High-throughput analysis of multimodal monitoring data: the role of machine learning in early warning systems for high-risk neonates Authors: Huiyi Huo, Yongxue Lu, Jinyu Zhou, Li Zhang, Jinjuan Pei, Chao Hu Published: 2026 ABSTRACT AbstractThe neonatal intensive care unit (NICU) generates vast amounts of high-throughput, multimodal monitoring data, offering unprecedented potential for identifying early signs of clinical deterioration in high-risk neonates. However, the traditional threshold-based alarm systems are plagued by high false alarm rates and alarm fatigue, failing to harness this data complexity. This narrative review examines the role of machine learning (ML) in transforming early warning systems (EWSs) by effectively analysing these complex data streams. We first characterise the diverse sources—including physiological waveforms, neuromonitoring signals, electronic health records and emerging behavioural data—and inherent challenges (eg, noise, heterogeneity, label scarcity) of NICU data. We then detail key ML technologies, from preprocessing and feature engineering to core algorithms like deep learning models (recurrent neural networks, convolutional neural networks, Transformers) and multimodal fusion strategies, emphasising their application in handling time-series data. The review catalogues empirical evidence of ML-driven EWS for critical conditions such as sepsis, necrotising enterocolitis, neurological injury and cardiorespiratory instability, highlighting performance improvements over conventional methods. Finally, we discuss the significant technical, clinical integration and ethical challenges that impede widespread adoption and outline future directions, including federated learning, digital twins and cloud-edge architectures. The integration of ML-based insights promises to shift neonatal care from a reactive to a proactive, personalised paradigm, ultimately aiming to improve outcomes for vulnerable infants. Introduction The neonatal intensive care unit (NICU) is a critical setting for the treatment of high-risk neonates. For the purposes of this review, this term defines the population of infants requiring NICU admission due to conditions that place them at significant risk for high morbidity and mortality. This group primarily includes preterm infants, low birth weight infants and neonates with specific conditions such as perinatal asphyxia. These infants, whose physiological systems are still immature, are highly vulnerable to disturbances in their internal homeostasis and face a range of complications with high morbidity and mortality, such as sepsis, necrotising enterocolitis (NEC), intraventricular haemorrhage and respiratory distress syndrome (RDS).1 A central challenge in clinical practice is that the early signs of these critical conditions are often subtle or non-specific. By the time obvious deterioration in vital signs occurs, the window for optimal intervention may have already passed. The existing monitoring paradigm primarily relies on bedside monitors for the continuous surveillance of basic vital signs such as heart rate (HR), oxygen saturation (SpO₂) and respiratory rate (RR). Its early warning mechanism is typically based on preset single-parameter thresholds, for instance, bradycardia or a drop in SpO₂.2 However, the effectiveness of this method is severely limited. Studies indicate that threshold-based alarm systems have an extremely high false positive rate, with over 90% of alarms in the NICU not corresponding to genuine clinical deterioration events, which directly leads to the widespread phenomenon of alarm fatigue. In direct response to this challenge, machine learning (ML) models have shown significant promise in reducing false alarm rates; for instance, recent work has demonstrated the ability to reduce false bradycardia alarms by 36% while maintaining high detection sensitivity.3 Nevertheless, healthcare providers become desensitised due to frequent invalid alarms and may overlook truly critical warning signals, thereby compromising patient safety. Furthermore, this reactive approach fails to capture the complex, dynamic interactions among multiple physiological systems that occur prior to the onset of illness. With the advancement of information technology, modern NICUs have evolved into data-intensive environments. Bedside monitors continuously capture high-resolution physiological data, such as ECG, photoplethysmography (PPG) and respiratory waveforms, at a hertz-level frequency. Concurrently, the electronic health record (EHR) system consolidates vast amounts of clinical information, including laboratory test results, medication records and nursing documentation. These high-throughput, multimodal data streams form a complex and dynamic dataset, which contains rich pathophysiological information far beyond the reach of traditional threshold-based methods. The core challenge lies in how to extract subtle yet decisive early warning signals from this massive and noisy sea of data. ML, as a branch of artificial intelligence (AI), provides a powerful technical framework for addressing this issue. Unlike traditional statistical models that rely on predefined rules, ML