Introduction
Over the past two decades, neonatal haemodynamic assessment has evolved from an approach primarily guided by clinical examination and conventional physiological monitoring to one centred on bedside functional imaging. Reference McNamara, Jain and El-Khuffash1,Reference Nestaas2 Targeted neonatal echocardiography has been instrumental in this transformation, enabling direct evaluation of cardiovascular function and supporting clearer individualised diagnosis and management of critically ill newborns. Reference McNamara, Jain and El-Khuffash1–Reference Miller, Stoller and Fraga3 As neonatal cardiovascular evaluation and treatment have become increasingly sophisticated, neonatologists are now required to integrate multiple functional, anatomical, and physiological variables into real-time haemodynamic decision-making, making image acquisition, quantitative assessment, and interpretation more complex and operator-dependent. Reference McNamara, Jain and El-Khuffash1,Reference Miller, Stoller and Fraga3,Reference Raissi-Dehkordi, Raissi-Dehkordi and Xu4 Consequently, variability in image quality, measurement reproducibility, and clinical interpretation remains a significant challenge, even in experienced centres. Reference Miller, Stoller and Fraga3,Reference Ouyang, He and Ghorbani5,Reference Sahashi, Ouyang, Okura and Kagiyama6
Against this background, artificial intelligence is emerging as the next step in the evolution of neonatal cardiovascular imaging. Reference Rallis, Baltogianni, Kapetaniou and Giapros7,Reference Nikolaou, Chatzigiannis and Kefala8 Advances in artificial inteligence have demonstrated the ability to automate image analysis, improve quantitative assessment, optimise clinical workflows, and support interpretation of complex imaging data. Reference Raissi-Dehkordi, Raissi-Dehkordi and Xu4–Reference Sahashi, Ouyang, Okura and Kagiyama6,Reference He, Kwan and Cho9,Reference Vukadinovic, Chiu and Tang10 Importantly, the greatest promise of artificial inteligence lies not in replacing clinician expertise but in augmenting clinical decision-making through greater efficiency, consistency, and integration of imaging information into bedside care. Reference Topol11–Reference Lekadir, Frangi and Porras15 This shift from automation to collaboration is particularly relevant in neonatal haemodynamic assessment, where echocardiographic findings must always be interpreted within the broader physiological and clinical context of the individual infant. Reference Strong, Rogers and Sun13,Reference Vasey, Nagendran and Campbell14,Reference Myhre, Grenne and Asch16 This Viewpoint discusses how the future of neonatal haemodynamic assessment will be shaped not by autonomous algorithms, but by human–artificial inteligence partnerships that enhance diagnostic precision while preserving clinical judgement. Reference Topol11,Reference Vasey, Nagendran and Campbell14,Reference Lekadir, Frangi and Porras15,Reference Collins, Moons and Dhiman17,Reference Moons, Damen and Kaul18 However, much of the evidence supporting these capabilities derives from general or adult echocardiography, whereas neonatal applications remain at an early stage and are supported predominantly by retrospective, technical, or proof-of-concept studies. Reference Raissi-Dehkordi, Raissi-Dehkordi and Xu4–Reference Vukadinovic, Chiu and Tang10 Accordingly, their relevance to neonatal haemodynamic care should currently be regarded as promising but not yet established.
Reimagining the neonatal haemodynamic workflow
Artificial intelligence could potentially support several stages of the neonatal haemodynamic workflow, including image acquisition, quantitative analysis, physiological interpretation, clinical decision support, and longitudinal monitoring. Reference Raissi-Dehkordi, Raissi-Dehkordi and Xu4,Reference Sahashi, Ouyang, Okura and Kagiyama6,Reference Rallis, Baltogianni, Kapetaniou and Giapros7,Reference Strong, Rogers and Sun13,Reference Myhre, Grenne and Asch16,Reference Sermesant, Delingette, Cochet, Jaïs and Ayache19 However, these applications are at different stages of development, and many remain extrapolated from experience in general echocardiography rather than validated specifically in neonatal care. Their current value therefore lies primarily in providing a framework for future clinical development rather than an established end-to-end neonatal workflow.
