Artificial intelligence-guided tools in adult ECMO: Current advancements,emerging Trends and future directions

Abstract

Introduction

Extracorporeal membrane oxygenation (ECMO) provides life support for patients with refractory cardiac or respiratory failure. The complexity of ECMO management and associated mortality necessitates high-accuracy clinical decision-making systems. Artificial intelligence (AI) has emerged as a potential approach to address challenges in ECMO management, from patient selection to real-time assessment and outcome prediction.

Objective

To synthesize the current evidence of AI application in adult ECMO, addressing predictive modelling for patient outcomes, real-time decision support systems, and complication prevention, as well as the evolving regulatory challenges governing medical AI deployment in critical care settings.

Methods

A narrative literature review was conducted across PubMed/MEDLINE, Embase, Cochrane Library, IEEE Xplore, and preprint servers (arXiv/medRxiv). The search strategy combined ECMO-relevant terms (“V-A ECMO”, “V-V ECMO”) with AI terminologies (“artificial intelligence”, “machine learning”, “deep learning”, “digital twin”). Studies were included if they focused on adult cohorts (age ≥18 years) and were published in English between 2018 and 2025.

Results

The review found several AI algorithms under development for different stages of ECMO therapy. AI algorithms have been developed to assist in the initiation, prognostication, complication detection, real-time control, and weaning of ECMO. However, none have been clinically translated thus far.

Conclusion

While AI for precision ECMO management is promising, several prerequisites remain unmet, including the integration of high-frequency device data, prospective external multicenter validation, and the development of robust regulatory frameworks. Securing these advances will bridge the gap between algorithm development and the clinical arena.

Introduction

Extracorporeal membrane oxygenation (ECMO) provides life-saving support for patients with refractory cardiac or respiratory failure. The complexity of managing these patients arise from the heterogeneous causes of organ failure, their distinct ECMO circuit and native circulation interactions, pathophysiological mechanisms, and the multitude of associated diagnostic and monitoring parameters. These factors generate large volumes of high-velocity data from laboratory tests, hemodynamic monitoring, mechanical ventilation, and the ECMO circuit itself.

Additionally, ECMO support is highly resource-intensive with substantial cost, various resource limitations across geographical regions, and high mortality rates despite technological and organizational advances. This clinical complexity, information overload and outcome uncertainty provides a need for more precise and reliable clinical decision support systems (CDSS). Conventional scoring systems to measure disease severity and predict survival such as SAVE (Survival After Veno-arterial ECMO; V-A ECMO) and RESP (Respiratory ECMO Survival Prediction; V-V ECMO) were validated for ECMO patients and typically rely on static discrete time-point patient data.^1,2 Although widely used, these algorithms are now over a decade old and have not been updated to reflect the significant advances in ECMO support. Additionally, they may fail to capture the dynamic interaction of rapid and wide fluctuations in hemodynamic instability, extracorporeal circuit parameters, and treatment intensity. Artificial intelligence (AI) has emerged as a promising approach as a CDSS in ECMO, from patient selection and cannulation timing, through real-time monitoring to complication prediction and decannulation. By leveraging large datasets from electronic health records (EHRs), registries and continuous device monitoring, it is possible to use machine-learning (ML), including deep-learning (DL) algorithms to identify non-linear patterns and interactions that may not be apparent to human providers. The major limitation regarding the clinical application of AI into ECMO management is the evolving stringent and regionally diverse regulatory landscape, which hinders the collection of large-scale datasets for training and the clinical deployment of models.

This narrative review categorizes AI applications in ECMO into three primary domains: indication and triage; prognosis and outcome prediction; and real-time control and decision support. This review provides commentary on the evolving AI regulatory landscape and identifies systemic challenges that must be addressed to facilitate the integration of AI as a CDSS. The sequential stages of the AI-augmented ECMO patient journey, as discussed throughout this review, are synthesized in Figure 1.

Figure 1.

