跳到正文
原文
Nature 健康护理主题·· 2026-10-02AI 评分40

人工智能与数字技术变革新生儿呼吸评估:范围综述

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

AI 导读

本范围综述系统梳理了2015至2026年间应用于新生儿呼吸评估的人工智能与数字技术。研究涵盖预测模型、自动肺图像分析及持续无创监测系统,旨在综合其临床应用、评估参数及现有知识差距,为未来研究及将这些技术安全有效地融入新生儿临床实践提供指导。

正文

Introduction

Neonatal respiratory disorders remain among the leading causes of morbidity and mortality during the neonatal period worldwide, contributing significantly to the approximately 2.3 million neonatal deaths reported annually across the globe [1,2,3]. Preterm newborns are particularly vulnerable due to the structural and functional immaturity of the respiratory system, presenting an increased risk of developing conditions such as respiratory distress syndrome, bronchopulmonary dysplasia, and prolonged need for ventilatory support. Even in high-income settings, these conditions are associated with short- and long-term complications, including chronic pulmonary morbidity and neurodevelopmental impairment.

Neonatal respiratory assessment therefore plays a central role in clinical decision-making. However, traditionally used methods, based on clinical examination, blood gas analysis, and conventional imaging tests, present important limitations, including interobserver variability, limited predictive capacity, and difficulty integrating complex and dynamic clinical data [4]. These limitations become even more critical in neonatal intensive care units, where rapid and accurate decisions are essential.

In this context, advances in digital technologies and artificial intelligence (AI) have been identified as an opportunity to transform neonatal respiratory assessment through the analysis of large volumes of multimodal data [5,6,7]. Machine learning and deep learning algorithms enable the identification of nonlinear patterns in physiological signals, medical images, and electronic health records, expanding the potential for prediction and clinical decision support.

Recent studies have demonstrated promising applications of these tools, including predictive models [8,9,10,11,12,13,14,15,16,17], automated lung image analysis [18,19,20,21,22], and continuous noninvasive monitoring systems [23,24,25]. However, the literature still presents significant heterogeneity regarding the methods used, evaluated parameters, and clinical application settings, highlighting important gaps for their safe and effective implementation.

Given this scenario, this scoping review aimed to map digital technologies and artificial intelligence-based approaches used in neonatal respiratory system assessment, as well as to synthesize their clinical applications, evaluated parameters, and knowledge gaps, thereby contributing to guiding future research and the critical incorporation of these technologies into neonatal practice.

Materials and methods

Study design

This scoping review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR) guidelines [26]. The review protocol was previously registered in the Open Science Framework (OSF): 10.17605/OSF.IO/H6RKC.

Search strategy

The search strategy was developed using a combination of controlled descriptors (MeSH/Emtree) and free-text terms and was refined through discussions among members of the research team. The following electronic databases were systematically searched: PubMed/MEDLINE, Embase, Scopus, Web of Science, and IEEE Xplore. The search included studies published between January 2015 and March 2026. Search terms included “neonate” OR “newborn,” “respiratory assessment,” “mobile application” OR “digital health,” “artificial intelligence” OR “machine learning,” and “thoracoabdominal mobility,” combined using Boolean operators (AND/OR). The complete search strategy for each database was documented and is available in the supplementary material. All retrieved records were exported and organized in a reference management software, followed by duplicate removal. The remaining studies were imported into a screening platform (Rayyan), where titles and abstracts were independently assessed for eligibility by two reviewers. Full-text review of potentially relevant articles was also independently conducted by two reviewers. Disagreements at both stages were resolved through discussion, with the participation of a third reviewer when necessary, ensuring methodological rigor and minimizing selection bias. Data extraction was performed using a previously defined standardized form, including the following variables: author and year of publication, country, study type, population, technology used, type of artificial intelligence approach, evaluated respiratory parameters, clinical objectives, application setting, main findings, and limitations.

Eligibility criteria

Studies involving neonatal populations (term and preterm newborns), the use of digital technologies and/or artificial intelligence, and the assessment of respiratory parameters were included. Studies focused on non-respiratory outcomes, without artificial intelligence or digital technology, editorials or non-original studies, and studies with a small sample size ( < 10 patients) were excluded.