algorithms can autonomously learn complex non-linear relationships and high-order feature interactions from large-scale data, thereby constructing high-precision predictive models.4 5 Particularly in handling time-series and multimodal data, deep learning models such as recurrent neural networks (RNNs) and convolutional neural networks (CNNs) have demonstrated exceptional performance. This narrative review aims to provide a comprehensive overview of the emerging field of leveraging ML technologies to analyse high-throughput multimodal monitoring data for constructing and optimising early warning systems (EWSs) for high-risk neonates. Recently, Millana et al published a systematic review on the use of AI in the NICU, comprehensively surveying AI applications across clinical domains and concluding that most models remain at the proof-of-concept stage with limited external validation.6 That work provides an invaluable panoramic view of the field. The present narrative review complements it by narrowing the focus to a specific, technically challenging subdomain: high-throughput multimodal monitoring data for EWS. While Millana et al offer breadth and evidence-quality assessment, we provide depth in ML methodologies—particularly for time-series analysis, multimodal fusion and handling of NICU-specific data challenges—and offer a forward-looking technical framework for researchers and engineers. Initially, the complex landscape of multimodal data within the neonatal intensive care environment will be examined, encompassing its diverse sources, distinctive characteristics and core challenges in processing. Subsequently, key methodological advances in ML applicable to such data analysis will be elaborated, with particular emphasis on deep learning models designed for time-series and multimodal fusion. Building on this foundation, the paper will categorise and detail the specific application outcomes and empirical research of these technologies in the early warning of major neonatal critical conditions such as sepsis, NEC and brain injury. Following this, current challenges in the field—spanning technical implementation, clinical translation and ethical considerations—will be prudently discussed, leading to proposed future research directions and development pathways. Finally, the prospects of an ML-driven paradigm shift in neonatal monitoring will be summarised and envisioned. A graphical abstract summarising the key concepts of this review is provided as figure 1. Graphical abstract. CNN, convolutional neural network; EEG, electroencephalogram; HIE, hypoxic-ischaemic encephalopathy; NEC, necrotizing enterocolitis; RNN, recurrent neural network; XAI, explainable artificial intelligence. Methods This manuscript is a narrative review. Its purpose is to provide a broad, educational and comprehensive overview of the application of ML for EWSs in the NICU, summarising the current state of the art, challenges and future directions. The literature included was selected based on its relevance to the core topics discussed—including data sources, ML methodologies, clinical applications and implementation challenges—drawing from the authors’ expert knowledge in the field. While literature searches were conducted in databases such as PubMed, Institute of Electrical and Electronics Engineers (IEEE) Xplore and Google Scholar, this review does not adhere to a formal, prespecified and reproducible protocol, such as the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. Consequently, it does not involve a systematic search strategy, predefined inclusion/exclusion criteria or a formal quality assessment of the cited studies. The goal is to synthesise the key concepts and evidence into a coherent narrative for clinicians, engineers and data scientists, rather than to answer a specific, narrowly-defined research question. Multimodal data sources and characteristics in the NICU The cornerstone of constructing an effective ML-based early warning model lies in a comprehensive and nuanced understanding of the available data. Data sources within the NICU are characterised by remarkable diversity, which can be broadly categorised into several distinct yet complementary types, as visualised in figure 2. Landscape of multimodal high-throughput monitoring data in the NICU. The data ecosystem comprises three primary, complementary categories: (1) continuous physiological waveforms (eg, ECG, EEG) providing high-resolution, real-time biosignals; (2) discrete clinical metadata from the electronic health record (EHR) and laboratory systems, offering contextual and episodic data points and (3) novel and emerging data streams, such as video analytics and ambient acoustics, which expand the monitoring paradigm beyond traditional biosignals. This heterogeneous, multimodal nature forms both the foundation and the core challenge for subsequent machine learning analysis. EEG, electroencephalogram; NICU, neonatal intensive care unit; SpO₂, oxygen saturation. Main data types Continuous physiological time-series data represent the primary data stream, captured non-invasively at the bedside. The ECG provides a high-frequency waveform of cardiac electrical activity, from which