Cystic fibrosis is a genetic condition that affects more than 10,000 people in the UK and causes a wide range of challenging symptoms which affect the entire body. Machine learning, training a computer to learn the mapping between inputs and desired outputs without explicitly programming it to do so, has the potential to improve treatment plans for people living with this life-limiting condition.
Explaining the science
The Cystic Fibrosis Trust’s video on what it’s like living with CF, in the words of someone with the condition
People living with cystic fibrosis (CF) frequently want to know why their health is like it is currently, and what will happen to their health in the future.
In partnership with the Cystic Fibrosis Trust, Turing researchers have access to an anonymised extract of UK CF Registry data. The Registry is a secure centralised database managed by the Cystic Fibrosis Trust, which holds consented health data from over 99% of the people with CF across the UK. Click here for more information about the Registry, including how the data is kept secure. The aim of this project is to utilise machine learning techniques to train a system to learn and make reliable predictions from this historical data, creating a method of generating personalised risk scores for people with CF.
The hope is that these scores will then be used by people with CF and their clinical teams to tailor treatments and other activities to effectively manage the condition. Intelligent risk adjustment methods would also support clinical teams to monitor and continuously improve the clinical care they provide.
Unlike current prognostic tools and practice, a data-driven model can provide a full risk profile for various timescales and potential treatments, as well as take into account multiple, competing, potential adverse events for an individual. For CF these adverse events can be severe, include lung transplant, related diabetes, respiratory, pancreatic and renal complications, and ultimately death. The model will also be able to be tailored to an individual’s demographic, environmental, microbiological, and genetic traits.
These more accurate forecasts of the course of the disease, for each individual patient, have the potential to improve treatment, quality of life, and survival for many people with CF.
This methodology also has the potential to be customisable to other diseases such as cardiovascular disorders and asthma.
A machine learning algorithm was used on the CF Registry data to learn an accurate 'decision-making policy' for when an individual with CF should be referred for a lung transplant. This was possible by looking at the health of individuals in the dataset both before and after they had a lung transplant. The weightings of the various health indicators and risks associated with CF were varied and rearranged, to find which combination would provide the most accurate predictor of whether having a lung transplant is the best treatment for any particular individual.
35% - Potential improvement in lung transplant referral accuracy
To test the accuracy of this new decision making policy it was re-applied to the CF Registry dataset, to see how well it would have hypothetically performed for the individuals included. If the data-driven policy had been used for these individuals: of those who would have been referred for a lung transplant, this would have been the best course of action for 65% of them. Using current, existing practice results in only 48% of individuals being correctly referred.
Therefore, as a percentage increase on current practice, the method could provide doctors with a 35% improvement in accuracy of lung transplant referral. Whilst this development is potentially exciting, further validation and testing will need to be conducted to confirm its efficacy.