AI Shows Promise for Newborn Lung Care, But Hospitals Aren’t Ready Yet

Artificial intelligence is quietly creeping into one of medicine’s most delicate environments: the neonatal intensive care unit, where the smallest patients fight for every breath. A sweeping new analysis published in the Journal of Perinatology has mapped, for the first time in a structured way, exactly where digital technologies and machine learning are being deployed to assess newborn respiratory health — and where they still fall short. The scoping review, led by Simone Nascimento Santos Ribeiro of Faculdade Ciências Médicas de Minas Gerais in Brazil, examined 35 studies published between 2019 and 2026 and concluded that while the technology is advancing at remarkable speed, it is not yet ready for widespread clinical implementation.

The stakes could hardly be higher. Neonatal respiratory disorders remain among the leading causes of illness and death in the first weeks of life, contributing significantly to the roughly 2.3 million neonatal deaths reported worldwide each year. Preterm infants are especially vulnerable because their lungs are structurally and functionally immature, leaving them exposed to respiratory distress syndrome, bronchopulmonary dysplasia, and prolonged dependence on ventilators. Even in wealthy countries with state-of-the-art intensive care, these conditions carry short- and long-term consequences, including chronic pulmonary disease and impaired neurodevelopment.

Traditional respiratory assessment in newborns relies on clinical examination, blood gas analysis, and conventional imaging — tools the review describes as fundamentally limited. Different clinicians can interpret the same chest X-ray or lung sounds differently, a problem known as interobserver variability. Conventional methods also have limited ability to predict how a baby’s condition will evolve, and they struggle to integrate the complex, constantly changing streams of data generated in an intensive care unit. In a setting where minutes can matter, those limitations become critical.

To map the emerging alternatives, the Brazilian team followed the PRISMA-ScR reporting guidelines and registered their protocol on the Open Science Framework. They searched PubMed/MEDLINE, Embase, Scopus, Web of Science, and IEEE Xplore for studies published between January 2015 and March 2026, combining controlled descriptors with free-text terms covering neonates, respiratory assessment, digital health, and artificial intelligence. From 199 initially identified records, independent reviewers whittled the pool down through duplicate removal, title and abstract screening, and full-text review, ultimately including 35 studies — just 18.5 percent of what they started with. Studies lacking a neonatal population, respiratory outcomes, or actual AI components were excluded, as were those with fewer than ten patients.

The technological landscape that emerged is dominated by machine learning. Twenty-one of the 35 studies — 60 percent — used machine-learning approaches, most commonly XGBoost, support vector machines, random forests, and ensemble models. Deep learning appeared in ten studies, or 28.6 percent, typically built on convolutional neural networks, ResNet architectures, BiLSTM networks, and U-Net for analyzing images and physiological signals. Three studies used hybrid machine learning and deep learning models, and one deployed an intelligent closed-loop control system. Electronic health records were the most common data source at 34.3 percent of studies, followed by physiological signals, chest radiography, digital auscultation, ventilator-derived data, and a scattering of other technologies including lung ultrasound, spectroscopy, and video imaging.

The review identified three major application domains. The largest was predictive modeling: algorithms trained largely on routinely collected clinical data to forecast outcomes such as bronchopulmonary dysplasia, respiratory distress syndrome, the need for mechanical ventilation, and extubation failure. Many of these models reported high predictive performance, and the authors suggest they are the closest to genuine clinical implementation. Yet the field is fragmented — datasets, outcome definitions, and methodologies vary so widely from study to study that comparing models head-to-head is nearly impossible, and most evidence comes from retrospective designs without robust external validation.

The second domain is medical imaging. Roughly a quarter to a third of the studies applied deep learning to chest radiographs and lung ultrasound, enabling automated classification of neonatal respiratory diseases. These tools showed particular promise for reducing interobserver variability and standardizing diagnosis — a meaningful advantage in hospitals with limited access to specialist radiologists. Some studies went further, assessing agreement between human readers and AI models interpreting lung ultrasound scans. But imaging-based AI carries its own dependencies: it requires standardized image acquisition, high-quality annotations, and multicenter validation to perform reliably across different populations and equipment, none of which is yet guaranteed.

The third domain, continuous physiological monitoring, may be the most futuristic. Researchers are developing noninvasive systems using sensors, video cameras, radar, wearable devices, and photoplethysmography to track oxygen saturation, respiratory rate, and ventilatory patterns in real time — some without touching the baby at all. Other creative approaches include separating heart and lung sounds from noisy recordings, detecting apnea from thermal video, analyzing newborn cry acoustics to distinguish sepsis from respiratory distress, and even using mid-infrared spectroscopy of gastric aspirates to predict the need for prolonged respiratory support. These systems demonstrated feasibility for early detection of clinical deterioration, but they are highly susceptible to motion artifacts, environmental interference, and variations in how a baby is positioned — and they remain at an early stage of clinical validation.

Perhaps the most significant trend the authors identified is the move toward multimodal AI, which fuses clinical, physiological, and imaging data into more comprehensive predictive models. Such integration could give clinicians a far richer picture of an infant’s respiratory trajectory than any single data stream. But it also raises the bar considerably, demanding data standardization, interoperability between hospital systems, and substantial technological infrastructure — resources that are unevenly distributed across healthcare settings worldwide.

The review’s bottom line is a sobering counterweight to the hype. Beyond methodological heterogeneity and small samples, the authors flag poorly explored questions of algorithmic bias, equity in model performance, and clinical safety, along with the low interpretability of many deep learning models — a serious barrier in environments requiring rapid, transparent decisions. Their prescription is clear: the field must move beyond proof-of-concept studies toward prospective, multicenter validation trials, standardized protocols, transparent model reporting, and careful integration into clinical workflows, with attention to ethical and regulatory dimensions. AI may eventually become an indispensable assistant in the neonatal unit, but for now, the babies’ breathing is still being judged by human ears, eyes, and hands.

Subject of Research: Artificial intelligence and digital technologies for neonatal respiratory assessment

Article Title: Artificial intelligence and digital technologies transforming neonatal respiratory assessment: a scoping review

Article References: Ribeiro, S. N. S., Vargas, M. E. R. R., Velame, R. D. C., Fernandes, A. E. R., Filho, M. A. T., Leão, R. N., Maia, H., Pereira, S. A., & Rodrigues-Machado, M. D. G. (2026). Artificial intelligence and digital technologies transforming neonatal respiratory assessment: a scoping review. Journal of Perinatology. https://doi.org/10.1038/s41372-026-02894-5

Image Credits: AI Generated

DOI: 10.1038/s41372-026-02894-5

Keywords: artificial intelligence, machine learning, deep learning, neonatology, respiratory assessment, bronchopulmonary dysplasia, lung ultrasound, chest radiography, continuous monitoring, neonatal intensive care, predictive models, scoping review