Using Machine Learning to Predict the Need for Postnatal Cerebrospinal Fluid Diversion in Fetal Ventriculomegaly

About the study

Neurosurgeons on our team recently performed a retrospective case-control study of 50 patients with cerebral ventriculomegaly (VM) diagnosed based on fetal MRI findings and managed in the postnatal period at our hospital between 2008 and 2014 to determine if analysis of fetal MRI and machine learning techniques could be used to predict the need for postnatal cerebrospinal fluid (CSF) diversion in patients with VM.

Multiple imaging features were extracted from fetal MRI and input into a machine learning algorithm that correctly classified postnatal CSF diversion status with 82% accuracy. In an independent replication cohort study, the model achieved 91% accuracy.

About the findings

These findings suggest that image analysis and machine learning techniques can be applied to fetal MRI to predict the need for postnatal CSF diversion in patients with VM with high accuracy and generalizability.

A significant strength of the computational approach is the ability to simultaneously assess multiple imaging features, most of which are not appreciable by visual inspection alone. Rather than examining a single imaging feature at a time, several types of measurements were incorporated into the model, including linear measures, area, volume and morphologic features. Integration of patterns of ventricle size, shape and configuration more fully characterized VM and the development of postnatal hydrocephalus (HC) compared with any single imaging feature.

Figure 1: Multiple imaging features were collected from the fetal MRI and input into a machine learning algorithm to predict which patients would need future surgery after birth.

Application in clinical setting and beyond

The image-based predictive model has several clinical applications. Counseling in the prenatal setting is challenging in VM, and the model can assist clinicians in providing families with prognostic information related to the anticipated need for CSF diversion in the postnatal period. After birth, patients diagnosed in utero with VM are monitored closely in the NICU for HC, including signs of increased intracranial pressure, such as lethargy, vomiting, apnea, upgaze paresis, bulging fontanelle, increasing head circumference and splayed cranial sutures. Patients deemed to have high risk by the model would warrant closer and possibly an extended duration of monitoring before leaving the hospital, whereas the inpatient observation period may be shortened in low-risk patients.

Figure 2: These pictures show the spread of the top imaging features for each patient, where each vertical bar represents a patient and each horizontal bar represents an imaging feature.

We envision this model to be used as a guide or supplementary clinical tool in addition to clinical judgment and monitoring of the severity and rate of increase of VM in the postnatal period. Furthermore, as fetal surgery techniques continue to advance, a noninvasive predictive model may aid in patient selection for future clinical trials to assess in utero CSF diversion. In addition to clinical benefit, the application of image analysis and machine learning techniques to create the model yielded novel imaging features associated with postnatal CSF diversion, which in turn provide further insight into the pathophysiologic features of VM.

Figure 3: These graphs show the performance of the model in samples of patients from two different institutions. For comparison purposes, the diagonal line shows model performance equivalent to flipping a coin to determine the outcome for each patient.

Work is currently underway to incorporate the fetal MRI-based predictive model into user-friendly software, and future studies will investigate the integration of the software into clinical workflow and prospectively assess model accuracy.

Reference

Pisapia JM, Akbari H, Rozycki M, Goldstein H, Bakas S, Rathore S, Moldenhauer JS, Storm PB, Zarnow DM, Anderson RCE, Heuer GG, Davatzikos C. Use of fetal magnetic resonance imaging analysis and machine learning to predict the need for postnatal cerebrospinal fluid diversion in fetal ventriculomegaly. JAMA Pediatr. 2018;172(2):128-135.