The use of deep learning and inference to predict patient admissions to intensive care in healthcare

Lead Research Organisation: University of Oxford
Department Name: Engineering Science


With the NHS coming under increasing financial pressure it has never been more important to find improvements in the way the service runs and to optimise the delivery of healthcare. It is estimated that to keep a patient in hospital for a day in the UK costs the NHS approximately £413 and to keep a patient in the intensive care unit (ICU) almost increases the cost by five times at £1932. Reducing the rate of unplanned admittance to ICU is a strong NHS priority, by identifying physiological deterioration in advance of it occurring, such that adverse outcomes (including unplanned ICU admission) can be avoided. Recent advances in machine learning have given us the capability to acquire and perform inference with large quantities of data simultaneously from multiple sources. An example of this in hospitals is electronic health record (EHR) data, which records a multitude of patient vital signs as well as the patient's condition along with the results of blood tests, etc. With access to these data "big data" techniques involving deep learning and probabilistic inference can provide prediction of patient condition rather than merely identifying adverse events as they occur, which is the current clinical standard. The other failure mode of current systems is identifying a patient when they need to be escalated to a higher level of care. In the USA 200,000 patients each year suffer cardiac arrest while in hospital, of which 75% die. Predictive inference has a key role to play in reducing these, and other related, avoidable mortality rates.

Given the wealth of data available from EHR systems, I propose to use deep learning to carry out the prediction task of when a patient will be admitted to the ICU. This will investigate the use of time-series models (such as Bayesian Gaussian processes), provided as input to deep recurrent neural networks, such as long short term memory (LSTM) networks. Advantages of this approach to predictive inference when compared with existing methods include the capacity of LSTM (and related architectures, such as recurrent versions of variational autoencoders) to perform the role of automatic feature extraction. As EHR data is recorded via a multitude of acquisition methods such that the data recorded re inherently noisy and irregularly sampled, this would allow the creation of a personalised, predictive monitoring system which would be compared with existing population-based methods. The reason for choosing Bayesian Gaussian processes combined with deep recurrent neural networks is due to their ability to not only handle time-series data but also due to their ability to cope with inconsistent time sampling of data.


10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/R512333/1 01/10/2017 30/09/2021
1939630 Studentship EP/R512333/1 01/10/2017 30/09/2021 Rasheed El-Bouri
Description We have developed novel machine learning algorithms for use with medical data from electronic health records and medical sensors, utilising deep learning, curriculum learning and reinforcement learning methods.

We have proposed a framework for improving patient flow through emergency departments in hospitals using predictive inference and have shown promising results. We are implementing an empirical test to see how real world results compare to performance on historical data.

We have also developed a method of predicting readiness to discharge from hospital for patients already within the hospital.
Exploitation Route The work developed here can be used by hospitals firstly for the prediction of where beds will be needed by incoming patients to the emergency department. Our work aims to predict this at point of entry and so provides hospitals with enough notice to ready the wards downstream for patient arrivals. We believe this will reduce crowding of the emergency department and improve waiting times.

The work can also be used by other researchers aiming to work in patient flow prediction and the scheduling of hospitals.

Our methodologies that we have developed can be used by people working in machine learning generally whether working with medical data or not.
Sectors Digital/Communication/Information Technologies (including Software),Healthcare