Smart Cardiac ICU: Digital Integration, Predictive Analytics, and Perioperative Inflammation

1. Introduction

Contemporary intensive is operating in an environment with high complexity cases, large volumes of information, and vast physiological, biological and therapeutical data. These data arise from multimodal monitoring, laboratory results, therapies for organ support - mechanical ventilation, continous renal replacement therapy, extracorporeal membrane oxygenation (ECMO). Real-time integration of this immense amount of data is beyond human capacity, resulting in an increase interest in artificial intelligence–based clinical decision support systems [1].

Conventional intensive care encounters several limitations. First, the overwhelming amount of data leading to decision-fatigue and delaying prompt therapeutic interventions. The burden of information overload is amplified by the lack of interoperability between the systems that generate those data, operating in parallel, representing a major drawback in critical situations where rapid decision-making is mandatory. Third, alarm fatigue adds to the burden of healthcare workers: 72-99% of alarms are either false, or clinically irrelevant [1,2,3]. Because of the abundance of clinical parameters, predictive scores, and continuous monitoring, intensive care units are considered key settings for implementing data-driven medicine.

The Smart ICU concept is showing promising results by integrating medical devices, information systems, Internet of Things technologies, and artificial intelligence into a unified system that allows real-time monitoring of a critical patient, enables risk stratification and prediction, offers decision support and potentially tele-ICU–based care [4,5,6]. However, routine clinical implementation remains limited.

A recent systematic review of 1263 studies reported that 74% of artificial intelligence (AI) models are at a technology readiness level (TRL) of 4 or below, and only 2% is feasible for clinical integration (TRL ≥ 6). The field is experiencing an annual growth rate of 39% in publication volume, but dominated by retrospective studies [7].

Cardiac surgery provides a suitable model for the Smart ICU. It combines the surgical intervention with hemodynamic, perfusion, and biological and inflammatory data, leading to measurable postoperative outcomes [8].

The aim of this narrative review is to present current data of Smart Cardiac ICU with a focus on digital integration of perioperative data, predictive analytics for decision-making support and monitorization of postoperative inflammatory response to cardiopulmonary bypass (CPB). We describe the digital architecture of Smart-ICU, related concepts, AI-tools used for prediction, and discuss current limitations and future directions of this concept. A dedicated chapter on the post-CPB inflammatory response emphasizes the feasibility of AI tools integrated in clinical practice in such a complex field, and how cardiac ICUs represent a key setting for data-driven critical care.

2. Methodology

This article is a narrative review of the Smart ICU concept, focusing on the cardiac ICU and aiming to present current evidence and research future directions. We performed a database search on PubMed/MEDLINE, Web of Science, for articles published between 2015 to 2026, using the following key-words: „intensive care unit”, „Smart ICU”, „intelligent ICU”, „clinical decision support”, „patient data management system”, „alarm fatigue”, „Internet of Things”, „tele-ICU”, „tele-critical care”, „machine learning”, „artificial intelligence”, „cardiac surgery”, „cardiopulmonary bypass”, „systemic inflammatory response”.

We included original articles, observational studies, randomized trials, systematic reviews, meta-analyses and consensus documents published in English or Romanian, with relevance to at least one of the central themes of the review: Smart ICU architecture, IoT and alarm management, artificial intelligence in ICU, tele-ICU and perioperative inflammation in cardiac surgery. We excluded articles focused on non-critical fields and reports with no clinical implementation of the smart-ICU concept.

We chose a narrative rather than a systemic design because the aim of this review is conceptual - to collect evidence from clinical integration of this concept, highlight the knowledge gaps from the current literature. Therefore, we did not follow the PRISMA methodology and did not perform a quantitative synthesis.

3.1. The Smart ICU concept and its application in the Cardiac ICU

In the last six decades, continuous monitoring and specific organ support therapies have evolved, and conventional intensive care units are now relying on numerous sources that generate clinical and paraclinical data. The major disadvantage is represented by the lack of interoperability between monitors, ventilators, infusion pumps and laboratory findings [9]. Therefore, the clinician remains the „brain” behind this heterogenous information, high amounts of information creating cognitive burden and increasing the risk of errors [10]. Evolving digital technologies – electronic records, computerized prescribing systems, tele-ICU platforms, are focusing on limiting this issue, to reduce diagnostic time and reduce the information overload encounter by clinicians [11,12].

The Smart ICU represents a paradigm shift from a fragmented, reactive model of care to an integrated and predictive ecosystem. The conventional intensive care is evolving through incorporation of communication technologies, Internet of Things (IoT) networks, artificial intelligence, robotics, and big data analytics, with the purpose of improving real-time monitorization, enable risk stratification and prediction and improve diagnostic accuracy. Given the scale and complexity of modern critical care systems, digital integration and standardization allows monitorization and analysis as an interconnected system, linked through IoT and high-speed network infrastructures [5].

Several related terms are often used interchangeably in the literature. Table 1 distinguishes them to clarify the scope of this review.