The AI-Augmented Patient Journey and Data Layer. Schematic representation of artificial intelligence applications across the ECMO clinical pathway. Candidate Selection utilizes data-driven models (e.g., PreEMPT-ECMO, ForecastECMO) to move from static scoring to dynamic early warning and triage. Monitoring & Control focuses on precision management through real-time surveillance and emerging closed-loop systems. Weaning & Outcome integrates prognostic tools for survival (ECMO PAL, RESCUE-24), decannulation readiness, and long-term readmission risk (XGBoost). The underlying Data Layer highlights the critical dependency on data granularity. While current models largely rely on static EHR & Clinical Context (green), future real-time applications require the integration of High-Frequency Vitals and Device Sensors (orange/red), which currently lack standardized interoperability.

Methodology

A narrative literature review was conducted. A database search across PubMed/MEDLINE, Embase, Cochrane Library, IEEE Xplore, and preprint servers, arXiv/medRxiv was performed. The search strategy combined ECMO-relevant terms (“V-A ECMO”, “V-V ECMO”), with AI terminologies (“artificial intelligence”, “machine learning”, “deep learning”, “digital twin”). This review included studies involving adults (≥18 years old) and published in English language between 2018 and 2025. Data extraction prioritized study setting (registry vs single-center), AI methodologies, and performance metrics with Area Under the Curve Receiver Operating Characteristic (AUROC) or accuracy.

AI-guided prediction for ECMO initiation and triage

Several recent studies have sought to move beyond retrospective outcome scores and develop data-driven models to predict optimal time for initiating ECMO.

Zhu and colleagues used EHR data from the United States National COVID Cohort Collaborative (N3C) to build a hierarchical DL system called ”Prediction, Early Monitoring and Proactive Triage for ECMO” (PreEMPT-ECMO). ³ Their model integrates static demographic and comorbidity data with high-frequency physiological time-series. In the N3C cohort (n = 101,400 ICU admissions) 1.28% received ECMO, the system provided continuous risk estimates ≤96 h before ECMO initiation. Model interpretation analyses highlighted dynamic variation in feature importance during the patients’ clinical course. The authors argued that continuous early warning could aid triage decisions and transfer planning. The study was limited to the retrospective design, focusing on COVID-19–associated respiratory failure and lacked any (prospective) validation. Moreover, the N3C database captures national practice variability only (within the USA), and does not include features such as ventilatory waveforms or imaging.

In an earlier study, Xue et al. developed a gradient-boosting model (ForecastECMO) using data from 6,247 COVID-19 ICU patients across a multi-hospital system. ⁴ The model predicted ECMO initiation up to 48h before cannulation and was compared with the PaO2/FiO2 ratio, SOFA and PRESET scores and logistic regression models. In the development and holdout cohorts (ECMO prevalence 2.89 % and 1.73 %), ForecastECMO achieved the best discrimination; at an 18 h horizon the area under the receiver-operating curve (AUC) reached 0.94–0.95 with precision-recall values of 0.54-0.37. These results suggest a clinically actionable window for early referral. However, limitations include the retrospective single-centre design (albeit spanning 15 hospitals) and the lack of external validation. Furthermore, the model relied on aggregated input features (e.g., worst values) within prediction windows, potentially losing granular temporal trends. Finally, the model’s applicability is strictly bounded by its development cohort: COVID-19 patients aged 3–70 years with a BMI ≤45 kg/m². While these parameters establish a clear intended use-case, they naturally preclude generalization to broader patient populations.

A radiomics-enhanced approach by Mirus et al. combined computed tomography (CT) imaging with clinical variables to predict the need for V-V ECMO in 375 COVID-19 ARDS patients. ⁵ Using a convolutional neural network (CNN) for automated lung segmentation, the authors extracted 592 quantitative features. These were paired with clinical parameters (age, mean airway pressure, lactate, and C-reactive protein) to develop three logistic regression models (imaging-only, clinical-only, and combined). In the temporally separated validation cohort (n = 203), the combined model achieved the highest AUC (0.705), improving sensitivity compared with clinical or imaging models alone. This study illustrates the potential of AI-supported quantitative imaging for early risk stratification. However, the study was limited by its retrospective, single-center design without external validation, and reliance on a logistic regression classifier that may not capture complex non-linear feature interactions. Furthermore, manual correction of CNN-segmentations was still required, limiting full automation.

