Artificial intelligence is being widely employed across research and applied medicine, offering attractive new opportunities for diagnosis and treatment. Recently, a team of research groups from the Netherlands, Cyprus and the UK proposed an algorithm to improve diagnosis of trisomy 21, or Down syndrome. Their results were published in the most recent ‘Ultrasound in Obstetrics and Gynecology’.
The team used an artificial neural network (ANN) approach; a type of machine learning. ANN is based on software that is fed information from past events. This information, or training data, is processed through an algorithm which allows the ANN to draw conclusions about this information. The more information the system is fed, the more conclusions it can draw. However, quantity isn’t everything – the system can provide conclusions that aren’t obvious to a human, since they involve complex relationships between variables.
Using the system to stratify risk of Down syndrome incidence meant the researchers could decide whether or not to carry out invasive diagnostic procedures with potential adverse effects. Sampling of chorionic villus, or placental fluid, is one of the most reliable methods to detect chromosomal abnormalities, but it carries serious risk of infection or even miscarriage. Less or non-invasive tests are available, such as estimation of maternal age, maternal blood tests or an ultrasound to detect the presence or absence of fetal nasal bone. However, these approaches have a substantial false-positive rate and could lead to the abortion of a healthy foetus. A non-invasive cell-free DNA (cfDNA) test is a recent breakthrough, but it is expensive and difficult to apply to an entire population.
The team trained an ANN using results from both invasive and non-invasive tests from around 5,000 pregnancies. The risks were stratified into two stages. In the first, the cases were marked as “no-risk” (no further action) and “risk” (taken to the stage 2). In the second stage, the process was repeated, allowing for a more nuanced stratification. For “no-risk”, again, no further action was required, and for “moderate-risk” a cfDNA test is required. Only those who are in the “high-risk” cohort would be recommended to undergo placental fluid sampling. The ANN training was validated by around 36,000 pregnancies.
After the first stage, around a half of the pregnancies were taken to the second stage testing, where additional markers were included. There, the majority were classed as “moderate-risk” (~10,000 pregnancies), slightly less were “no-risk” (~7,000 pregnancies) and the rest (less than 200 pregnancies) were “high-risk”. One of the ANN’s biggest achievements is the reduction of false-positives in transition between the stages. The proportion of pregnancies requiring the invasive procedure was less than 1% of the starting population. The method could also accurately predict other chromosomal abnormalities, such as Turner syndrome or triploidy (three sets of chromosomes rather than two). As the authors stated, the main weakness of the approach is the classification of the “moderate-risk” group, which can still be improved, with the numbers potentially further reduced.
While still limited, this new approach has improved the accuracy and cost-effectiveness of testing during pregnancy. The team are focusing their future work on other pregnancy complications associated with chromosomal abnormalities, such as preeclampsia.
Further information can be found here http://onlinelibrary.wiley.com/doi/10.1002/uog.17558/epdf
This article was written by Alina Gukova and edited by Sam Stanfield.
Kindly share the updated research on AI and DS.