Use of multiple outputs from a NICU incubator for feature set development and utilisation of an automated sepsis warning system via

Neonatal Intensive Care Units (NICUs) look after the sickest of newborn babies. They are often very premature and underdeveloped and may have several life-threatening conditions including sepsis. The baby will be in an incubator and be comprehensively monitored by sensors that independently output multiple streams of key physiological data such as heart rate & rhythm, oxygen saturation levels, respiration, and temperature every second as well as ECG data at 500Hz. This information is displayed on monitors to identify immediate problems. However, this extensive data has the potential to provide real-time warnings for forthcoming physiological changes in the baby that may require intervention. Published literature and recent investigations
by MSc students in our group, have suggested that subtle changes in physiology suggestive of sepsis can be detected as early as 48 hours before a clinical suspicion of sepsis and changes in one or more of the recorded streams could trigger an alarm long before the baby exhibits the usual symptoms the medical team might recognise. The proposed research builds on the existing three-party collaboration between COMSC, MEDIC and the SKS Hospital in India and we have set up a repository at Cardiff University to store NICU data continuously collected from our partners at SKS. Preliminary results from our four recent
MSc projects suggest that sepsis can be detected significantly earlier than the appearance of clinical signs and symptoms. This studentship aims to exploit rich NICU data and create novel machine learning models to aid the development of a Sepsis early warning alarm system that could be commercially available and in use in NICUs worldwide in the next decade. The objectives of this research are to:
- Prepare the data for downstream tasks and conduct exploratory data analysis to gain
insights to assist feature engineering and data modelling.
- Through data manipulation and signal processing extract features from raw data in both
the time domain and frequency domain to allow machine learning model development.
- Create accurate machine learning models for early detection of neonatal sepsis. This will
include general machine learning (e.g., SVM, ensemble learning) and deep learning (e.g.,
recurrent networks and transformers).
- Carry out extensive evaluation using new collected data from SKS Hospital and publicly
sourced datasets.