Neonatal intensive care units (NICUs) are specialized hospital units providing intensive medical care for critically ill newborns, particularly those born prematurely or with medical and surgical conditions, where close monitoring and specialized interventions and treatment are needed. One in 10 babies is born prematurely or sick, emphasizing the essential role of neonatal care in promoting their well-being and survival [1]. With an estimated 15 million premature births annually and around 1 million child deaths due to preterm complications each year [2], the significance of specialized care is evident. Surviving infants may encounter lifelong challenges, highlighting the need for effective neonatal care to ensure healthy development and improved long-term outcomes.

Within the complex ecosystem of the NICU, health care providers rely on the expertise of skilled clinicians and essential medical devices to deliver specialized care. The implementation of electronic medical records has made vast volumes of data more accessible in recent years, contributing to significant advances in medical research and technology, leading to improved survival rates for NICU patients [3]. However, these advancements come with increased demands on health care resources, including specialized staff, equipment, and facilities. Innovative approaches are essential to address these challenges and enhance overall health care efficacy.

One avenue for improvement is the prediction of the length of stay in the NICU. Accurate length of stay predictions are crucial for resource allocation, discharge planning, and optimizing care pathways [4]. Prolonged stays can have significant implications on neonatal development, parent-newborn interactions, and family well-being [5,6]. Additionally, predicting clinical outcomes is equally important, as it allows for monitoring essential developmental milestones, potential delays or improvements, and the lasting effects of medical interventions on NICU patients [7-12].

Health care providers face several challenges when predicting length of stay or clinical outcomes in the NICU setting. The vast amounts of data, including patient vitals, laboratory results, imaging data, and clinical notes, can be overwhelming to process manually [13]. Limited availability of human resources in high-stress NICU settings further compounds the challenge, where teams must balance heavy workloads and rapid decision-making [14].

Furthermore, the dynamic nature of neonatal care, marked by rapid changes in health status, adds another layer of complexity to accurate predictions. In this context, artificial intelligence (AI) technology emerges as a potential solution. AI refers to the capability of computer systems to carry out tasks that typically require human intelligence, including prediction, learning, and decision-making [15]. While traditional software and other computer systems rely on predetermined rules and instructions to accomplish specific tasks, AI sets itself apart by its capacity to learn from data and continually enhance performance without explicit programming. This capability is primarily achieved through a branch of AI called machine learning [16].

Machine learning encompasses a diverse array of techniques, including supervised learning, where models are trained on labeled datasets (eg, historical patient outcomes); unsupervised learning, which aims to discover hidden structures and patterns within unlabeled data; and reinforcement learning, a paradigm focused on optimizing decision-making through iterative interactions with an environment and the receipt of rewards. Deep learning, a specialized form of machine learning, uses multilayered neural networks that mimic the human brain’s processing abilities to uncover complex relationships in large datasets. In the context of this study, these techniques are especially valuable for predicting neonatal outcomes, such as length of stay or health conditions, using data from electronic medical records. By leveraging machine learning and deep learning, AI has the potential to transform decision-making in NICUs, providing more accurate, timely, and personalized insights to support neonatal care [12].

In health care, predictive analytics powered by AI has proven effective for disease diagnosis and risk stratification, offering insights from patient data, medical records, and imaging results [17,18]. However, AI integration in the NICU environment is not without its challenges. Concerns related to data privacy, algorithm bias, and the need for transparent and interpretable models raise ethical considerations [19,20]. Understanding these challenges is essential for the responsible and successful implementation of AI for predictive analytics in neonatal care.

Despite the growing body of research, existing reviews in this field [12,21,22] have primarily focused on technical performance metrics, such as model readiness and diagnostic accuracy, while overlooking critical themes such as stakeholder engagement, ethical considerations, and data integration. For example, Schouten et al [22] focus on model evaluation but overlook broader integration opportunities, Adegboro et al [21] highlight health impacts but neglect barriers like trust and collaboration, and McAdams et al [12] emphasize predictive performance without addressing implementation challenges such as data quality and explainability. To address these gaps, this study synthesizes recent research (2017-2023) through a thematic lens, providing actionable insights into opportunities such as predictive modeling and personalized neonatal care, as well as challenges such as ethical issues and stakeholder trust. A systematic review design was chosen to ensure a comprehensive and methodologically rigorous synthesis of available literature on AI applications in NICU settings. This approach allows for structured evidence appraisal and thematic synthesis across diverse study designs and outcomes, providing robust and evidence-based recommendations to guide clinical, policy, and research decisions in this rapidly evolving field.