Collectively, these investigations demonstrate some evidence that ML models can facilitate data-driven decision making and thereby support triage decisions. To date, PreEMPT-ECMO represents the most comprehensive effort, leveraging large multi-centre data and continuous time-series analyses to provide early warning signals. ³ The ForecastECMO and the radiomics-enhanced models demonstrate that both traditional boosting algorithms and imaging-derived features can offer substantial discrimination.^4,5 However, the evidence remains nascent. All three studies are retrospective and largely limited to COVID-19 populations; neither without prospective validation. Single-centre designs and exclusion criteria can restrict generalisability, which is why clear use-case boundaries must be established for each proposed AI model. Further, important data types, for example, continuous ventilatory parameters, sequential echocardiography, or biomarkers are rarely included in model development. Therefore, while AI-guided prediction of ECMO indication is promising, further prospective, multicentre research across diverse indications is needed before these tools can be integrated into clinical practice.

Prognosis and outcome prediction models

The majority of AI research in ECMO focuses on predicting survival and complications, aiming to provide clinicians with more accurate prognosis than existing scoring systems based on traditional statistical methods.

Mortality and survival

Machine learning models have achieved high accuracy in predicting in-hospital and short-term mortality.^6–10 The most notable example is ECMO PAL a survival prediction tool for VA ECMO. ¹¹ ECMO PAL uses only pre-ECMO variables to produce a percentage likelihood to survive to hospital discharge. ECMO PAL was trained on a retrospective cohort of 18,167 patients (2017-2020) from 543 contributing hospitals within the Extracorporeal Life Support Organization (ELSO) registry. ¹² The developed model demonstrated good accuracy (75.5%) and AUROC (0.80) on a validation cohort of 5,015 new patients who were entered into the registry prospectively after model development (2021). Novel at the time, ECMO PAL included SHAP explainer theory to provide explanations about overall model behaviour as well as individual patient prediction explanations. ECMO PAL was released as an online tool at ecmo-pal.icu. This model was developed using a large international cohort, greatly improving generalisability over other AI ECMO tools, however it was not prospectively validated and was limited by the use of ELSO registry data, which only included patients who were already put on ECMO (no baseline comparison with non-ECMO patients).

Another prognostication tool is the RESCUE-24 score (Risk Evaluation Scoring for Critical Utilization of ECMO Within 24 Hours), ¹³ derived from a Random Survival Forest (RSF) model in a Taiwanese cohort (2025, N = 1,748). This tool achieved an AUC of 0.953 for predicting first-day and in-hospital mortality, significantly outperforming traditional regression models.

Complications (bleeding, thrombosis, neurological)

Complications remain a major determinant of ECMO outcomes, another focus of ECMO researchers using AI. Recent work utilizing the ELSO registry (N = 37,473) applied gradient boosting models (CatBoost, LightGBM) to predict acute brain injury, specifically intracranial hemorrhage and central nervous system (CNS) ischemia. ¹⁴ Reaching AUROC values around 0.70, demonstrates the significant challenge in accurately predicting neurological events. This is due to the inherent limitation of using variable-accuracy registry data alone, which could instead be supplemented with high-resolution physiological signals to meaningfully improve prediction accuracy.

The prediction of bleeding risk similarly represents a research priority. A multicenter cohort study from Japan (2025) utilized a large dataset (N = 470) and LightGBM to predict bleeding risk, employing red blood cell (RBC) transfusion as a surrogate marker. ¹⁵ The model achieved an AUROC of 0.70. Earlier prospective work from the USA (2020) explored predicting hemorrhage using Random Forests in a mixed ECMO cohort, reporting accuracies between 58–80%. ¹⁶ This moderate performance suggests that while AI can identify high-risk patients, current models likely miss important real-time coagulopathy indicators.

