Machine learning tool effective in early disease risk identification

A machine learning tool has inferred risk factors associated with over one thousand diseases, offering the possibility for improving early disease detection and drug discovery.

The tool, known as MILTON, was trained on data from around 500,000 participants from the UK Biobank, which included 67 biomarkers that are routinely collected at diagnosis such as blood pressure, body size measurements, and respiratory functions. From over 3000 diseases which MILTON was trained to predict, the model was considered 'highly predictive' for 1091 diseases, and 'exceptional' for 121.

'Our research demonstrates MILTON's capabilities and how it is able to identify disease risk cases in large biobank datasets, which in the future, could enable us to detect illnesses earlier and at more treatable stages' said Dr Slavé Petrovski, the head of the Centre for Genomics Research at AstraZeneca who co-led the study.

He added, 'Improving our ability to detect illnesses earlier and at more treatable stages is critical for early interventions in clinical care.'

MILTON was also designed to identify patient cases in clinical studies who may have been incorrectly labelled as controls, for example due to self-reported data. This was achieved by first training the algorithm on pre-diagnosed patients, and then applying it to the original set of control patients.

Retraining the model with these reclassified samples improved the statistical power of the subsequent genetic analysis and led to the uncovering of over 180 gene-disease relationships which were not identifiable using the original cohort.

Data on 3000 proteins found in blood plasma were also available for a subset of around 50,000 participants. Including these proteins when training MILTON led to improved prediction of many diseases, while likely not significantly increasing the costs of diagnosis.

The results of the analyses, published in Nature Genetics, have also been made publicly available, allowing other researchers to leverage them to improve their own diagnostic tools.

While AstraZeneca has clarified that MILTON is currently a research tool, the fundamental goal will be to apply the knowledge garnered from this tool within a clinical environment. One likely improvement may involve having to apply this approach to larger sample sizes, which currently limits the predictive confidence for some diseases.

Additionally, the tool's capabilities as a 'predictive' tool are being taken into consideration. Timothy Frayling, professor of human genetics at the University of Geneva, Switzerland, said 'We need to take care when claiming we can "predict disease" when we really mean "we can give you a slightly better idea of your chances of developing a disease but there are still many unknown factors."'

Furthermore, Dusko Ilic, professor of stem cell science at King's College London, warned '… while MILTON's capabilities are impressive, I have some concerns regarding its ethical use. The powerful predictive abilities of this tool could, if unregulated, be misused by health insurance companies or employers to assess individuals without their knowledge or consent. This could lead to discrimination and a breach of privacy' (see BioNews 1256).

He continued: 'Therefore, it is essential that the use of such systems be carefully regulated to protect individuals' rights and ensure that they are not subjected to unwarranted evaluations. Strict guidelines and oversight will be critical in ensuring that the benefits of MILTON are realised in an ethical and responsible manner.'

However, future work on MILTON may enable additional valuable insights to be made with regards to our understandings of disease risk and prediction.