Selection and primary evaluation of studies

  • The selection and evaluation of the studies occurred in four stages:

  • Stage 1: Search in the PubMed, Cochrane Library, LILACS, Web of Science, SciELO, and Science Direct databases using the main strategy and its variations.

  • Stage 2: Exclusion of duplicates - All references obtained from the database search were imported into the reference management Rayyan systematic review software, where duplicate studies were initially excluded.

  • Stage 3: Title and Abstract Screening - Two independent reviewers (M.E.R.R.V and R.C.V) conducted this step. In cases of disagreement, a third independent reviewer (S.N.S.R) reviewed the article and then confirmed or rejected the inclusion of the specific publication. All reviewers conducted the screening according to the eligibility criteria for article selection.

  • Stage 4: full-text reading and data extraction - Full texts were independently assessed by two reviewers. For studies meeting the inclusion criteria, data extraction was independently performed by two reviewers using a standardized extraction form. The collected information included bibliographic data (title, author, and publication year), study location, study type, objective, evaluated techniques, outcomes, and conclusions. Any disagreements during study selection or data extraction were resolved through discussion with a third reviewer (S.N.S.R).

Data synthesis

The characteristics of the interventions, as well as categorical variables, were synthesized descriptively and presented in tables. Numerical variable results extracted from the articles were presented as means and standard deviations or medians and interquartile ranges.

Results

A total of 199 studies were identified across the selected databases. After the removal of 10 duplicates, 189 records remained for the screening stage.

During title and abstract screening, 63 studies (33.3%) were excluded for not meeting eligibility criteria, including publication year criteria (n = 15; 23.8%), study type (n = 22; 34.9%), and title irrelevance (n = 26; 41.3%). Consequently, 126 articles (66.7%) were selected for full-text review.

Following the eligibility stage, 91 studies (72.2%) were excluded for absence of a neonatal population, application of artificial intelligence to non-respiratory outcomes, absence of respiratory parameter assessment, lack of artificial intelligence use, and small sample size ( < 10 patients). At the end of the selection process, 35 studies (18.5% of those initially identified) published between 2019 and 2026 were included in the scoping review [1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23, 25, 27,28,29,30,31,32,33,34,35,36,37].

The complete study selection process is presented in the PRISMA-ScR flowchart (Fig. 1).

Fig. 1

PRISMA flowchart. Summary of the screening process.

The 35 included studies comprised predominantly prospective observational and retrospective cohort designs, with fewer diagnostic and crossover studies (Table 1). Machine-learning approaches were used in 21 of the 35 studies (60.0%), deep-learning approaches in 10 studies (28.6%), hybrid machine learning/deep learning models in 3 studies (8.6%), and one study (2.9%) employed an intelligent closed-loop control system. Electronic health records represented the most common data source (34.3%), followed by physiological signals (20.0%), chest radiography (8.6%), digital auscultation (8.6%), ventilator-derived data (8.6%), and other technologies, including lung ultrasound, spectroscopy, video imaging, prenatal ultrasound, and clinical decision-support systems (Table 1).

Table 1 Characteristics of included studies on digital technologies and artificial intelligence in neonatal respiratory assessment.

Full size table

Main findings

Analysis of the 35 included studies revealed three major application domains of digital technologies and artificial intelligence in neonatal respiratory assessment: predictive models, imaging technologies, and continuous physiological monitoring.1. Reference numbering throughout the manuscript follows the complete reference list and therefore does not correspond to the sequential number of included studies.

As summarized in Table 1, the included studies were predominantly conducted in neonatal intensive care units and comprised prospective observational, retrospective cohort, and diagnostic designs. Machine learning was the most frequently used artificial intelligence approach, particularly XGBoost, support vector machines, random forests, and ensemble models, whereas deep learning was mainly applied to imaging and physiological signal analysis using convolutional neural networks, ResNet architectures, BiLSTM, and U-Net. Electronic health records represented the most common data source, followed by imaging modalities, physiological signals, ventilator-derived data, and digital auscultation. Prediction of respiratory outcomes was the predominant clinical application, followed by diagnosis and continuous respiratory monitoring.

Artificial intelligence-based predictive models

Predictive models represented the largest application domain identified in this review, encompassing studies that used predominantly machine-learning algorithms and, to a lesser extent, deep-learning approaches to predict respiratory outcomes such as bronchopulmonary dysplasia, respiratory distress syndrome, need for mechanical ventilation, and extubation failure [1,2,3, 5, 7, 9, 12, 15, 17, 18, 25, 27, 28, 31, 33].