critical metrics such as HR and heart rate variability (HRV) are derived. HRV, reflecting autonomic nervous system (ANS) regulation, has been established as a sensitive early biomarker for conditions like sepsis, with a reduction in its complexity often preceding clinical deterioration. PPG is used for the continuous, non-invasive monitoring of peripheral SpO₂ and pulse rate. The stability, frequency and depth of SpO₂ fluctuations serve as vital indicators of respiratory compromise. Respiratory signals, obtained via thoracic impedance or flow sensors, enable the calculation of RR and the critical identification of apnea events, distinguishing between central, obstructive and mixed types.7 Blood pressure (BP), measured either invasively via arterial lines or non-invasively via cuffs, provides a direct assessment of circulatory function. Core or peripheral temperature monitoring offers a continuous readout of thermal stability, where instability can be an early sign of systemic infection or other pathologies. For neonates at risk of neurological injury, specialised neuromonitoring data are indispensable. The electroencephalogram (EEG) remains the gold standard for assessing cortical electrical activity, employed to detect subclinical seizures and evaluate brain maturation and injury. Its processed, simplified form, the amplitude-integrated EEG, facilitates continuous bedside monitoring of background pattern trends, which is crucial for managing conditions like hypoxic-ischaemic encephalopathy (HIE).8 9 Near-infrared spectroscopy offers a non-invasive means to monitor regional cerebral oxygen saturation, providing real-time insights into cerebral perfusion and oxygenation status, which is particularly critical following perinatal asphyxia. Clinical and laboratory data, primarily sourced from the EHR system, supply essential contextual information. The EHR encompasses demographic details (eg, gestational age, birth weight), diagnostic codes, medication administration records, fluid balance, ventilator settings and nursing notes. Laboratory results, though typically sampled intermittently, provide biochemical evidence of the infant’s status, including complete blood counts, C-reactive protein levels, blood gas analyses, electrolytes and glucose levels.10 Emerging data modalities are expanding the monitoring paradigm beyond traditional biosignals. Although research using these modalities is largely still in the proof-of-concept and early clinical validation stages, they show significant promise. Video and audio data, captured via cameras and microphones, enable the computational analysis of neonatal behaviour. Specific application examples include using convolutional neural networks to assess facial expressions for automated, continuous neonatal pain assessment,11 and the acoustic analysis of crying characteristics to detect early signs of neurological distress. Additionally, algorithms can track spontaneous body movement patterns to aid in the early prediction of motor disorders like cerebral palsy. These behavioural phenotypes offer a novel window into neurological integrity and overall state of comfort or distress, potentially correlating with underlying pathological states. Challenges of neonatal monitoring data Leveraging these multimodal data for ML modelling faces a series of inherent challenges. The data are characterised by high dimensionality and heterogeneity, encompassing diverse types and structures such as continuous waveforms, discrete numerical values and textual records, with sampling frequencies ranging from millisecond-level (eg, ECG) to hourly or even daily intervals (eg, laboratory tests). Effectively aligning and integrating these heterogeneous data sources presents a significant difficulty. Furthermore, the data are contaminated by high levels of noise and artefacts; frequent neonatal movements, routine care procedures (like suctioning or diaper changes) and poor electrode contact can introduce substantial interference into physiological signals. These artefacts risk being misinterpreted as pathological events or obscuring genuine pathological signatures, necessitating robust identification and processing methods. Compounding these issues are data sparsity and irregularity: unlike continuous physiological monitoring, laboratory results and clinical assessments are acquired sparsely and at non-uniform intervals, requiring models capable of handling this mixed sampling regime. Perhaps most critically, the field grapples with a scarcity of annotated data. The efficacy of supervised learning paradigms is fundamentally contingent on the availability of extensive, high-fidelity annotated datasets. However, within clinical contexts, the precise temporal demarcation of disease onset for complex syndromes, such as sepsis, constitutes a significant and persistent challenge. This determination often relies on retrospective judgement, which is subject to subjectivity and delay, thereby posing a major obstacle for model training. Additionally, the relative rarity of critical clinical events leads to severe class imbalance, further complicating the development of robust predictive models. Key ML technologies for high-throughput