AI Predicts Lung Disease in Premature Babies
Identifying bronchopulmonary dysplasia (BPD) in premature babies remains a challenge. Lung function tests usually require blowing out on request, which is a task babies cannot perform. Current techniques require sophisticated equipment to measure an infant’s lung ventilation characteristics, so doctors usually diagnose BPD by the presence of its leading causes, prematurity and the need for respiratory support.
Researchers at the University of Basel in Switzerland trained an ANN model to predict BPD in premature babies.
The team studied a group of 139 full-term and 190 premature infants who had been assessed for BPD, recording their breathing for 10 minutes while they slept. For each baby, 100 consecutive regular breaths, carefully inspected to exclude sighs or other artifacts, were used to train, validate, and test an ANN called a Long Short-Term Memory model (LSTM), which is particularly effective at classifying sequential data such as tidal breathing.
Researchers used 60% of the data to teach the network how to recognize BPD, 20% to validate the model, and then fed the remaining 20% of the data to the model to see if it could correctly identify those babies with BPD.
The LSTM model classified a series of flow values in the unseen test data set as belonging to a patient diagnosed with BPD or not with 96% accuracy.
“Until recently, this need for large amounts of data has hindered efforts to create accurate models for lung disease in infants because it is so difficult to assess their lung function,” Delgado-Eckert said. “Our research delivers, for the first time, a comprehensive way of analyzing infants’ breathing and allows us to detect which babies have BPD as early as 1 month of corrected age.”
The study presented by Delgado-Eckert received funding from the Swiss National Science Foundation. Narayanan and Pinnock reported no relevant financial relationships.
A version of this article appeared on Medscape.com.