A predominance of techniques such as deep neural networks, random forests, and support vector machines was observed, with high predictive performance reported in most studies. However, significant heterogeneity was identified in the datasets used, variability in analyzed outcomes, and limited methodological standardization, making direct comparison among models difficult [1,2,3, 5, 7, 9, 12, 15, 17, 18, 25, 27, 28, 31, 33].

Artificial intelligence-supported imaging technologies

Approximately 25–30% of the studies used deep learning techniques applied to imaging examinations, mainly chest radiography and lung ultrasound. These approaches enabled automated classification of neonatal respiratory diseases, with potential reduction in interobserver variability and increased diagnostic accuracy [4, 8, 10, 11, 32].

Comparatively, imaging-based studies demonstrated greater standardization of input data; however, they still face limitations related to external validation and model generalizability across different clinical settings and populations [4, 8, 10, 11, 32].3.

Physiological monitoring and noninvasive systems

Approximately 25–30% of the studies focused on the development of technologies for continuous monitoring of respiratory parameters using noninvasive systems. Sensors, video-based systems, radar, wearable devices, and photoplethysmography were used to assess parameters such as oxygen saturation, respiratory rate, and ventilatory patterns [6, 13, 14, 16, 19,20,21,22,23, 29, 30, 34,35,36,37].

These approaches demonstrated feasibility for real-time monitoring and potential for early detection of clinical deterioration. However, compared with predictive models, they presented greater technical variability and stronger dependence on environmental conditions and signal quality [6, 13, 14, 16,19,20,21,22,23, 29, 30, 34,35,36,37].

Comparative synthesis of findings

Overall, the findings demonstrated that:

  • Predictive models were the most frequently used approaches, with greater focus on clinical outcomes and decision support.

  • Imaging technologies showed greater potential for diagnostic standardization, although with a smaller volume of studies.

  • Continuous monitoring systems stand out for their real-time applicability, although they remain at an early stage of clinical validation. In addition, a growing trend toward multimodal integration of different data types (clinical, physiological, and imaging) was identified, although still limited by interoperability and standardization challenges.

Table 1 summarizes the characteristics of the included studies, encompassing methodological aspects, employed technologies, artificial intelligence approaches, and evaluated respiratory parameters.

Figure 2 provides a conceptual overview of the principal application domains identified in this scoping review, illustrating the relationship between digital technologies, artificial intelligence approaches, and neonatal respiratory assessment [1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23, 25, 27,28,29,30,31,32,33,34,35,36,37].

Fig. 2

Synthesis of findings (technology + AI + parameters).

Discussion

This scoping review demonstrates that digital technologies and artificial intelligence are being progressively incorporated into neonatal respiratory system assessment, representing a transition from essentially reactive models toward predictive, automated, and data-driven approaches. The findings indicate that these technologies are predominantly applied across three domains: prediction of respiratory outcomes, medical image analysis, and continuous physiological monitoring [1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23, 25, 27,28,29,30,31,32,33,34,35,36,37].

The distribution of the included studies suggests that AI development in neonatal respiratory care is currently driven primarily by predictive analytics based on structured clinical data rather than autonomous diagnostic systems. Machine-learning models were predominantly applied to routinely collected electronic health records for prediction of respiratory outcomes, whereas deep-learning approaches were mainly used for image interpretation and physiological signal analysis. These findings highlight that AI applications in neonatal respiratory care are at different stages of technological maturity, with predictive models appearing to be closer to clinical implementation than imaging- and monitoring-based approaches.

Artificial intelligence-based predictive models constituted the main application domain identified, being widely used to estimate the risk of bronchopulmonary dysplasia, the need for mechanical ventilation, and extubation failure [8,9,10,11,12,13,14,15,16,17,18,19]. These models, which frequently incorporated random forests, support vector machines, gradient boosting methods, and deep neural networks, demonstrated high predictive performance.

However, their clinical applicability remains limited by the predominance of retrospective studies, lack of robust external validation, and heterogeneity in the analyzed outcomes. Furthermore, the low interpretability of many deep learning models represents an important barrier to their adoption in clinical practice, particularly in settings requiring rapid and transparent decision-making.

