An Intelligent Assistant for Medical Doctors when Prioritising Pathology Results

The project is supported by the University of Liverpool Doctoral Network in Future Digital Health, which is directed at creating and maintaining a community of AI health care professionals that can realise the benefits that AI can bring to Health Care. The network will be providing doctoral training, culminating in a PhD, in a collaborative environment that features, amongst other things, peer-to-peer and cohort-to-cohort based learning.

In any hospital, millions of computerised pathology results are produced every year. The process involves robotics, technicians, analysers and scientists to provide a view of patient conditions to the medical team who are looking after them. When the pathology results have been generated, the current hospital system calibrates the results against the normal range for a patient of their age and gender, and makes them available to doctors for decision making. Some doctors may well have hundreds of such results to review on each shift and it can take hours to go through them all and understand if they are worthy of action or simply to be endorsed as appropriate. Doctors use their medical training to decide what should happen based on:

• Working diagnoses/conditions
• Which medications the patient is taking
• Trends in previous results

Based on the result they may decide to take further action such as a change in medication, a change in the level of care or even that it is safe to discharge the patient home.

The PhD programme of work will be directed at investigating machine learning techniques that can be applied to help prioritise pathology results to save the doctor time and, potentially, to advise on further course of action. More specifically to investigate how ordinal relationships across priority levels can be learnt and applied in a real setting. Fundamentally there are two approaches that can be adopted; either the problem can be treated as a multiple regression problem or as an ordinal classification problem. Both approaches have advantages and disadvantages. Regression operates using a single weight vector where the weights are learnt from examples; what is known as the “one model” assumption. Using standard regression [1] or support vector regression [2] the result is a sequence of equally spaced parallel decision boundaries, an unlikely eventuality given a large number of data attributes. The difference between “excellent” and “good” might be small, while the difference between “poor” and “unacceptable” might be large. One solution is to use an ordinal regression approach, such as Support Vector Ordinal Regression [3], which avoids the equal spacing assumption but still entails a single weight vector and parallel decision boundaries. An alternative approach, which avoids the “one model” assumption is to adopt an ordinal classification approach [4]; although in this case, instead of predicting a continuous variable, discrete class labels (with an ordering imposed) will be used. The precise nature of the priority learning to be adopted will form a central element of the research. Once prioritised an additional step will be to propose a tentative course of action.

The evaluation of the techniques to be considered will have to take into account the ordinal nature of the prediction/classification; the error between “excellent” and “good” is not the same as the error between “excellent” and “unacceptable”. In the first instance Mean Absolute Error and Mean Square Error would seem to be appropriate although further investigation is warranted.

The work to be undertaken will be conducted in close collaboration with Wirral Teaching Hospital who will be providing consultancy and, more importantly, structured data spanning the last 10 years giving pathology results, the diagnoses, the outcomes, complications, lengths of stays in hospital, death rates etc. Whilst Wirral Teaching Hospital have comprehensive data for every patient, to start with the proposed study will focus on specific conditions and/or specific data attributes. For example the most common pathology tests might be considered first, such as Diabetes, Microbial Infection and Kidney or Liver disease (all of which have specific tests which would narrow down the related investigation on the subject).

The successful applicant should have (or be expect to achieve) a UK honours degree at 2.1 or above (or equivalent) in Computer Science. The successful applicant should also be familiar with the theory and practice of machine learning.

To apply for this opportunity, please visit: https://www.liverpool.ac.uk/study/postgraduate-research/how-to-apply/