Supported via arm's length funding from Bayer

Highlights from

ISTH 2019

The International Society of Thrombosis and Haemostasis

Melbourne 6-10 July 2019

Could machine learning improve upon the Wells score for DVT risk stratification?

Take-home messages
  • A machine-learning algorithm has been shown to safely exclude DVT without the need for ultrasonography
  • The algorithm identified young age and male sex as the highest risk factors for DVT
  • Future trials are warranted to test and improve the algorithm, which was developed for this proof-of-concept study
"There are many fields of science where machine-learning algorithms have really taken a hold and are driving forward research. I don’t think that's yet the case in our field."

Dr John Willan, Consultant Haematologist, Haemophilia and Thrombosis Centre, Oxford University Hospitals, Oxford, UK.

Machine-learning algorithms may have improved capabilities in assessing the risk of deep vein thrombosis (DVT) compared with current methods, show new findings presented at the International Society on Thrombosis and Haemostasis (ISTH) 2019 Congress.

When a patient presents with suspected DVT, the current approach is to use a points-based score, such as the Wells score, typically combined with a D-dimer test.

"There are many fields of science where machine-learning algorithms have really taken a hold and are driving forward research. I don’t think that's yet the case in our field, but I do wonder if this may become more important," explained Dr John Willan, Consultant Haematologist, Haemophilia and Thrombosis Centre, Oxford University Hospitals, Oxford, UK, who presented his findings.

Machine-learning algorithms do not require linear relationships between variables, and they can analyse complex data without a pre-existing hypothesis, so should, theoretically, be an ideal diagnostic tool.

"I was interested to know whether we could use machine-learning algorithms in our field to improve the risk assessment of patients who have been referred to a DVT clinic with a possible DVT… could we safely exclude DVT in more patients than we currently do, without the need for scanning?"

Dr Willan retrospectively analysed data from 11,490 records of patients with suspected DVT in a single diagnostic clinic. Of these, 7,080 records contained the required 13 data points: the ten components of the Wells score, age, sex, and D-dimer. Of this, a 'training set' of 5,270 records were used to develop the algorithm, and 1,810 'unseen' records were used to test its performance. All records included an ultrasound result; 12.5% of patient records studied had confirmed DVT.

The algorithm was able to exclude DVT in 37.5% of patients, with no need for ultrasound (false negative rate: 0.22%), compared with 8.3% using the Wells score and D-dimer. "We're actually picking up additional patients here that wouldn’t be scanned by conventional systems, but the network would recommend, actually, these are quite high risk… [some] with a low D-dimer; less than 300 [ng/mL]."

The study also produced some unexpected data, noted Dr Willan. "One thing I wanted to draw to your attention to here is the difference between men and women that has come out of this algorithm" he continued, "[there is] a higher D-dimer threshold before a woman would be recommended for a scan using this algorithm. For instance, a patient with a Wells score of 1 because of a previous DVT, who has a D-dimer of 500 [ng/mL] and age of 60: if they're male, they’d be scanned, but if they're female, it would recommend not scanning in this particular algorithm. And that came up consistently."

The algorithm also generated intriguing data with regards to DVT risk in active cancer. Because D-dimer is elevated in patients with cancer (even if they have not had a previous DVT), the threshold indicating the need for a scan is increased. "I think what this algorithm is saying is, if you've got active cancer but your D-dimer is only 1,000 [ng/mL], then really, you're at very low risk of having a clot."

Further studies are warranted, however, as this analysis included retrospective, single-centre data. "This is not a model we can use immediately - it's a proof of principle. What I'd then like to do is gather prospective data from multiple centres," he explained, noting that the algorithm could be improved to include additional factors, such as oestrogen use, height and weight, ethnicity, neutrophil and platelet count.

"I think these kind of algorithms may represent a way forward," he concluded. The tool is available online and can be used to experiment with hypothetical cases (https://oxforddvt.github.io), though is not validated for clinical use.

Dr Willan briefly explored the potential reluctance people may have to uptake new artificial intelligence-directed diagnostic approaches, in general, explaining that people tend to distrust the 'black box' or tools they cannot scrutinise themselves.

Founder of the Wells score, Phillip Wells, however, commended Dr Willan's work after his talk at the ISTH Congress, and recommended that clinicians embrace machine learning and its huge potential for the future of diagnostic medicine.

Based on Willan J, Katz H, Keeling D. Machine-learning algorithms can improve upon current methods of risk-stratification of patients presenting with suspected deep vein thrombosis (abstract OC 63.3). Presented on Tuesday 9 July 2019.

Top image: PhonlamaiPhoto

The content and interpretation of these conference highlights are the views and comments of the speakers/authors.

Supported by