves many challenges. The AInfants project aims to close this gap by addressing two of the most pressing challenges: First, an efficient utilization of unlabelled data, and second, resolving the inherent ambiguity from inter-rater variability in annotated datasets. The developed methods are mainly applied in the context of infant movement analysis based on novel intelligent wearables.