analysis To address the aforementioned challenges and extract actionable insights from this complex data, researchers have developed and applied a series of ML techniques. Figure 3 provides a schematic overview of a typical pipeline for constructing an ML-based EWS, from raw data ingestion to clinical insight. Schematic pipeline of a machine learning-based early warning system for high-throughput neonatal data. The pipeline begins with data acquisition and synchronisation of multimodal streams, followed by critical preprocessing steps (eg, artefact removal, imputation). Feature engineering bifurcates into extracting domain-knowledge-driven hand-crafted features (eg, HRV metrics) and employing deep learning models (eg, CNNs, RNNs) for automated representation learning. These features are integrated within an integrated model layer, where advanced strategies like attention-based multimodal fusion synthesise information before the final decision layer produces outputs such as a real-time risk score or a classification alert, ultimately aimed at generating actionable clinical insight. CNN, convolutional neural network; EWS, early warning system; HRV, heart rate variability; RNN, recurrent neural network. Data preprocessing and feature engineering Data preprocessing and feature engineering, as the starting point of the modelling pipeline, directly determine the upper limit of model performance. This primarily involves artefact detection and removal, where a variety of algorithms have been developed to handle noise in physiological signals, such as methods based on the signal quality index, wavelet transform denoising and independent component analysis.12 13 For heterogeneous data with different sampling frequencies, time alignment through upsampling or downsampling is necessary to achieve data synchronisation, along with reasonable interpolation methods for sparse data, such as simple forward filling or linear interpolation.14 In terms of feature extraction, one approach involves hand-crafted features, where domain experts define a series of highly interpretable features based on pathophysiological knowledge, for example, extracting time-domain, frequency-domain and non-linear HRV features from ECG signals; the other approach involves representation learning, which uses deep learning models—particularly CNNs and autoencoders—to automatically learn high-level abstract representations of data directly from raw or semi-processed signals, thereby avoiding the cumbersome and potentially suboptimal process of manual feature engineering.15 Core ML methods Building an effective ML-based early warning model is fundamentally reliant on a comprehensive understanding of the available data. Data sources within the NICU exhibit a high degree of diversity and can be broadly classified into the following categories. Regarding model construction, traditional ML models—such as logistic regression (LR), support vector machines, random forests and gradient boosting trees (eg, XGBoost)—have demonstrated excellent performance in numerous predictive tasks when coupled with effective feature engineering, while typically maintaining good interpretability and computational efficiency.4 16 17 Concurrently, technologies represented by deep learning offer powerful modelling capabilities. CNNs are primarily used to process data with a grid-like topological structure; in one-dimensional applications, they can effectively extract local patterns and morphological features from physiological time series, while in two-dimensional applications, they can be employed to analyse time-frequency spectrograms of signals or medical images.18 For time-series data, RNNs and their variants, such as long short-term memory (LSTM) networks and gated recurrent units (GRUs), have become the standard choice for modelling the dynamic progression of diseases due to their internal gating mechanisms, which effectively capture long-term temporal dependencies in the data.19 Furthermore, the Transformer model, originally designed for natural language processing, can weigh the importance of different time points in a sequence through its self-attention mechanism, excelling at capturing ultra-long-term dependencies; this architecture has now been successfully applied in the field of physiological signal analysis.20 Multimodal fusion strategy Building an effective ML-based early warning model hinges on the comprehensive integration of information from diverse data sources. Data fusion strategies can be primarily implemented in several ways. The foundational approach is early fusion, which involves concatenating raw data or extracted features from different modalities at the input stage and feeding them into a single model for training; while this method is straightforward, it necessitates meticulous data alignment across modalities. A more flexible alternative is late fusion, where individual models are trained on each data modality separately, and their outputs are subsequently combined at the decision level through techniques like voting or weighted averaging. However, this strategy might fail to capture the nuanced, low-level interactions between different data streams.21 22 