Development and Feasibility Evaluation of a New Standard of Care for Investigation of Patients with Kidney Stone Disease: A Multi-Omic Machine Learning Approach

Kidney stone disease (KSD) is the common endpoint of a heterogenous group of disorders.

It is painful and potentially life threatening, with well documented impacts on quality of life, as well as high costs to the NHS (~£0.25 billion in England alone each year). Half of all patients with KSD will have a recurrence within 5-years.

Current standards of care for investigating underlying causes of KSD are limited to serum and urine biochemistry, which are demonstrably ineffective screening tests.

NICE recognise this and have recently called for research into the clinical and cost effectiveness of a full metabolic assessment compared to standard advice alone.

I have previously built machine learning (ML) models using the current standard of care (full metabolic assessment) to predict recurrent disease. These models were ineffective, and suggest that alternative predictors are needed. Our PPI work has demonstrated that patients want to know their risk of recurrence, and avoid emergency and operations.

This has therefore guided the rationale behind developing a new standard of care in the form of an AI model for recurrence prediction. The outcome of this model being a composite including emergency presentations and operations. There is increasing evidence of a genetic influence on KSD.

In a preliminary study I built a ML model predicting recurrence utilising genetic data (based on a published gene panel) along with serum biochemistry. This had a significant increase in prognostic accuracy compared to serum biochemistry alone.

However, this model utilises a relatively small dataset derived only from the UK Biobank with a large number of input variables, which is clearly not clinically practicable.

I therefore aim to develop and externally validate a clinically useful and cost effective ML model for recurrence using demographic, biochemical, dietary and genetic data from both UK Biobank (UKB) encompassing known risk factors for KSD including whole exome sequencing (WES). The outcome will be composite as guided by our PPI group.

I will perform extensive feature engineering to ensure the model is clinically practicable by reducing the number of input factors required, whilst maximising performance. I will then externally validate this model with data from All of Us (AoU - American equivalent). I will examine the cost effectiveness of implementing the ML model in practice with Markov modelling.

I will also examine various thresholds of predicted risk to facilitate early discharge. This will allow cost saving, reassurance to low risk patients and resource reallocation to high risk patients. An effective model has the potential to be practice changing. PPI has been essential in guiding this project, and will continue to be integral as it moves forward.

I aim to expand and formalise the existing PPI group in Newcastle. I aim to widely distribute project findings.

This will be to academics through peer-reviewed journal articles and conferences, and to patients and the public via social media, websites and in-person meetings as guided by PPI members.