AI-enabled real-time control and decision support

This domain represents the frontier of autonomous or semi-autonomous ECMO, transitioning from passive prognostic modelling to the active modulation of the ECMO circuit components and integrated physiological management. Closed-loop control systems are currently in preclinical development. A 2024 non-AI study demonstrated a closed-loop automated control system capable of regulating hemodynamics in a canine model of cardiogenic shock supported by V-A ECMO combined with a left ventricular assist device (LVAD). ¹⁹ The system utilized a “circulatory equilibrium framework” to automatically adjust pump flow rates, fluid volume, and vasoactive agents (norepinephrine/nitroprusside) to maintain physiological targets. While standard VA-ECMO resulted in significant left ventricular distension with a median left atrial pressure (PLA) of 43.0 mmHg (IQR: 25.7–51.4), the automated system successfully unloaded the ventricle, reducing PLA to 12.5 mmHg (IQR: 12.0–13.4) within 60 min while maintaining a mean arterial pressure of 69 mmHg (target: 70 mmHg). However, the study was limited by the short duration of the automated control (60 min) and the use of an acute animal model, necessitating further validation before clinical translation.

In the realm of pulsatile support, a 2025 in vitro study utilized filter-type neural networks (f-NNs). Filter-type neural networks are specialized architectures designed to emulate biological reflexive arcs for high-fidelity, rapid real-time signal processing and can analyze blood pressure waveforms for counter-pulsation (CP) therapy. ²⁰ In a mock circulation loop setup, these networks successfully distinguished intrinsic heartbeats from ECMO pulses with an accuracy of over 87%. The system adjusted the pump’s pulsation timing (phase) to synchronize with the cardiac cycle, achieving a CP success rate of 78.62% even under heart rate variability, compared to only 25.75% without active control. This demonstrates that AI can process complex, overlapping waveform data with high fidelity to optimize hemodynamic support, although a limitation remains the processing delay of approximately 0.48 s. While full clinical digital twins for ECMO are not yet operational, concepts like these simulate patient-device interactions to test interventions in silico or in vitro before clinical application.

Weaning and readmission

In addition to survival and complications, milestone achievement during ECMO support is becoming an important research endpoint for outcome prediction. Deep learning models, such as Long Short-Term Memory (LSTM) networks, have been applied to continuous device data combined with EHR features to predict successful ECMO decannulation in a 118 single-center V-V ECMO patient cohort (USA, 2024). ¹⁷ Model validation was performed on synthetic data. Synthetic data is powerful in early-stages of model training (e.g., data augmentation), with the gold standard remaining an external clinical validation cohort of unseen data. While the model’s true performance lies in its ability to effectively stratify patients by decannulation risk, the modest performance (average AUROC 0.6937) is too low to integrate it into a clinical setting. Moreover, research is currently expanding to predict the successful readiness for discontinuing invasive mechanical ventilation (IMV) during V-V ECMO, an essential weaning step readily overlooked in outcome-focused models.

Machine learning is increasingly integrated in studies that examine the long-term sequelae and burden of ECMO support. One such analysis of the US Nationwide Readmissions Database (NRD) (2024, N = 22,947) predicted 90-days non-elective readmissions. ¹⁸ Of this cohort, 19.6% were readmitted, and the dataset was randomly split 80% for training and 20% for validation. The XGBoost model performed statistically significantly (p < 0.05) better than logistic regression (AUROC 0.64 vs 0.49). However, the model’s predictive power is not yet robust enough for high-confidence clinical application, as indicated by the modest AUROC value. This is likely due to the lack of incorporating granular time-series data (e.g., ECMO and IMV settings, imaging), which is not captured by the readmissions database, setting the stage for the next step in research. it represents an important step toward understanding post-discharge trajectories and resource utilization.