Currently, the most advanced and increasingly adopted strategy is hybrid/attention-based fusion. This approach leverages mechanisms like attention to enable models to dynamically and selectively integrate information across modalities at various processing layers. By assigning higher weights to features from critical modalities or at pivotal time points, it achieves more sophisticated feature interaction and information synthesis, leading to enhanced model performance and interpretability.23 24 Key technologies for addressing special challenges The foundation of building an effective ML-based early warning model lies in a comprehensive understanding and proper handling of the inherent challenges within the data, among which class imbalance and label scarcity are two critical issues. During the data preprocessing stage, techniques such as oversampling (eg, synthetic minority over-sampling technique), undersampling or cost-sensitive learning can be employed to adjust the data distribution. By assigning a higher misclassification penalty to minority class samples, these methods balance class weights and prevent the model’s predictions from being excessively biased toward the majority class.25 26 In response to the costly and scarce data labelling scenarios prevalent in healthcare, weak supervision methods enable model training using incomplete or imprecise labels. Furthermore, self-supervised learning designs pretext tasks that allow models to learn general, transferable feature representations from large-scale unlabelled data. Subsequently, fine-tuning on a small amount of labelled data adapts the model to downstream tasks. This approach has been proven effective in leveraging unlabelled data resources, enhancing model performance while significantly reducing dependence on precise annotations, thereby offering a highly promising technical pathway for medical data analysis.27 Explainable artificial intelligence (XAI) In high-stakes clinical decision support scenarios, model transparency and interpretability are paramount. Providing a prediction outcome without an accompanying explanation makes it difficult to earn the trust of clinicians. XAI techniques, such as SHAP (SHapley Additive exPlanations) and LIME (local interpretable model-agnostic explanations), can quantify the contribution of each input feature to an individual prediction.28 29 This helps physicians understand the rationale behind the model’s decision, thereby enhancing its reliability and clinical utility. ML-driven EWS: application scenarios and empirical evidence ML techniques have been extensively employed in EWS research for a variety of critical neonatal conditions. Sepsis/late-onset sepsis Neonatal sepsis is a leading cause of mortality and long-term neurological sequelae in the NICU. Due to the difficulty of early clinical diagnosis, the prediction of this condition has become one of the most focused areas of EWS research. Early representative achievements in related studies were early warning models based on HRV analysis, the most notable of which is the Heart Rate Observation (HeRO) Score. By quantifying abnormal patterns in HRV, this system can predict sepsis risk approximately 24 hours before clinical diagnosis and has been confirmed in multiple randomised controlled trials to effectively reduce mortality in preterm infants.30 Recent research trends have shifted toward constructing multimodal ML models, aiming to enhance early warning performance by integrating multidimensional data sources. For instance, some studies have employed gradient boosting tree models that integrate HRV, respiratory patterns, temperature variability and laboratory indicators, demonstrating predictive performance significantly superior to traditional models that rely solely on HRV.31 Concurrently, deep learning models have been applied to directly analyse raw ECG and respiratory impedance, further improving the lead time and accuracy of warnings through the automatic extraction of predictive features.2 However, for clinical application, both approaches possess distinct advantages and disadvantages. Feature-based systems like the HeRO score rely on well-understood pathophysiological metrics (eg, HRV), making them highly interpretable, computationally lightweight and easier to gain clinicians’ trust. In contrast, end-to-end deep learning models directly processing raw waveforms can uncover subtle, complex patterns missed by human-defined features, potentially yielding higher accuracy. Yet, they often suffer from a lack of interpretability and require significant computational resources. Consequently, more complex models do not invariably lead to greater clinical benefit. The ultimate clinical value of an EWS depends not only on its predictive accuracy but equally on its transparency, interpretability and seamless integration into clinical workflows. The core advantage of these advanced models lies in their ability to identify subtle, prodromal patterns that manifest hours before overt clinical symptoms, a concept illustrated in figure 4. The window of opportunity: machine learning (ML)-based early warning systems (EWS) versus traditional threshold alarms. This graph compares the