In the medical imaging domain, the application of deep learning algorithms to chest radiography and lung ultrasound demonstrated relevant potential for automating the classification of neonatal respiratory diseases and reducing interobserver variability [14,15,16,17, 30,31,32]. These tools may contribute to greater diagnostic standardization, particularly in settings with limited availability of specialists. Unlike prediction models, imaging-based AI depends on standardized image acquisition, high-quality annotations, and multicenter validation to ensure robust performance across different clinical settings, factors that remain important barriers to widespread implementation [38,39,40].

Continuous physiological monitoring using digital technologies emerged as a promising field, with the use of noninvasive sensors, video-based systems, radar, and respiratory signal analysis [6, 13, 14, 16, 19,20,21,22,23, 29, 30, 34,35,36,37]. These approaches enable real-time assessment of parameters such as respiratory rate, oxygen saturation, and ventilatory patterns, offering potential for early detection of clinical instability, in line with recent advances in digital monitoring and artificial intelligence applied to healthcare [38,39,40,41,42]. However, their implementation in routine clinical practice remains challenging because respiratory measurements are highly susceptible to motion artifacts, environmental interference, variability in neonatal positioning, integration with hospital systems, and the need for validation in real-world clinical settings.

Another important finding of this review was the increasing adoption of multimodal artificial intelligence approaches integrating clinical, physiological, and imaging data to develop more comprehensive predictive models, as observed in the included studies [8,9,10,11,12,13,14,15,16,17,18,19,20,21,22, 29,30,31]. Although promising, these approaches strongly depend on data standardization, interoperability between systems, and the availability of technological infrastructure, factors widely discussed in the digital health literature [38,39,40], which may limit their applicability across different healthcare settings.

Overall, the included studies presented significant methodological heterogeneity, with a predominance of retrospective designs and limited sample sizes [1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23, 25, 27,28,29,30,31,32,33,34,35,36,37], restricting the generalizability of the findings. Considerable variability was also observed in the AI methodologies adopted, with studies evaluating similar respiratory outcomes using different algorithms, predictors, preprocessing techniques, and validation strategies, limiting direct comparison between models and thereby reducing the comparability of findings across studies. Furthermore, fundamental aspects such as algorithmic bias, equity in model performance, and clinical safety remain poorly explored, as highlighted in recent studies on artificial intelligence in healthcare [24, 38,39,40,41,42], representing critical gaps for the safe implementation of these technologies.

From a clinical perspective, the results of this review suggest that artificial intelligence may serve as a complementary tool in neonatal respiratory assessment, contributing to clinical decision-making and optimization of care [1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23, 25, 27,28,29,30,31,32,33,34,35,36,37]. However, effective incorporation into clinical practice depends on robust external validation, development of standardized protocols, and integration with existing clinical workflows, as recommended in methodological and implementation literature [26, 38,39,40]. Although the search strategy was designed around the central concepts of artificial intelligence, digital technologies, and neonatal respiratory assessment, additional search terms specifically related to prediction models (e.g., “prediction model” or “logistic regression”) might have identified additional relevant studies. This should be considered when interpreting the comprehensiveness of the predictive modeling literature included in this review.

Therefore, future studies should prioritize prospective designs, multicenter validation, and assessment of the clinical impact of these technologies, in addition to considering ethical, regulatory, and safety aspects [24, 38,39,40,41,42], which are essential to ensure responsible and sustainable implementation in neonatal care.

Collectively, these findings suggest that the next stage of AI development in neonatal respiratory care should move beyond proof-of-concept studies toward prospective multicenter validation, transparent model reporting, and integration into routine clinical workflows. Only through rigorous validation and implementation studies will these technologies be able to demonstrate meaningful clinical benefit beyond technical performance.

Conclusion

Digital technologies and artificial intelligence represent promising tools for neonatal respiratory assessment, with the potential to improve outcome prediction, diagnostic accuracy, and continuous monitoring. However, clinical application remains limited by the lack of robust external validation, methodological heterogeneity, and challenges related to model interpretability. Future prospective, multicenter, and standardized studies are needed to evaluate the effectiveness, safety, and clinical impact of these technologies. The development and implementation of standardized clinical protocols will be essential to safely and effectively integrate these technologies into neonatal practice.

来源:Nature 健康护理主题 · nature.com