Unlike conventional ICUs, smart ICUs integrates and analyses a large volume of data enabling prediction and decision support. AI models with AUC frequently >0.80 are capable of early warning and decision support, but only in the presence of robust integration with EHR; calibration and external validation of these models remain sub optimally reported [13]. Specifically, AI models dedicated to sepsis prediction can anticipate onset 6–12 hours before conventional clinical diagnosis, exceeding in accuracy established scores such as SOFA and qSOFA, and the prediction of acute kidney injury 12–48 hours in advance, intraoperative hypotension 15 minutes in advance, and the optimization of ventilator weaning — with a reduction in the duration of mechanical ventilation of about 21 hours — concretely illustrate the transformative potential of AI integration into clinical flow [14].

The Smart ICU requires three interdependent dimensions – infrastructural, including medical devices, high-speed networks and digital platforms; analytical dimension, including machine learning (ML), Internet of Things (IoT) and predictive algorithms, and the human-organizational dimension, that implies staff training and focuses on algorithm-assisted decision-making [5,10,11,12,13,14,15,16].

Digital integration of data carries significant benefits. Recent data from the literature shows that implementation of a patient data management system (PDMS) is associated with a reduction in ICU length of stay, lower mortality rates and lower rates of nosocomial infections, including elimination of Acinetobacter spp. infections [11]. Electronic prescribing has been proven to lead to earlier de-escalation of antimicrobial therapy and a reduction in antimicrobial resistance [11,17], and among other benefits of digitalization, it has been shown that by reducing the documentation errors and optimizing the resource allocation, smart-ICU enhances patient outcomes and clinical workflow. Telemedicine has become a component of the Smart ICU concept, developed and accelerated by the COVID-19 pandemic, where telerounds, telemonitoring and teleconsultations became a necessity [18].

Cardiac surgery can be a key setting for implementing these tools in clinical practice. Starting from the surgical procedure as a well-defined procedural event, with a defined inflammatory trigger – cardiopulmonary bypass, dynamic monitoring of postoperative data may predict postoperative outcomes. These features are underlie the feasibility of cardiac ICU to be used as a demonstrative model.

3.2. Digital Integration and Interoperability

The Smart ICU uses a layered digital architecture, that combines signals into clinical decisions.

The data acquisition layer, collecting data from specific monitoring systems – monitors, infusion pumps, ventilators, ECMO systems, point-of-care analyzers, and enables synchronization of EHR data in order to obtain information about trends and changes in the clinical course of the patient [19,20].

Interoperability layer, based on specific universal systems, like HL7, DICOM and FHIR that allows synchronization of collected data [21,22].

Data storage layer, that allow storage and transmission of patient information, including Patient Data Management Systems (PDMS) that improves data completeness and enable trend analysis over time [11,22].

Analytic layer, acting as the Smart ICU’s “brain” and integrating clinical decision support systems (CDSS), AI algorithms, and predictive models, in order to optimize the decision-making process by anticipating specific complications, such as acute kidney injury and delirium [19,23,24,25].

Clinical interface layer, that converts data to specific information targeting specific healthcare workers – physicians, nurses and pharmacists, while a specific command center provides guidance and coordination, leading to a better adherence to clinical protocol that are currently in use [25,26].

Governance and cybersecurity layer, managing access and security policies and limiting the risk of bias, ensuring ethical an legal requirements.

A recently published Romanian study concluded that automated PDMS documentation was associated with a reduction in length of stay in the hospital and improved mortality rates [11]. AI tools using multimodal data have shown better outcomes compared to conventional triage tools and predict specific infection – surgical site infections and urinary tract infections, but validation of these models and feasibility of implementing them into the clinical practice still remains a challenge [27,28,29].

Multimodal predictive models have shown better discrimination than conventional triage tools for specific infections, such as surgical site and urinary tract infections, although validation and clinical feasibility remain a challenge [27,28,29].

Table 2 summarizes the main data sources of the Smart Cardiac ICU and their potential predictive value.

3.3. Predictive Analytics and Clinical Decision Support

Predictive analytics can support clinical decision, making digitalization reshape clinical care and ICU a leading field. Continuous monitoring, laboratory data, and therapeutic interventions generate a large volume of data, exceeding human cognitive capacity [30,31]. Machine learning (ML) and deep learning can recognize patterns and predict severity scores and early deterioration. Models such as RNN-LSTM (Recurrent Neural Networks – Long Short-Term Memory) can collect data in a dynamic matter, decreasing the use of static scores and intermittent assessment [32,33]. AI applications in ICU care can be organized by outcome.

Concerning mortality, a systematic review compared scores commonly used in the ICUs – APACHE, SOFA and SAPS II scores, with XGBoost and Random Forest – models of machine learning algorithms, using data from the first 24 hours after admission. ML models showed better results compared with conventional scoring in predicting sepsis mortality and length of stay in the ICU [34,35,36].