During image acquisition, artificial inteligence-assisted computer vision can support real-time view recognition, image quality assessment, and operator guidance, potentially reducing variability and facilitating standardised examinations across different levels of expertise. Reference Raissi-Dehkordi, Raissi-Dehkordi and Xu4–Reference Sahashi, Ouyang, Okura and Kagiyama6,Reference He, Kwan and Cho9 Once high-quality images are obtained, automated chamber segmentation, functional quantification, Doppler analysis, and calculation of haemodynamic indices can substantially reduce analysis time while improving measurement reproducibility. Reference McNamara, Jain and El-Khuffash1,Reference Miller, Stoller and Fraga3,Reference Ouyang, He and Ghorbani5,Reference He, Kwan and Cho9,Reference Vukadinovic, Chiu and Tang10
Neonatal-specific research remains considerably narrower than the broader echocardiography literature and has focused mainly on defined technical or disease-specific tasks. Examples include automated recognition and extraction of standard cardiac views, machine-learning models combining clinical and echocardiographic variables to identify haemodynamically significant patent ductus arteriosus, and deep-learning analysis of pretreatment echocardiographic clips to predict pharmacological ductal closure. These studies demonstrate technical feasibility but remain largely retrospective or developmental. By contrast, automated comprehensive Doppler analysis, longitudinal multimodal interpretation, and bedside decision support have not yet been sufficiently validated for routine neonatal hemodynamic care. Reference Nikolaou, Chatzigiannis and Kefala8,Reference Leke, Sop Deffo and Wirsiy20–Reference Sharma, Gearhart and Luo22
Multimodal artificial inteligence systems developed predominantly outside neonatal care suggest the potential to synthesise heterogeneous data streams and identify evolving hemodynamic patterns. Reference Vukadinovic, Chiu and Tang10,Reference Rajpurkar, Chen, Banerjee and Topol12,Reference Myhre, Grenne and Asch16 In neonatal intensive care, however, this remains largely a future direction rather than an established clinical capability. Prospective neonatal studies are needed to determine whether integration of serial echocardiography with physiological monitoring and laboratory data can reliably support earlier recognition of circulatory deterioration and more individualised cardiovascular management. Reference Raissi-Dehkordi, Raissi-Dehkordi and Xu4,Reference Topol11,Reference Myhre, Grenne and Asch16,Reference Sermesant, Delingette, Cochet, Jaïs and Ayache19
Artificial intelligence as a clinical partner rather than a replacement
While these technological advances are rapidly transforming neonatal echocardiography, their role should be viewed as complementary rather than substitutive. The rapid expansion of artificial intelligence in medical imaging has fuelled concerns that automation may eventually replace human expertise. Reference Topol11,Reference Rajpurkar, Chen, Banerjee and Topol12 However, growing evidence suggests that the greatest clinical value of artificial inteligence lies not in autonomous diagnosis but in augmenting human cognition by automating repetitive tasks, improving measurement consistency, identifying subtle imaging patterns, and supporting standardised interpretation. Reference Raissi-Dehkordi, Raissi-Dehkordi and Xu4–Reference Sahashi, Ouyang, Okura and Kagiyama6,Reference He, Kwan and Cho9,Reference Vukadinovic, Chiu and Tang10
This collaborative model is particularly relevant in neonatal hemodynamic assessment, where echocardiographic findings cannot be interpreted in isolation. Clinical decision-making requires integration of imaging findings with gestational age, transitional physiology, pathophysiology, respiratory support, pharmacological therapies, laboratory data, and the infant’s evolving clinical condition. Reference McNamara, Jain and El-Khuffash1–Reference Miller, Stoller and Fraga3 Consequently, identical echocardiographic measurements may have different clinical implications depending on the physiological context, making clinician judgement indispensable despite advances in automated image analysis. Reference McNamara, Jain and El-Khuffash1,Reference Topol11,Reference Strong, Rogers and Sun13
In practice, safe human–artificial inteligence collaboration will require predefined pathways for managing system limitations and failures. Artificial inteligence systems should identify inadequate image quality, low-confidence outputs, and cases that may fall outside their training distribution rather than provide apparently definitive results. In such situations, the clinician should repeat or modify image acquisition, perform an independent assessment, or seek expert review. When an artificial inteligence output is physiologically implausible or conflicts with the infant’s clinical condition, it should not guide treatment until the discrepancy has been resolved, and clinicians must be able to override or suspend the system. Final responsibility for interpretation and therapeutic decision-making should remain with the treating physician. Reference Topol11,Reference Strong, Rogers and Sun13,Reference Myhre, Grenne and Asch16
