Using machine learning to predict treatment pathways in end stage kidney disease (ESKD)

Lead Supervisor
Dr Mariam Molokhia
Lecturer in Psychosis Studies
Population Health Sciences, SPHES, FoLSM, King’s College London
mariam.molokhia@kcl.ac.uk

Co-supervisor
Professor Claire Sharpe King’s College London
Dr Kateie Vinen King’s College London
Dr Kathleen Steinhöfel (joint 1ry) King’s College London
Professor Dorothea Nitsch London School of Hygiene & Tropical Medicine (LSHTM)

Project Details

Background: Chronic kidney disease (CKD) is common, found in 14% of people aged 65-74 and 33% of people aged ‚â•75 years in England [1, 2], and the prevalence increases with age. [3] In some patients, kidney function continues to slowly decline, resulting in end stage kidney disease (ESKD). Patients with ESKD either choose to forgo renal replacement therapy and be managed conservatively for symptom control and quality of life,  or choose dialysis. Currently, we are unable to identify patients who are being managed conservatively, although a single previous survey suggested significant variation between centres for uptake of this pathway. [4] This results in an inability to define and measure quality standards for these patients and may result in suboptimal care.  As the majority of these conservative care patients are elderly, the lack of data has been particularly highlighted in the  context of COVID-19, but also limits our ability to provide renal care and to understand management of these patients and their outcomes.

Aims: To develop risk prediction algorithms for differentiating individuals preparing for dialysis from those receiving comprehensive conservative care. 

Investigation plan:

This PhD project will uniquely attempt to determine the number of people who have advanced kidney failure who have opted against dialysis using UKRR data by using machine learning (ML) methods. To inform the ML approach we will use a local database, Lambeth DataNet and then apply to the UKRR dataset. Once these patients have been identified,  we can prospectively investigate their outcomes including during the COVID-19 epidemic. 

Objectives:

  1. To use a training dataset from a single renal-centre’s record (KCL) of individuals preparing for ESKD to develop algorithms predicting whether a person is preparing to receive dialysis or comprehensive conservative care.
  2. To use a training dataset split from a Primary Care record (Lambeth DataNet) to develop algorithms that are able to predict whether a person is preparing to receive dialysis or comprehensive conservative care.
  3. To identify data items in the RenalWare (CKD data warehouse) which enhance algorithm accuracy.
  4. To internally validate these algorithms using records from the original datasets.
  5. To compare the accuracy of the ML approaches with traditional statistical methods for modelling planned treatment using individual patient data.
  6. To use the findings of 3) to inform expansion of the UKRR Dataset, to develop personalised risk prediction algorithms identifying the UKRR treatment pathways.

Machine learning approaches have been successfully used in risk prediction in hospital for acute kidney injury  including neural network assessing sequential data inputs over time. Analyses will apply the artificial neural network (ANN) learning algorithm by fitting an optimized model to training data. 

The project will explore differences and commonality between results found using traditional statistical methods and ANN methods to determine whether results are similar and evaluate the robustness of each model using receiver operating characteristic curve (ROC statistic) for sensitivity and specificity of both models in training and validation data cohorts.

Skills training: epidemiological, statistical evidence synthesis and machine learning approaches, using two large UK population datasets. 

References:

  1. England, P.H., Chronic Kidney Disease (CKD) prevalence model 2014 P. 3 number: 2014386. p. P. 3 number: 2014386.
  2. registry, R.: p. Chapter 2 Figs  2.1-2.7.
  3. Chronic Kidney Disease Prognosis, C., et al., Association of estimated glomerular filtration rate and albuminuria with all-cause and cardiovascular mortality in general population cohorts: a collaborative meta-analysis. Lancet, 2010. 375(9731): p. 2073-81.
  4. Okamoto I, Tonkin-Crine S, Rayner H, Murtagh FE, Farrington K, Caskey F, Tomson C, Loud F,Greenwood R, O’Donoghue DJ, Roderick P. Conservative care for ESRD in the United Kingdom: a national survey. Clin J Am Soc Nephrol. 2015 Jan 7;10(1):120-6. 

PhD Stages:

Yr1 Develop risk prediction algorithms in  primary and secondary care and data items to enhance accuracy.
Yr2 To internally validate these algorithms using records split from the original datasets.
Yr3 To pilot and validate use of personalised risk prediction algorithms for identifying patient treatment plans, comparing traditional and Machine Learning methods.  

Datasets

Renal Registry: The UK Renal Registry (UKRR) charity is part of the Renal Association. The UKRR collects, analyses and reports on data from 71 adult and 13 paediatric renal centres. UKRR has data on renal patients who are under the  care of a renal consultant, those on dialysis, and those with a kidney transplant.  It has only limited data on patients with stage 4 and 5 kidney disease who are not yet receiving kidney replacement therapy.   Participation is mandated in England through the NHS National Service Specification and the Chief Executive of each Trust is responsible for adherence to this contract. We are currently in the process to adapt our existing permissions to link to PHE infection data to include COVID-19, and research ethics permissions are being sought.

Keywords

renal replacement therapy, dialysis, conservative care,chronic kidney disease (CKD)