The importance of advancing research in this area becomes even more apparent when considering that other researchers have highlighted the lack of investigation into the practical implementation of AI in intensive care units. For instance, Vellido et al [23] lament the paucity of research on success factors for implementing machine learning in intensive care units, while Kim et al [24] point out that although challenges to implementing these technologies are acknowledged, they remain underexplored. This study seeks to address these gaps by providing a comprehensive and updated synthesis to guide future research and practical implementation in NICUs.

This review aimed to address the research question, “What are the opportunities and challenges in using AI in the NICU to predict length of stay and clinical outcomes?”

The objective was to analyze the current AI research landscape for predicting clinical outcomes and length of stay in the NICU, exploring the benefits and challenges of using AI in the NICU for these predictions. This will help identify gaps in AI applications for predicting clinical outcomes and length of stay in the NICU and offer recommendations for future research.

A systematic review methodology was applied to provide a robust and transparent synthesis of evidence related to AI applications in neonatal intensive care. This approach followed the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines and involved systematic search, screening, data extraction, quality appraisal, and thematic analysis across multiple databases [25]. The review was tailored to the digital health context by focusing on studies with predictive outcomes in the NICU, aligning with current research priorities and ensuring methodological depth.

To enhance relevance and address recent advances, a targeted supplementary search of literature published post–March 2023 was conducted to identify key developments for discussion. These findings were integrated thematically into the Discussion section.

English language peer-reviewed studies between January 2017 and March 2023 were searched in 6 databases (Embase, MEDLINE, CINAHL, Cochrane, Informit, and La Trobe Library). The search used relevant search strings and Medical Subject Headings (MeSH). The keywords included artificial intelligence, NICU, length of stay, and outcome, as detailed in Table 1.

The inclusion criteria focused on NICU patients and the adoption of AI technology for predicting outcomes and length of stay. Outcomes were defined as measures related to the adoption of AI technology for predicting length of stay and clinical outcomes in NICU patients, including mortality, morbidity, growth and development, and other clinical parameters up to 5 years of age. This framework was consistently applied during the eligibility assessment to ensure a clear and systematic approach. Both retrospective and prospective studies were included for a comprehensive view. Retrospective studies offer historical insights through electronic health records, while prospective studies provide real-time data for dynamic observations. The combination ensured a comprehensive evaluation, considering AI's opportunities and challenges in the NICU. Studies within the last 5 years, in English, and with full-text availability were included, aiming to understand AI's role in predicting NICU outcomes and length of stay. Multimedia Appendix 1 details the reasons for exclusion of studies during the screening process.

The quality appraisal process was conducted using a structured approach to ensure rigor and minimize bias. Multiple authors were involved in the selection and coding of studies to enhance reliability and reduce subjective influence. Initially, ST screened titles and abstracts for relevance, whereas URK and RB independently reviewed the screening, following which discrepancies were resolved through discussion or consultation during regular meetings.

The included studies underwent quality assessment using modified WHO (World Health Organization) guidelines [26] and preferred QCC (Quality Checklist for Clinical Case Series) [27] due to varied study designs. Multimedia Appendix 2 [28-51] contains detailed ratings, evaluating bias and quality.

Relevant data from the selected studies were systematically extracted and recorded in Microsoft Excel. A customized template, designed collaboratively by the research team specifically for this systematic review, was used for data extraction. The template included key features, such as study characteristics (background, methods, results, conclusions, and impact), study details (type of study, keywords, country, and number of participants), technology-related information (technology type and algorithm or model details), outcome (length of stay or outcome), evaluation measures, as well as insights on potential opportunities, challenges, discussed gaps, further research needs, and suggested improvement areas. The extracted data were organized and analyzed for synthesizing key findings and identifying and coding emerging themes relating to the research objectives using Thomas and Harden’s method of thematic analysis [52]. Studies identified for full-text review were coded to identify recurring concepts, which were grouped into descriptive themes. Analytical themes, such as those related to challenges and opportunities, were developed to interpret the findings and provide actionable insights. This method ensured a structured and transparent synthesis of diverse qualitative data. The identification and development of themes were conducted iteratively, with refinements made after each round of discussions among the research team to achieve consensus and ensure accuracy. This collaborative process ensured a comprehensive and balanced synthesis of the evidence, leveraging diverse expertise to strengthen the review's findings.