Clinical implementation will also require physician training to combine continued competence in image acquisition and physiological interpretation with basic artificial inteligence literacy, including awareness of model limitations, uncertainty, bias, and automation-related errors. In practice, clinicians should verify image quality and automated measurements, interpret outputs within the infant’s evolving clinical context, and override recommendations when appropriate. Final responsibility for interpretation and management must remain with the treating physician. Reference Topol11,Reference Strong, Rogers and Sun13–Reference Moons, Damen and Kaul18
From innovation to clinical implementation
Despite rapid advances in artificial intelligence, successful integration into neonatal haemodynamic assessment will depend less on algorithmic performance than on clinical trustworthiness. Reference Vasey, Nagendran and Campbell14–Reference Myhre, Grenne and Asch16 Although many artificial inteligence models achieve excellent performance under experimental conditions, few have been prospectively validated in neonatal intensive care settings, where limited datasets, heterogeneous imaging protocols, and the unique physiology of newborn infants challenge model robustness and generalisability. Reference McNamara, Jain and El-Khuffash1,Reference Raissi-Dehkordi, Raissi-Dehkordi and Xu4,Reference Sahashi, Ouyang, Okura and Kagiyama6,Reference Nikolaou, Chatzigiannis and Kefala8,Reference He, Kwan and Cho9,Reference Myhre, Grenne and Asch16
Implementation therefore requires more than technical accuracy. TRIPOD+AI can improve transparent reporting of neonatal populations, input data, outcomes, and model performance, whereas PROBAST+AI can help identify risks of bias and limited applicability arising from small, heterogeneous datasets. DECIDE-AI provides a framework for early prospective evaluation within real clinical workflows, including human–artificial inteligence interactions and unintended effects, while FUTURE-AI extends this perspective across the system’s lifecycle by emphasising fairness, explainability, robustness, and continuous performance monitoring. Reference Vasey, Nagendran and Campbell14,Reference Lekadir, Frangi and Porras15,Reference Collins, Moons and Dhiman17,Reference Moons, Damen and Kaul18 Together, these frameworks offer complementary guidance for determining whether an artificial inteligence system is sufficiently transparent, clinically relevant, and safe for neonatal intensive care.
Clinical translation should follow a staged pathway, including multicentre external validation, prospective evaluation under real-world conditions, and clinical-impact studies assessing safety, workflow integration, and patient-relevant outcomes in addition to accuracy. Initial deployment should involve clearly defined tasks, mandatory clinician review, and predefined escalation and override procedures. Continuous post-deployment monitoring will also be necessary to identify performance drift, subgroup-specific errors, and unintended clinical consequences. Reference Vasey, Nagendran and Campbell14,Reference Lekadir, Frangi and Porras15,Reference Collins, Moons and Dhiman17,Reference Moons, Damen and Kaul18
Future artificial inteligence systems should therefore move beyond isolated image analysis toward longitudinal, multimodal models capable of learning from serial echocardiography, physiological monitoring, laboratory data, and therapeutic interventions. Reference McNamara, Jain and El-Khuffash1,Reference Nestaas2,Reference Myhre, Grenne and Asch16,Reference Sermesant, Delingette, Cochet, Jaïs and Ayache19 Achieving this vision will require multicenter collaboration, standardised imaging protocols, interoperable datasets, and prospective validation before widespread clinical adoption can be justified. Reference Vasey, Nagendran and Campbell14,Reference Lekadir, Frangi and Porras15,Reference Collins, Moons and Dhiman17 Ultimately, the success of artificial inteligence in neonatal haemodynamic assessment will be measured not by its analytical performance alone, but by its ability to improve clinical decision-making, reduce unwarranted variability, and strengthen clinician confidence while preserving human judgement at the point of care. Reference Topol11,Reference Strong, Rogers and Sun13,Reference Myhre, Grenne and Asch16
Conclusion
Artificial intelligence has the potential to become an integral component of neonatal haemodynamic assessment by enhancing image acquisition, quantitative analysis, and clinical decision support. Realising this potential, however, will require trustworthy, transparent, and prospectively validated systems that integrate seamlessly into clinical workflows while maintaining clinician oversight and accountability. As neonatal cardiovascular care becomes increasingly data-rich and physiologically complex, the true value of artificial inteligence will be measured not by its analytical performance alone, but by its ability to improve clinical decision-making, reduce unwarranted variability, and ultimately contribute to better outcomes for critically ill newborn infants.