Market readiness versus research reality

Despite rapid methodological progress, the translation of AI into bedside ECMO practice remains limited; Figure 2 conceptualizes this ‘implementation gap’, spanning from data foundations and in-silico development to validation, regulation, adoption, and ultimate clinical reality. To date, there are no Food and Drug Administration (FDA)-approved, or CE-marked (European Conformity) AI tools indicated for autonomous ECMO control or standalone prognostication for CDSS in adult ECMO patients.

Figure 2.

The Implementation Gap in Adult ECMO AI. Visualization of the translational bottleneck. Despite a high volume of Retrospective Model Development entering the research pipeline (left), clinical translation is impeded by distinct filters. Progress is first stalled by the Validation Gap, characterized by a reliance on noisy, single-center datasets and a lack of standardized benchmarking. Subsequently, models face the Regulatory Wall, comprising stringent safety and liability requirements for high-risk software (e.g., EU AI Act, FDA SaMD). Finally, barriers in Knowledge & Adoption, such as the necessity for “human-in-the-loop” workflows, further restrict deployment. Consequently, no tools for autonomous control or complex prognostication currently reach Clinical Reality at the bedside (right).

While the FDA has granted premarket notification, known as 510(k) clearance, to modern ECMO consoles; however, these do not currently incorporate any AI features. ²¹ Tools such as ECMO PAL ¹¹ or the ECMO-SVC Score AI Tool are currently deployed as web-based educational or research calculators, labeled “not for clinical use”, therefore not falling under regulatory classification as a medical device.

Moreover, the translation gap remains substantial. Many models described in Section 4 are strictly in silico, yet to be integrated into the daily clinical workflow via corresponding systems (e.g., EHRs, bedside monitors, mechanical devices) for prospective validation and real-time decision support. A comparison of modelling approaches is given in Table 1.

Table 1.

A brief comparison of artificial intelligence (AI) models explored in detail in this review. Note: In the case of model name, where none is given, the first author is used for the etymology. The AI methodology column outlines the best performing model if multiple were reported. Accuracy and AUC columns correspond to the best reported values at the highest level of model validation for each study.

Model Name	Use Case	AI Methodology	Data Time Point	Sites Included	# Patients	Accuracy	AUC	Validation	Ref #
PreEMPT-ECMO	COVID-19 ECMO Eligibility	Cascaded LSTM	0-96 Hours	National Registry (USA)	101,400 (1,298 on ECMO)	N/A	0.94	None (Training Only)	3
Forecast ECMO	COVID-19 ECMO Eligibility	Gradient Boosted Trees	0-48 Hours	15 Hospitals across two states in USA	2,318 (Training)	N/A	0.94 (Testing)	Holdout Set	4
					3,929 (Testing)
ECMO PAL	VA-ECMO Outcome Prediction	Deep Neural Network	Immediately Pre-ECMO	International Registry (474 Sites)	18,167 (Training)	72.7% (Testing)	0.80 (Testing)	Temporal Validation Set	11
					5,015 (Testing)
RESCUE-24	VA-ECMO Outcome Prediction	Random Forest	Immediately Pre-ECMO	Single Centre	1,064 (Training and Testing)	N/A	0.80 (Cross-Validation)	None (Cross-Validation)	13
Kalra Score	ECMO-Associated Acute Brain Injury	Final Model Unclear (Tree-Based Approach)	Pre-ECMO up to 24 Hours	International Registry	37,473 (Training and Testing)	85% (LOOCV)	0.70 (LOOCV)	None (LOOCV)	14
Kamio Score	On-ECMO Bleeding Risk	LightGBM (Gradient Boosted Model)	ECMO Initiation	6 Hospitals (Japan)	357 (Training and Testing)	71% (Cross-Validation)	0.70 (Cross-Validation)	None (Cross-Validation)	15
Abbasi Score	ECMO-Associated Haemorrhage and Thrombosis	Random Forestr	Continuous	Single Centre	44 (Training and Testing)	64% (LOOCV)	N/A	None (LOOCV)	16
CEVVO	ECMO Weaning	LSTM	Continuous (Up to 25 Days)	Single Centre	118 (Training and Testing)	N/A	0.74 (Cross-Validation)	None (Cross-Validation Stratified by Outcome)	17
Balian Score	90-Day Readmission Prediction	XGBoost	At Hospital Discharge	National Registry (USA)	18, 358 (Training)	N/A	0.64 (Cross-Validation)	None (Cross-Validation)	18
Jang Controller	Continuous ECMO System Control	Filter-Type Neural Network	Continuous	None (In Vitro)	4,589 (Testing)	N/A	N/A	Benchtop Validation Using Mock Loop	20
					0 (In Vitro Study)

