Machine learning in emergency care

The intensive care unit (ICU) is a specialist department of a hospital or healthcare facility that treats acutely ill patients in need of comprehensive and continual observation and treatment. ICUs were created so that specialised units, utilising close monitoring and treatment of patients, could improve outcomes including risk of sepsis (which can lead to organ failure) and mortality, by maintaining a constant understanding of each patient's current state. It has been shown that the higher number of staff and the increased level of supervision in ICUs have lead to reduced incidents of mortality, lower hospital length of stay, and fewer illness complications, supporting the efficacy of an intensive monitoring approach. Nevertheless, increased support incurs increased costs, and modern hospitals are forced to restrict the number of nurses and doctors attending to patients in the ICU. For example, ICUs cost $81.7 billion in the US, accounting for 13.4% of hospital costs and 4.1% of national health expenditures. Moreover, between 2000 and 2005, the number of hospital beds in the United States shrank by 4.2%, but the number of critical care beds increased by 6.5% with occupancy increasing by 4.5% - underlining the escalating demand placed on ICUs [1]. As such, a major aim of this PhD will be to demonstrate that deep learning is now advanced enough to reliably intervene in the ICU, combatting the strain placed on trained specialists in critical care.

During the next decade, the data-hungry appetite of deep networks will be sated by the wealth of measurements generated through comprehensive monitoring of ICU patients. The information present in modern ICU databases will, in principle, detail the majority of factors required to optimally diagnose and medicate a patient. However, the procurement, analysis and interpretation of this data in a clinically relevant and comprehensible format is a major challenge of data analysis in critical care. Current inconsistencies exist due to various reasons, including: differing in-hospital procedures, human error, and unrecorded observations which affect the tests chosen by clinicians. Additionally, the information is present in a plethora of formats including lab results, clinical observations, imaging scans, free text notes, genome sequences, continuous wave-forms and more. Therefore, this PhD will also endeavour to use network architectures that can effectively accommodate cross-modal data, while maintaining simplicity and interpretability.

The clinical aids developed during this project would present several pragmatic strengths upon deployment. Deep learning models are free from factors such as overconfidence, fatigue, and time pressure. They avoid any overreliance on intuition when there may be insufficient expertise to justify its use, and can incorporate vast amounts of data into one decision.

[1] Alistair EW Johnson et al. "Machine learning and decision support in critical care". In: Proceedings of the IEEE 104.2 (2016), pp. 444-466.