Artificial neural network has ‘high performance’ in detecting bronchopulmonary dysplasia
Click Here to Manage Email Alerts
Key takeaways:
- Researchers trained an artificial neural network with air flow data during tidal breathing to identify bronchopulmonary dysplasia.
- This model had high accuracy, specificity, sensitivity and precision.
An artificial neural network demonstrated high accuracy in detecting bronchopulmonary dysplasia in infants using air flow data, according to a presentation at the European Respiratory Society International Congress.
“Bronchopulmonary dysplasia (BPD) is one of those few diseases that are diagnosed based on the presence of one of its main causes (ie, external respiratory assistance over a certain period of time),” Edgar Delgado-Eckert, PhD, adjunct professor in the department of biomedical engineering at the University of Basel and research group leader at the University Children’s Hospital in Switzerland, told Healio. “Our models could help clinicians detect BPD just a few weeks after birth, that is, before the external respiratory assistance is fully administered.
“Early detection would thus enable a more informed decision regarding the mode of respiratory support to be administered to the infant during the first weeks of life,” Delgado-Eckert said. “Furthermore, through early detection of BPD, a long-term prognosis may become available earlier, allowing for the early planning of follow-up assessments and potential interventions.”
In this study, Delgado-Eckert and colleagues trained, validated and tested an artificial neural network (ANN) — specifically a Long Short-Term Memory model — using air flow/breathing data from 190 preterm infants and 139 term-born infants taken at an average of 44.7 weeks postmenstrual age to see if the model could correctly identify BPD.
Researchers collected 10 minutes of inspiratory and expiratory air flow data for each infant while they were sleeping. These recordings were then looked through to capture instances of 100 consecutive regular breaths, or tidal breathing, for each infant to use in this study.
The training cohort involved 60% of the data, the validation cohort involved 20% and the testing cohort was made up of the last 20% of the data.
Most of the preterm infants in this study had either mild (n = 47), moderate (n = 54) or severe (n = 31) BPD, leaving only 58 without the disease.
Based on tidal breathing data from the testing cohort, researchers found that the trained model had 96% accuracy in detecting infants with and without BPD.
“We were surprised by the high performance of the trained ANN model at classifying unseen time series of air flow data measured during tidal breathing as originating from a BPD patient or from a non-BPD patient,” Delgado-Eckert told Healio.
Several other performance measures achieved high values in the testing cohort, including specificity (100%), sensitivity (96%), precision (98%) and F1-score (97%).
“In this study, we used data from only one hospital; future studies need to investigate the performance of the model on unseen data from a different hospital,” Delgado-Eckert told Healio. “This would help us understand how general are the insights the model has learned from the data. Also, future studies combining data from different health care centers may allow us to train even better models.”
Future studies may also use a different ANN model type or obtain tidal breathing data at a different time, Delgado-Eckert said.
“In the present study, the data were obtained at about a month of corrected age,” Delgado-Eckert told Healio. “In future studies, it would be interesting to try to obtain tidal breathing flow data from preterm infants just a couple of weeks after birth. This would enable us to generate models capable of making very early predictions of BPD.”
Reference:
- Artificial intelligence can be trained to detect lung disease in premature babies. https://www.ersnet.org/news-and-features/news/artificial-intelligence-can-be-trained-to-detect-lung-disease-in-premature-babies/. Published Sept. 10, 2024. Accessed Sept. 10, 2024.