Fetal MRI Analysis, Machine Learning Identifies Candidates for CSF Diversion
The model will assist in the selection of candidates for potential fetal surgical intervention.
The use of image analysis and machine learning of fetal magnetic resonance imaging (MRI) features can predict the need for postnatal cerebrospinal fluid (CSF) diversion among patients with fetal ventriculomegaly with an accuracy of at least 82%, according to results of a recent retrospective, case-control study published in JAMA Pediatrics.
An institutional database of 253 patients with fetal ventriculomegaly was used to generate a predictive model. Data analysis took place from January 1, 2008, through December 31, 2015. A total of 50 patients with fetal ventriculomegaly were evaluated — 25 of whom required postnatal CSF diversion and who were matched, based on gestational age, with 25 patients who had fetal ventriculomegaly but did not require CSF diversion (the discovery cohort). The model was applied to a sample of 24 consecutive patients with fetal ventriculomegaly who were evaluated at a different institution (the replication cohort) between January 1, 1998, and December 31, 2007. Data analysis occurred from January 1, 1998, through December 31, 2009.
The main study outcomes included accuracy, sensitivity, and specificity of the model to classify patients who were in need of postnatal CSF diversion correctly. Multiple imaging characteristics were obtained from fetal MRI analysis and integrated by machine learning to yield a model with the ability to classify postnatal CSF diversion status with a high level of accuracy.
A total of 74 patients (41 girls and 33 boys; mean gestational age, 27.0±5.6 months) from both cohorts were included in the analysis. Median time to CSF diversion was 6 days in the discovery cohort. Patients with fetal ventriculomegaly in whom no symptoms developed were followed for a median of 29 months. The discovery cohort model classified patients who required CSF diversion correctly with 82% accuracy (95% CI, 0.64-0.96), 80% sensitivity (95% CI, 0.64-0.96), and 84% specificity (95% CI, 0.70-0.98).
In the replication cohort, on the other hand, the model attained 91% accuracy (95% CI, 0.77-1.07), 75% sensitivity (95% CI, not applicable), and 95% specificity (95% CI, 0.85-1.04).
The investigators concluded that an MRI-based predictive model with high levels of accuracy and generalizability may provide clinicians with prenatal prognostic information and help guide postnatal management of infants with fetal ventriculomegaly, and may also be helpful in the selection of candidates for potential fetal surgical intervention.
Pisapia JM, Akbari H, Rozycki M, et al. Use of fetal magnetic resonance image analysis and machine learning to predict the need for postnatal cerebrospinal fluid diversion in fetal ventriculomegaly [published online December 18, 2017]. JAMA Pediatr. doi: 10.1001/jamapediatrics.2017.3993