The maturity levels of AI adoption in the included studies were assessed using a structured framework defining “exploring,” “emerging/activating,” or “integrated ecosystem” stages [53]. These stages were defined as follows:

Explorer (exploring): This stage reflects ad hoc efforts to leverage AI in health care, with no established benchmarks or strategic frameworks. Stakeholders, including policymakers and governments, may have begun exploring AI's potential but have not yet developed or drafted relevant policies or guidelines.

Emergent (emerging or activating): At this stage, efforts to use AI in health care are more systematic and linked to a national AI strategy with clearly defined priorities. Policies and guidelines supporting AI inclusion in public health are drafted, demonstrating progress toward structured integration.

Leader (integrated ecosystem): This stage represents the highest maturity level, where AI adoption is embedded in national strategies and aligned with public health goals. Policies and guidelines are implemented and continuously updated, reflecting a well-established and operational AI ecosystem in health care.

These indicators encompassed several dimensions, such as the presence of national AI strategies, systematic exploration of AI applications, and the existence of policies or guidelines supporting AI integration. Careful consideration was given to the content, including mentions of AI strategies, the level of systematic exploration, and any references to comprehensive policies or frameworks. Each article was then categorized according to the maturity level that most accurately reflected its alignment with these dimensions. ST initially performed the categorization, which was then reviewed by URK and ML.

This methodical process ensured that the categorization of articles was based on discernible criteria, promoting consistency and reliability in assessing the state of AI adoption in neonatal care research. By applying these definitions and criteria, the assessment provided a clear understanding of the progress and gaps in AI maturity across the included studies.

Participant cohort sizes were categorized as small (fewer than 100 participants), medium (100 to 1000 participants), and large (more than 1000 participants) to standardize the classification of studies based on the scale of their population data.

In evaluating the performance measures for the AI models used in the included studies, a systematic approach was used to rate their effectiveness based on the results obtained and the type of measure used, supported by statistical norms for machine learning [54]. The evaluation results were categorized as follows.

The PRISMA 2020 checklist was used to guide the reporting of this systematic review, ensuring a comprehensive and transparent presentation of the methodology and findings (Multimedia Appendix 3). Certain items (such as statistical synthesis or meta-analysis) were not applicable due to the heterogeneity of study designs and outcomes, but all relevant PRISMA elements were addressed in line with best practices for qualitative systematic reviews.

A total of 811 studies were obtained and imported into the Covidence Systematic Review Software (Veritas Health Innovation). Among them, 85 duplicates were removed, resulting in 726 studies for screening. Out of the initial 726 studies, 169 were retained following a title and abstract review. The final selection comprised 24 articles deemed appropriate and included following a full review. A PRISMA flow diagram (Figure 1) illustrates the process of article selection and exclusion.

The objective of this systematic review was to analyze the current AI research landscape for predicting clinical outcomes and length of stay in the NICU, exploring the benefits and challenges of using AI in the NICU for these predictions. The surge in studies between 2020 and 2021 reflects a growing recognition of AI's potential to reshape neonatal care, possibly accelerated by the COVID-19 pandemic's impact on health care systems and driving innovative solutions to mitigate disruptions. The sustained research momentum, including studies in 2022, signifies an enduring commitment to exploring AI's role in the NICU.

This study contributes to the growing field of AI in NICUs by addressing gaps left by previous reviews and advancing the discourse on critical factors for successful AI adoption. Unlike earlier reviews, which predominantly focused on technical aspects and model performance, this study emphasizes emerging themes, including the importance of explainability, multidisciplinary collaboration, and stakeholder engagement. For instance, while [22] and [12] underscore technical barriers and model readiness, they lack focus on real-time data integration and actionable pathways for clinical implementation. This study builds on their findings by highlighting the transformative potential of AI in enabling predictive modeling, personalized care, and multidimensional data synthesis, essential for advancing NICU practices. Furthermore, this review synthesizes research through a broader thematic lens and incorporates the most recent evidence from 2017 to 2023. The systematic review design ensures a methodologically rigorous and transparent synthesis of the latest research, enabling relevance and depth in addressing practical challenges such as ethical concerns, stakeholder trust, and system integration. By using structured quality appraisal and thematic analysis across diverse study types, this review offers actionable insights that are not only academically robust but also critically important for guiding health care professionals, policymakers, and technologists in implementing AI solutions tailored to the unique complexities of NICU environments.

