Machine learning to identify temporal patterns of neonatal movements which predict neurodevelopmental outcomes

Neonatal movements are necessary for the correct wiring of the developing sensorimotor nervous system, according to animal models. Concordant with this, in human neonates, reduced or abnormal motor activity, including after neurological compromise like hypoxia-ischemia, makes it slightly more likely that they will grow up with motor control problems (e.g. difficulty walking, clumsiness). However, the link is not strong enough for doctors to make accurate predictions. You will use machine learning to improve the way that movements are analysed and make predictions more accurate, including by taking account of the temporal variability which reflects sleep cycling. (In Figure, the transition from high to low motor activity approximately half-way through is a sleep state switch). You will learn about perinatal and sleep neuroscience and neurology, physiological time series analysis, machine learning approaches like Rough Path Theory, and deep learning methods such as Convolutional or Recurrent Neural Networks and transformers, together with Dynamic Time Warping.
In Year 1, you will use Dr Whitehead’s existing dataset of neonatal movement time series to model long- and short-range temporal fluctuations in motor activity. (These time series comprise video frame-by-frame changes in pixel intensity within regions-of-interest (neonate’s limbs)). In Year 2, you will use the neurodevelopmental outcome data associated to these neonatal movement recordings to identify which temporal characteristics are prognostic. In Year 3, there will be potential to develop the project in line with your interests. One direction is to extend data analysis to fetal movements, in conjunction with Dr Whitehead’s collaborator Prof Mary Rutherford’s large MRC-funded fetal research project ‘MiBirth’. Another option is to examine how neonatal movements interact with cortical activity, in recordings acquired synchronised to EEG.