temporal predictive capacity of advanced ML models against conventional monitoring paradigms. Curve A (blue) represents traditional threshold-based alarms, which typically remain silent during the subclinical phase and only trigger at the point of clinical symptom onset, often leading to a reactive response. Curve B (red) illustrates the ML prediction trajectory, which captures subclinical physiological fluctuations to provide earlier risk identification. The shaded golden window for early intervention represents the critical time gained (eg, hours before onset), allowing healthcare teams to shift from a reactive to a proactive care paradigm. Necrotising enterocolitis (NEC) NEC represents the most common and life-threatening gastrointestinal emergency among preterm infants.32 Research has established that prior to the onset of NEC, ANS dysfunction occurs in affected infants, manifesting as alterations in HR and BP variability. ML approaches have been applied to analyse continuous HR and BP data to identify specific patterns associated with NEC, enabling early risk stratification.33 Further advancing this field, other studies have developed multimodal models that integrate vital signs, feeding data and textual reports from abdominal radiographs to predict the risk of NEC occurrence.34 Neurological injury and seizures The timely detection and intervention of brain injury constitute a central clinical objective for improving the long-term neurological outcomes of neonates. In this field, ML technologies, particularly deep learning models, have demonstrated significant value in two key application areas. Firstly, in the detection of neonatal seizures, continuous EEG monitoring is crucial, as epileptic seizures often manifest as electrical discharges without apparent clinical convulsions. Deep learning algorithms based on CNNs have been developed to automatically and accurately identify seizure events from multichannel EEG signals.35 Their detection performance approaches or even surpasses that of expert human interpretation, thereby substantially alleviating the burden on neurophysiologists tasked with analysing vast volumes of EEG data. Second, for infants with HIE, early and accurate prediction of long-term neurodevelopmental outcomes is critical for guiding therapeutic decisions.36 A study published in 2024 employed an ML model to integrate and analyse quantitative MRI data from the brain, along with demographic and laboratory data, of 286 HIE infants who had undergone therapeutic hypothermia. The aim was to predict their cognitive, language and motor function scores at 18 months of age. The results demonstrated that the model could significantly improve prediction accuracy. Particularly when quantitative MRI data were incorporated, its explanatory power surpassed that of models relying solely on traditional clinical data. This approach provides a robust tool for early and objective assessment of prognosis and for formulating individualised intervention plans.37 These advancements signify a viable pathway towards employing objective, data-driven methods for the precise and automated assessment of early risk and prognosis in neonatal brain injury. Cardiorespiratory instability Apnoea, bradycardia and desaturation (ABD) events in preterm infants are manifestations of cardiopulmonary immaturity and constitute a core challenge in clinical monitoring. Among these, the accurate and reliable detection of bradycardia is particularly critical, as frequent false alarms can trigger alarm fatigue, severely undermining healthcare providers’ trust in EWS. The research by Rahman et al directly addresses this clinical need.3 They developed a ML framework incorporating an output correction module, applying sequence models such as LSTM and GRU to analyse over 440 hours of real-time data from preterm infants. While achieving a high recall rate of 0.992, the model also maintained a false positive rate as low as 0.007, reducing false alarms by 36% compared with baseline methods. This outcome demonstrates that optimising the detection accuracy of individual events can effectively enhance the clinical credibility and usability of comprehensive ABD EWS, serving as a crucial technical prerequisite for their development. Other critical conditions: RDS and PPHN Beyond the aforementioned conditions, ML is increasingly being explored for the early warning and management of other significant neonatal morbidities, particularly RDS and Persistent Pulmonary Hypertension of the Newborn (PPHN). For RDS, predictive models using multimodal data (respiratory waveforms, SpO₂ variability, HRV) have shown early prediction of CPAP failure and need for mechanical ventilation in preterm RDS, with areas under the curve (AUCs) up to 0.93.38 Similarly, for PPHN, recent work has demonstrated that ML can predict PPHN severity in high-risk populations such as infants with congenital diaphragmatic hernia. An XGBoost model integrating prenatal ultrasound and MRI features achieved 82% accuracy for predicting severe PPHN, with an AUC of 0.82.39 Although EWS for these conditions are currently less mature than those for sepsis or seizures, they represent a crucial frontier for