In terms of technologies used, deep learning emerges as a dominant technology across various clinical outcomes, highlighting its effectiveness in automatically identifying relevant features from raw data, especially in medical imaging and clinical analysis. However, the field's maturity in AI integration within neonatal care remains in the early exploration stage, lacking a mature and cohesive ecosystem.

The growing presence of various national guidelines and policies highlights the progress of AI implementation within health care [67-69]. However, there is a need to prioritize the development of a national AI strategy tailored for neonatal care due to the unique challenges of this environment. Specific considerations for neonatal care include the need for specialized models that account for the unique physiology and comorbidities of premature infants, as well as the heightened need of ensuring patient safety and data privacy in a patient population that have continuous and complete medical records from birth.

The study reveals substantial opportunities for the use of AI in the NICU to predict length of stay and patient outcomes. These cover diverse areas: advancements in medical imaging, data-driven insights and predictive models, improved understanding and risk stratification, and personalized neonatal care and intervention.

The first opportunity focuses on AI's potential to transform neonatal care in the NICU by leveraging machine learning to perform predictive analytics to help inform the clinician during the clinical decision-making process, aiding in earlier interventions. However, there are research gaps in translating AI's theoretical effectiveness into real-world NICU settings [3-6,70-74]. This brings model generalization and validation, the third challenge theme, which highlights the importance of ensuring AI models perform reliably across diverse patient populations. External validation across different populations is necessary for reliability and adaptability. The unique patient populations and scenarios in the NICU demand tailored approaches for model generalization and validation, vital for trustworthy predictions.

The second opportunity highlights the use of AI in NICU medical imaging. Six studies primarily use deep learning to analyze medical images, notably for conditions such as ROP. AI-driven imaging enhances accuracy, facilitating early intervention and personalized care plans. These findings not only affirm anticipated AI benefits in medical imaging but also underline its relevance in the NICU, particularly focusing on neonatal conditions. AI-powered imaging is key for precise diagnosis, standardized disease assessment, and predicting critical outcomes, aligning closely with literature expectations [15,75,76]. It enables tailored, timely care plans for individual neonatal needs, a significant advantage in this setting. The challenge here is handling clinical and diagnostic variability in these settings, where even subtle variations can have an impact on outcomes. Addressing this challenge involves the development of models and algorithms that can adapt to diverse clinical contexts, accounting for nuances that might otherwise be overlooked. It calls for an ongoing effort to refine models and diagnostic tools to accommodate the inherent variability in neonatal care, ultimately improving the precision and reliability of clinical decision-making.

The third and fourth opportunities highlight the transformative power of data-driven insights and predictive models in the NICU. Thirteen studies within these themes leverage AI to translate health care data into actionable insights, aiding decision-making. They encompass predictive modeling, spanning length of stay, disease severity, and mortality risk. These studies deepen the comprehension of neonatal conditions and their determinants, aligning with the literature's focus on advancing scientific understanding and medical care in the NICU [44,76-81]. By using data-driven approaches, they pave the way for redefining disease paradigms, ultimately enhancing interventions and care strategies. The challenge here is data quality and quantity, as it involves acquiring high-quality, sufficient data for AI model training. Data incompleteness and inconsistency, common issues in health care data, align with existing literature. Overcoming these challenges requires advanced algorithms and robust data management strategies within the NICU [80].

The final opportunity, personalized neonatal care and intervention, signals a significant shift in NICU health care. Studies focusing on tailoring care plans and interventions for individual neonatal needs explore personalized care plans, risk assessment, predicting motor outcomes, and distinguishing critical neonatal conditions. These efforts, using predictive models and continuous monitoring, aim to optimize interventions. Personalized care plans carry substantial implications, potentially reducing unnecessary treatments, cutting health care costs, and improving clinical outcomes. These align with literature stressing the importance of individualized neonatal care [71,82]. This is closely related to challenges of clinical interpretability and usability, as well as managing clinical and diagnostic variability. Clinical interpretability and usability focus on the transparency and practicality of AI-driven models within the NICU. Health care providers in the NICU demand accurate predictions and a clear understanding of AI outputs.

