Using Neural Networks to Assess Risk for Deep Vein Thrombosis

venous thromboembolism
The neural network developed in this study yielded a false-negative rate of 0.22%.

Neural network analysis may allow for improved risk assessment for deep vein thrombosis, according to results published in the British Journal of Hematology.

Artificial neural networks utilize machine learning to characterize nonlinear interactions and patterns in complex datasets. Risk assessment for DVT involves several factors, including patient demographics, clinical history, and D-dimer results. Most assessment methods aim to exclude a diagnosis of DVT without performing a compression ultrasound scan, which can be expensive and time-consuming.

Researchers hypothesized that neural network analysis may be more successful than conventional methods for diagnosing DVT without ultrasound scanning. They developed a network with an input layer consisting of 13 dimensions: age, sex, D-dimer, and the 10 components of the Wells’ score. The network was trained using data from 5270 patients who were evaluated on suspicion of DVT between January 2011 and December 2017.

Related Articles

The network returned a score between 0 and 1 for each patient, indicating the relative likelihood of that patient having DVT. The researchers chose a threshold of 0.1, below which they excluded a diagnosis of DVT without performing an ultrasound scan and above which an ultrasound scan was recommended.

The network was then tested using data from 1810 patients who were assessed for DVT alongside the training cohort. With a threshold of 0.1, ultrasound scanning was recommending for 62.5% of patients, with a false-negative rate of 0.22%. In comparison, ultrasound scans were recommended for 87% of patients using an age-adjusted D-dimer assessment strategy and 92% of patients using a D-dimer cutoff of 500 µg/FEU.

The authors noted that this network has not been validated and is therefore not currently clinically applicable. Furthermore, the network has limitations, as “it is not possible to demonstrate the influence of all possible combinations of patient criteria upon risk of DVT.” However, the network is available online for further experimentation to address this shortcoming. The authors concluded that with further prospective testing, a neural network approach may yield personalized risk assessment for DVT and decrease the number of recommended ultrasounds.


Willan J, Katz H, Keeling D. The use of artificial neural network analysis can improve the risk-stratification of patients presenting with suspected deep vein thrombosis [published online February 6, 2019]. Br J Haematol. doi:10.1111/bjh.15780

This article originally appeared on Hematology Advisor