Machine Learning Techniques to Predict Deterioration of Patients with Cirrhosis in Hospital Wards

Lead Supervisor
Dr Zina Ibrahim
Lecturer
Department of Biostatistics and Health Informatics, King’s College London
zina.ibrahim@kcl.ac.uk

Co-supervisor
Dr Mark McPhail
Senior Lecturer and Consultant in Liver Critical Care and Hepatology in the Institute of Liver Studies, King’s College London and Department of Hepatology at Imperial College

Project Details

Background:

Cirrhosis refers to scarring of the liver caused by continuous, long-term liver damage. Cirrhosis is a condition with slow progression, usually taking years to reach its end stages, and is often asymptomatic until the liver function becomes so severely compromised that the patient develops life-threatening complications such as jaundice, abdominal infections and gastrointestinal bleeding. Transplant is usually the only option for cirrhosis patients, without which only up to half of the patients exceed their 5-year life expectancy post diagnosis.

The project proposes the inception of novel explainable Machine Learning techniques, combining sophisticated knowledge representation models with algorithmic insight, to model and learn the complex temporal interactions embedded within a patients treatment timeline derived from Electronic Health Records (EHRs), to estimate the 2-, 5- and 10- year risk of cirrhosis in adult secondary care patients with abnormal liver blood test results. 

The project will use data from Kings College Hospital (KCH) EHRs as well as the Clinical Record Interactive Search (CRIS), a de-identified replica of the EHRs of the South London and Maudsley NHS. By linking physical and mental health EHRs, and given that depressive symptoms are known to be associated with worsened liver function, the project will focus no building robust, efficient and interpretable prognostic models for cirrhosis risk, using temporally-anchored physiological as well as psychiatric indicators mined from the sparse, noisy and irregular observations housed within EHRs. 

Primary aims:

This project aims to: 

a) Develop and evaluate a knowledge representation platform to enable modeling the multimodal, unevenly sampled, temporal and sparse physiological and mental health indicators of liver deterioration found in patient EHRs. 
b)  Develop and evaluate an interpretable cirrhosis risk evaluation predictor using machine learning models which make use of the knowledge representation framework developed. 
c)  Evaluate the models effectiveness in making personalized recommendations about cirrhosis prognosis to healthcare practitioners with regards to patients with abnormal liver blood results.  

Novelty and Importance: 

In the UK, driven by the upward trend in obesity and alcohol consumption, the prevalence of non-viral liver disease and those who develop cirrhosis is increasing, and whereas death rates in many other chronic diseases such as cardiovascular disease have improved over the last three decades, the number of premature deaths from cirrhosis continues to rise. The algorithms resulting from this work embody a much-needed tool for cirrhosis prediction. To our knowledge, apart from elementary prediction models, no such tool exists. The outcomes of this project go beyond patient stratification to developing explainable models of patient digital phenotypes based on patient-level longitudianal data. The scalability, immediacy and low cost of these digital phenotypes make their implementation into clinical practice highly likely if they are found predictive and instructive. Moreover, the algorithms resulting from this project will enable the creation of automated and explainable decision support tools providing patient-specific prognostic assessments for cirrhosis. 

Datasets

This project will use data from King’s College Hospital (KCH) EHRs as well as the Clinical Record Interactive Search (CRIS), a de-identified replica of the EHRs of the South London and Maudsley NHS. Permissions will be sought through existing governance infrastructures

Keywords

Machine learning, outcome prediction, Cirrhosis