Machine learning to help cystic fibrosis decision-making
03 Aug 2018by Chloe Green
New research claims to have demonstrated that machine learning techniques can predict with a 35% improvement in accuracy – in comparison to existing statistical methods – whether a cystic fibrosis patient should be referred for a lung transplant.
, led by Professor Mihaela van der Schaar of the Alan Turing Institute
at the University of Oxford, has been generated through a partnership between The Alan Turing Institute and charity the Cystic Fibrosis Trust
It is the first machine learning study to make use of a dataset representing 99% of CF patients living in the UK, the CF Registry, scientists have said.
Cystic fibrosis is a genetic condition that affects more than 10,000 people in the UK, and causes a wide range of symptoms that affect the entire body. Lung transplants may be recommended as the last treatment option, when others no longer have an impact. However, the procedure comes with the risk of post-transplant complications, and the availability of lung donors is scarce compared to demand.
To explore whether the lung transplant process could be improved through machine learning methods, the Cystic Fibrosis Trust supplied lead author Mihaela van der Schaar with access to an anonymised extract of UK CF Registry data.
Using the dataset, Mihaela and co-author, PhD student Ahmed Alaa (UCLA), developed an algorithmic framework that leverages machine learning to automate the process of constructing a prognostic model for CF patients – the point at which a clinician assesses a patient and calculates the risks of taking a certain pathway against the projected benefits.
The new algorithmic model, called AutoPrognosis, is capable of achieving a positive predictive value of 65%. Existing practice results in only 48% of individuals being correctly referred, so AutoPrognosis yields a 35% increase in accuracy overall in comparison to current methods, it is claimed.
The authors and experts from the Cystic Fibrosis Trust believe this research could be used to significantly augment clinicians’ decision-making processes. Use of machine learning to develop improved prognostic tools, provides additional information for clinicians and patients to assist in discussing future treatment options.
Furthermore, the machine learning methods developed in this study could be applied for other diseases in future, for example heart attacks or cancer diagnosis, they added.