Building on the findings of this review, more recent AI research in NICU settings (post-2023) demonstrates continued advancement across key clinical domains, particularly neurological development, ophthalmological conditions, and respiratory outcomes. A growing body of literature emphasizes the integration of longitudinal and multisource datasets to strengthen predictive modeling. For instance, studies focused on neurodevelopmental outcomes have begun incorporating nationwide longitudinal clinical and sociodemographic data to more accurately identify risks of cognitive delay [83,84]. In the respiratory domain, pilot studies have explored the early detection of BPD by analyzing volatile organic compounds in exhaled breath, identifying novel biomarkers to support early diagnosis and intervention [85]. Similarly, machine learning techniques have been applied to uncover predictive clinical variables spanning the perinatal and neonatal periods, enhancing early risk stratification for respiratory morbidity [86]. Recent developments also highlight a shift toward multifactorial models that enable personalized care planning. Notably, the Mendelian Phenotype Search Engine (MPSE) uses natural language processing and machine learning to automatically identify high-risk neonates likely to benefit from whole genome sequencing, improving diagnostic yield and enabling timely interventions within the first 48 hours of NICU admission [87]. Parallel advancements in ophthalmology include AI-assisted diagnosis for ROP, with recent studies validating model outputs through clinician review to enhance diagnostic accuracy and trust [88,89]. These post-2023 developments align with our review findings and reinforce the trajectory toward integrated, real-time, and patient-centered AI solutions in neonatal care.

Regardless of any application, ethical and regulatory considerations and challenges remain in all applications. It emphasizes the complex landscape requiring careful navigation. Key ethical concerns include data sharing dilemmas and safeguarding patient data privacy and security. Regulatory hurdles, especially in validating AI models across diverse populations and settings, require adaptable frameworks. Only 4 studies directly mention these ethical and regulatory challenges, suggesting a potential research gap in the NICU field of AI, requiring more exploration and consideration. Ensuring responsible AI adoption is pivotal for equity in health care access and maintaining high standards of care.

Recent studies confirm these challenges and suggest strategies to address them. For instance, learning models, enabling collaborative training across institutions without sharing raw data, have emerged as a practical solution to mitigate privacy risks while enhancing model robustness [21,90]. Large-scale neonatal research networks, such as the Vermont Oxford Network, have emerged as examples that offer collaborative frameworks for aggregating high-quality data to overcome issues of data scarcity and variability [12]. Challenges like clinical usability and model generalization can be addressed through explainable artificial intelligence (XAI), which enhances transparency and builds clinician trust [22]. This can be achieved by transparent data practices and clear disclosures about imbalanced datasets, which are essential for meeting the needs of underserved populations [90]. Additionally, integrating ethics review boards early in the AI development process can guide data governance and ensure compliance with privacy laws [12]. Addressing ethical challenges also requires diverse teams of researchers, engineers, informaticists, and neonatologists to tackle these problems using equity-focused frameworks, particularly for addressing historically understudied challenges in neonatal care. Multicenter validation and standardized data collection are also important to ensure model reliability and address variability across clinical environments [12]. Lastly, centralized AI platforms could be used to streamline model integration into workflows, support continuous updates, and address performance degradation over time [22]. Regular audits, along with clear documentation of model development and trade-offs, further ensure responsible implementation [90]. These strategies provide actionable pathways for advancing robust, ethical, and equitable AI adoption in NICUs, paving the way for impactful neonatal care.

Future directions for NICU opportunities present key research areas. First, integrating AI with medical imaging, especially for conditions like ROP, demands refined technology, larger datasets, and real-world validation. These challenges pertaining to data integration have also been highlighted in previous reviews [12,22]. Second, enhancing data-driven insights and predictive models requires broader clinical scenarios and improved data quality via collaborative NICU efforts. Building on recommendations in prior reviews to explore clinician trust in AI systems, this study highlights the need for actionable strategies to enhance transparency and multidisciplinary collaboration [21]. Understanding neonatal conditions better, exploring diverse risk factors, and fostering multidisciplinary collaborations are key priorities. Practical AI application in NICUs, optimizing resource allocation and care, requires real-world implementation and thorough assessment. Lastly, refining scalable personalized care plans and interventions, while ethically considering AI personalization, remains essential for expanding AI's role and improving neonatal outcomes.