expanding the comprehensiveness of data-driven neonatal care. Challenges and future directions Despite significant progress made, the wide application of ML-driven EWS in clinical practice still faces many challenges. Technical challenges The development of effective ML-based early warning models faces several critical challenges, among which ensuring model generalisability and robustness is paramount. Models trained on data from a single medical centre often exhibit significant performance degradation when deployed at other centres due to systematic disparities in patient population characteristics, monitoring equipment specifications and clinical practice standards—a phenomenon known as domain shift. Consequently, enhancing the cross-centre and cross-population applicability of these models represents an urgent research priority. A concurrent technical challenge lies in achieving real-time and efficient multimodal data fusion. The practical implementation of bedside EWS demands that algorithms process and synchronously analyse heterogeneous data streams from multiple sources in real time. This imposes stringent constraints not only on the computational efficiency of the model architecture but also on the engineering realisation of the underlying system. Furthermore, the majority of current models are constructed based on population-level characteristics, failing to adequately account for the inherent individual variability in physiological baselines and stress responses among neonates. Future advancements in EWS should therefore focus on establishing dynamically evolving, personalised physiological baseline models for each infant, enabling more precise and sensitive detection of deviations from their normative physiological state. Furthermore, a critical challenge is the need to rigorously demonstrate the incremental clinical value of complex ML models over simpler, established methods. The literature is susceptible to publication bias, where studies with positive results are more likely to be published than those with null or negative findings, making it difficult to gauge the frequency of ‘failed attempts’. In various medical domains, it has been shown that complex models (eg, deep learning) do not always significantly outperform well-calibrated traditional statistical models (eg, LR) or existing clinical scoring systems, particularly when datasets are limited or feature engineering is robust. Therefore, it is imperative that validation studies for any proposed NICU EWS include a direct and fair comparison against relevant, simpler benchmarks. This ensures that the pursuit of algorithmic complexity translates into genuine, demonstrable improvements in predictive performance and is not merely an exercise in technical sophistication. Clinical transformation and ethical considerations The successful integration of ML-driven EWS into the clinical practice of the NICU depends not only on the performance of the models themselves but also on the ability to overcome a series of complex technical, regulatory and ethical challenges. At the level of technical integration, the core issue lies in designing effective human-computer interaction interfaces to achieve seamless alignment with existing clinical workflows. The method of presenting warning information is crucial and should avoid introducing new alarm fatigue. For instance, employing visualised risk trend charts or explanatory reports that incorporate the rationale behind model inferences can provide intuitive and contextually relevant information support for clinical decision-making, rather than relying on simple numerical alerts or disruptive alarms. From a regulatory approval perspective, such systems are typically classified as Software as a Medical Device and must adhere to stringent regulatory pathways to obtain market authorisation, such as through the U.S. Food and Drug Administration (FDA) or the European Union’s Conformité Européenne (CE) marking. This process requires developers to provide robust clinical validation evidence, including data from prospective randomised controlled trials, to demonstrate clinical efficacy and safety. Furthermore, data privacy, security and algorithmic fairness are foundational requirements that must be upheld throughout. Across the entire data lifecycle, strict compliance with relevant health data privacy regulations is mandatory. Simultaneously, model development necessitates rigorous bias assessment and mitigation to ensure that warning performance remains equitable across patient groups of different gestational ages, races and socioeconomic backgrounds. For instance, models trained primarily on late preterm infants may underperform in extremely preterm neonates due to distinct physiological baselines. Likewise, if certain racial or socioeconomic groups are underrepresented in training data, the model may perpetuate existing care disparities. A concrete example is pulse oximetry, a key input feature for many EWS, which is known to overestimate SpO₂ in individuals with darker skin pigmentation, potentially delaying hypoxemia detection in non-White infants if uncorrected.40 Therefore, fairness audits—including stratified performance