COVID-19 National Notification & Digital Tracing Service

Pioneering patient predictions.
Chronic obstructive pulmonary disease (COPD) is a group of lung conditions that cause breathing difficulties. It affects 1.2 million people in the UK with a yearly £1.9 billion cost to the NHS.
Our collaboration with the NHS looked to help patients with COPD. A clinical study used machine learning to identify patients at highest risk of adverse events related to COPD. The study builds on earlier work to develop a COPD app, which helps patients manage their condition. Users of the app have been given the option to take part in the study.
Number of emergency hospital admissions that are related to COPD symptom flare ups
Reduction in admissions during clinical trial
Annual bed days saved per patient
"This is the first time we’re bringing predictive AI insight for COPD into live clinical practice."
Storm ID pioneered the development of machine learning models that can proactively identify high-risk patients and provide explainable, actionable insights. Moving from a reactive model of care to a preventative one could have hugely positive impacts. Both on patient outcomes and emergency care requirements.
We trained the algorithms used on a massive dataset of historical health records, whilst ensuring any pre-existing biases in the data were accounted for. This ensured fairness across age groups, gender, deprivation, and ethnicity.
It's important to present the outputs and insights from these models to patients and clinicians in a clear and consistent way. The design system we built underpins the user experience. Repeatable, carefully designed components - navigation, forms, charts, data visualisations - have been tested and validated extensively with patients and clinicians. These components were key in aiding understanding of patient health.
To support all this work, we also conducted interviews with users of the COPD app. We wanted to understand how they feel about the use of machine learning and AI to make risk predications about their health.
Interim results from the clinical study showed the models to be clinically safe, technically feasible and highly predictive in identifying patients at risk of mortality and hospital admission.
From qualitative research we were surprised by how much patients already knew and were comfortable about AI and its application. Transparent, human explanations of a service's benefits are more important to patients than the model or method used. As one participant put it, "Any tool that a doctor can use to make better predictions has got to be a good thing."
We anticipate that the learnings from this clinical study in relation to model development, model version control, explainability, clinical acceptability and data interoperability will support the adoption of AI across chronic conditions.
Indeed, the use of LLMs like GPT4 that are capable of analysing and interpreting large amounts of text and image data, as well as perform complex reasoning tasks represent the new state of the art for predictive analytics and will prove to be far more generalisable across healthcare systems.