AUC – area under the receiver operating characteristic curve; LSTM – long- and short-term memory; ECMO – extracorporeal membrane oxygenation; LOOCV – leave one out cross-validation.

Regulatory, ethical, and implementation considerations

The transition of AI-guided tools from development to bedside deployment for adult ECMO is inhibited by the stringent requirements of the regulatory environment. AI systems (i.e., algorithm and application) that may affect ECMO initiation, parameter adjustment, or weaning will almost invariably fall into the high-risk medical or medical software category under the European Union AI Act, which imposes stringent requirements on risk management, data quality, transparency, and post-market surveillance in addition to existing medical device regulation. ²² In the United States, the FDA has concurrently adopted a full product lifecycle approach for AI- and ML-enabled software as Medical Devices (SaMD), evidenced by initiative such as the AI/ML SaMD Action Plan and recent updates formalizing Predetermined Change Control Plans (PCCPs) for adaptive algorithms.^23,24 As ECMO represents the most invasive life support for the most critically ill patients with virtually no margin for error, these regulatory frameworks translate into substantial evidentiary and documentation burdens. In some jurisdictions, burdens associated with obtaining clinical approval to deploy AI algorithms are nearing the same level of stringency as regulating the ECMO device itself (a class III medical device). This context partly explains why no AI systems are currently approved for autonomous ECMO control or complex, standalone prognostication in adults.^25,26

Beyond regulatory compliance, the ethical landscape is paramount in determining the acceptability of AI in ECMO practice. A central issue is generalisability: many critical-care AI models are trained on registry or single-center or limited region datasets, raising the risk of systematic under-performance in under-represented groups and amplifying existing disparities in access to high-resource therapies.^25–29 In high-stakes scenarios, such as ECMO triage or de-cannulation, even small performance differences between subgroups may have direct consequences on survival. Accordingly, ethical analyses of prognostic AI in intensive care underscore the need for explicit definition of target populations and usage by the AI developers, bias assessment, and ongoing performance monitoring and calibration after deployment.^27–29

Transparency and explainability are closely linked to fairness and trust. Historically, AI models have been labelled as “black box” models, due to the data-driven and nonlinear systems which makes them so powerful. In the clinical realm, this has led to concern about the accuracy, equity, and transparency of decisions offered up by AI systems. However, the reality is that explainer algorithms have been developed for all but the most sophisticated AI models (generative pre-trained transformers), with popular model-agnostic methods such as SHAP and LIME capable of providing powerful insights to the vast majority of proposed AI models for healthcare.^30,31 These explainable AI methods are capable of giving explanations about overall model performance (equivalent to β coefficients in logistic regression), as well as explanations about how individual patient predictions were made. Despite these advances made by the AI community in the last decade, the “black box” stigma continues to be perpetuated as a major roadblock to AI integration into healthcare systems.

Beyond the requirement for model transparency, it is essential that AI applications undergo robust prospective validation and meet rigorous regulatory and ethical standards. Successful integration further requires compatibility with existing clinical workflows to promote adoption by clinical staff while maintaining the continuity of patient treatment.^25,32