analysis by gestational age, race and hospital type—should become routine in model validation to prevent algorithmic discrimination. Future development direction The foundation of building an effective ML-based early warning model lies in a profound understanding of data, while synergistic innovation across algorithms, systems and data architecture serves as the key driver for its advancement. At the algorithmic level, federated learning, as an emerging paradigm, enables multiple distributed medical institutions to collaboratively train a global model without exchanging raw sensitive data. This provides a viable pathway to address the limitations of multicentre research and collaboration under data privacy regulations, helps overcome the constraints of small sample sizes and insufficient representativeness inherent in single-centre data, and thereby enhances the generalisation capability of models. Another cutting-edge direction is digital twins technology. It is important to distinguish a digital twin from a highly personalised dynamic predictive model. While a predictive model forecasts clinical risks based on learnt correlations, a true digital twin is a bidirectional, mechanism-grounded framework that enables ‘what-if’ simulations—for example, predicting how an infant’s oxygenation would respond to a specific ventilator adjustment. In the NICU context, however, true digital twins remain a long-term vision due to the complexity of neonatal physiology and the scarcity of validated mechanistic models. Our use of the term is intended to highlight this aspirational direction rather than to claim near-term feasibility. At the system architecture level, future developments rely on constructing an intelligent, cloud-edge collaborative monitoring platform. This architectural design deploys lightweight real-time data processing and emergency event inference capabilities at bedside or proximate edge computing devices, while offloading complex model training, iterative updates and in-depth analytical tasks to cloud servers. Such a division of labour not only ensures the immediacy of critical warnings, reduces network dependency and latency, but also provides the necessary computational power for large-scale data aggregation and complex algorithm execution. Furthermore, the long-term vision for intelligent systems is to evolve from passive early warning to active closed-loop intervention. Under stringent safety constraints and clinical oversight mechanisms, such systems could dynamically adjust therapeutic parameters based on predicted risks. For example, automatic fine-tuning of ventilator support levels in response to early predictions of respiratory deterioration could form a complete control loop encompassing perception, analysis and intervention. Finally, infrastructure development at the data level forms the cornerstone supporting progress across the field. Creating standardised, large-scale, high-quality and rigorously de-identified public datasets, and establishing open benchmark testing platforms based on them, play an irreplaceable role in promoting the reproducibility of algorithmic research, enabling fair comparisons and accelerating rapid development. Successful experiences from platforms such as PhysioNet, achieved through organising open challenges, demonstrate that such initiatives can effectively stimulate research community engagement and accelerate the generation and clinical translation of universal solutions. Conclusion ML provides an unprecedented and powerful tool for parsing high-throughput multimodal monitoring data of high-risk neonates and uncovering underlying early warning signals. By integrating and analysing the complex patterns hidden within vast datasets, ML-based EWS have the potential to transform the paradigm of neonatal intensive care—shifting it from the traditional, threshold-based, reactive model to a more precise, personalised, proactive predictive and interventional model. Current research in this field has already demonstrated considerable application potential across multiple critical clinical areas, including sepsis, NEC and brain injury. However, the path from algorithmic validation to routine clinical application remains challenging. Future success will depend on close interdisciplinary collaboration: clinicians must define clinically meaningful questions and provide high-quality annotated data; data scientists and engineers need to develop robust, interpretable and efficient algorithms and systems and the involvement of regulatory bodies and ethicists is essential to ensure the safety, efficacy and fairness of these technologies. The ultimate objective is to establish an intelligent NICU environment where data-driven insights are seamlessly integrated into the clinical decision-making process. This integration aims to assist healthcare teams in earlier risk identification and more precise intervention implementation, thereby effecting a fundamental shift in clinical practice from a reactive alarm-response paradigm to a proactive, insight-driven approach to patient care. Such a transformation is anticipated to tangibly improve the health outcomes of millions of high-risk neonates globally.