Responsibility and liability remain undefined. Conceptual, medical, and legal analyses converge on a human-in-the-loop paradigm: AI systems must support but not replace clinical judgement, ensuring intensivists retain final responsibility for ECMO-related decisions.^25,27–29 Meanwhile, unresolved liability remains a major concern in relevant commentaries, focusing on how responsibility will be allocated across clinicians, hospitals and manufacturers when AI-assisted decisions contribute to harm, especially given the challenge of continuously updated algorithms.^23,24,28 Closely related remain concerns regarding patient autonomy and informed consent: ICU patients and their surrogates are not typically informed that AI systems may guide clinical decisions about ECMO initiation, continuation or withdrawal, despite the potential impact of such tools.^27,28

Finally, successful implementation depends on trust and acceptance among the interdisciplinary ICU team. Current surveys suggest that most intensivists remain uncertain regarding the reliable use of AI as a decision-making tool, and that opaque models engender scepticism.^29,33 Conversely, over-reliance on algorithmic outputs risks inducing automation bias and the erosion of essential clinical skills. Therefore, ethical and implementation frameworks advocate for human-centred design, iterative user testing, education of ICU teams and robust external validation on realistic, noisy ICU data.^25,29,32 For ECMO specifically, where most AI models remain retrospective and non-clinical, these regulatory and ethical hurdles are likely to be decisive bottlenecks that determine if sophisticated algorithms will ever be trusted to guide this high-risk extracorporeal support at the bedside.^25,26

Knowledge gaps and other hurdles to adoption

Despite rapid progress in AI modelling, several hurdles still prevent these tools from implementation. ICU data is notoriously noisy, characterized by vital signs highly prone to artifacts and an inherent tendency toward error in human EHR documentation. Most current AI models are trained on cleaned retrospective datasets excluding the real chaotic live clinical data streams.

Another critical barrier is the limited standardisation and interoperability of ECMO data across institutions and device platforms. Most successful models still rely on static EHR variables, leaving the rich, high-frequency parameter signals from ECMO devices largely untapped because they are stored in heterogeneous, often proprietary formats, lacking standardised interfaces. Although generic data standards such as the OMOP (observational medical outcomes partnership) Common Data Model, HL7 FHIR (health level seven fast healthcare interoperability resource), and clinical terminologies (e.g. SNOMED CT – systemized nomenclature of medicine) already exist, they are not yet consistently implemented in critical care and ECMO datasets; consequently, centre-specific registries and vendor-specific data exports remain challenging to align and compare. For ECMO, domain-specific extensions have only recently started to emerge: Rieder et al. proposed the Extracorporeal Life Support Common Data Model (ECLS CDM), an open-source expansion of OMOP that encodes ECMO concepts and ELSO registry definitions in a harmonised, machine-readable structure. ³⁴ To successfully translate these safety concept into routine data capture, robust regulatory action will likely be essential to require that ECMO device manufacturers adopt common data models and provide standardised access to high-frequency device outputs. Large multicentre databases systematically capturing high-frequency ECMO signals are only beginning to emerge. Open-source initiatives (e.g., ECMO Reader) illustrate how shared, high-fidelity machine-data registries could form the backbone of a dedicated high-frequency ECMO register. ³⁵ This register would help to overcome current interoperability barriers and enable more realistic training, transportability assessment and the crucial validation of AI models in real-world practice.

Importantly, there is a significant prevalence of AI model development studies relying solely on internal validation techniques (e.g., leave-one-out cross-validation, LOO-CV). LOO-CV often focuses on training performance rather than true generalizability. Robust external validation is frequently missing. There is a notable dearth of prospective, randomized trials comparing AI-aided care versus standard of care. Without this level of evidence, clinical adoption will remain limited.

Conclusion

While integration of AI adult ECMO is rapidly evolving from identification of novel associations toward the development of robust predictive tools, clinical deployment has not yet reached the bedside. The technology remains fundamentally constrained by issues of market readiness, a highly dynamic regulatory environment, data interoperability, and a deficit of prospective validation among developed models.

Future research regarding AI in ECMO research should now move decisively beyond retrospective in silico accuracy. The transition requires the integration of multicenter, multiregional, high-frequency device data, and rigorous external validation to ensure safety, allowing AI to evolve from a theoretical promise to a practical CDSS in the ICU for ECMO support.