Concerning length of stay - a meta-analysis of 33 studies reported a pooled AUROC of 0.9005 for the prediction of ICU or hospital length of stay [36].

Assessments of perioperative risk using AI-tools also shown promising results, by predicting severe complications, outperforming conventional risk scores. The MySurgeryRisk and POTTER systems have shown promising results regarding morbidity and mortality in over 380.000 patients [37,38]. In a large cohort of over 220000 patients, XGBoost achieved better results compared to EuroSCORE II for in-hospital mortality [39].

Concerning ventilation and pulmonary complication, boosting models may detect postoperative pneumonia earlier, and may support the need for mechanical ventilation or the weaning off the ventilator. ML may also predict prolonged ventilation in cardiac ICUs [40].

Among other syndromes encountered in the postoperative period, acute kidney injury can be detected earlier with the use of multimodal models acquiring preoperative and intraoperative data [38].

Predictive analytics using neuromonitoring and specific biomarkers can be used for postoperative delirium, by detecting the earliest alterations preceding cognitive dysfunction [41].

Perioperative inflammation will be discuses in the following sections.

Challenges encountered by ICU predictive models are the result of incomplete and irregular data. Model performance depends on preprocess capacity, missing values and clear definitions of the intended outcome [42,43]. Explainable AI tools such as SHAP and LIME, used to interpret the predictions of machine learning models, can improve the predictive accuracy [43,44].

However, AI implementation into clinical practice remains limited due to the bias risk (50%), data heterogeneity and ethical concerns. To overcome these challenges, smart ICUs require standardized data, with immediate availability to improve its effectiveness, and explainable AI frameworks [7,27,45,46]. In this context, cardiac surgery provides an ideal setting to integrate physiological, procedural, and immunological data into bedside predictive models.

3.4. Alarm Management, IoT, and Command-Centre Models

The Internet of Things (IoT) and the Internet of Medical Things (IoMT) are the technological foundation of the Smart ICU, collecting continuous data from different medical devices. One IoT architecture was design capable to collect and process approximately 22 TB of medical data per year [26].

Alarm overload remains a significant problem in ICUs, generating up to 152.5 alarms/bed/day, 72–99% of which are false positives [48,49,50], taking up to 35% of nursing time and therefore affecting clinical performance [51,52,53]. AI-based solutions implies filtering strategies and mobile notifications that may improve response times and reduce alarm fatigue [2,50,54]. Wearable smartwatches, by prioritizing alarms according to the urgency, are reported to reduce unnecessary notifications, increase response rates and support the concept of „Silent ICU” – replacing bedside alarms with targeted alerts [2,50,51,58].

Command-center models extend these functions by centralising monitoring and coordination, which may extend the access to critical care experts and improve adherence to protocols [26].

3.5. Tele-ICU and Distributed Cardiac Critical Care

Telemedicine in the intensive care unit (Tele-ICU) enables real-time remote monitoring and clinical support for patients in separated ICUs. Hub-and-spoke architecture allows a central hub to provide support and clinical guidance over smaller units. Other models have been developed for hybrid approaches and consultative interventions, showing promising results in different healthcare environments.

Recent literature emphasizes that Tele-ICU should be considered a system adaptable to local resources and organizational structures [56]. It may be particularly helpful in units with limited access to intensive care physicians, where remote support can improve the decision-making process and use of resource, facilitating treatment plans and shortening diagnostic times, with evidence showing benefits on critical care capacity, but no impact on reducing ICU transfers [57].

The TELESCOPE trial, the largest randomized trial evaluating Tele-ICU, tested the effectiveness of daily multidisciplinary rounds conducted by an intensivist through telemedicine in 30 ICUs, reporting no significant decrease in ICU length of stay and hospital mortality [58]. The result was attributed to the operational model and implementation protocol, rather than to a lack of value of tele-critical care in general [58]. Another systematic review including 26 studies reported low mortality rates reduction in ICU settings, highlighting the heterogeneity of data and lack of standardized methodological an clinical protocols. Future studies should compare different Tele-ICU models to determine which approaches are most effective in specific critical care settings [59].

Nevertheless, Tele-ICU has the potential of becoming a connectivity platform, once certain needs are met – international standards in definitions, specific protocols regarding training and implementation of tele-critical care models [60]. In cardiac ICU settings, it has been demonstrated that preoperative echocardiographic parameters are independent predictors of the need for inotropic and length of stay [61]. These findings emphasizes the need for structured data collection in order to proper integrate Smart-ICU systems in clinical practice, and obtain valuable information to support clinical decisions.

A recent systematic review found that hub-and-spoke and hybrid models provided the most consistent clinical and staffing benefits, tele-ICU nurses playing a major role in enhancing care quality [62]. The COVID-19 pandemic accelerated the need and implementation of tele-ICU technologies, emphasizing the benefits of telemonitoring, teleconsultation and leading to a major shift in infrastructure designs [63].

Overall, Tele-ICU has significant potential to expand access to expertise and standardize care, but challenges are still encountered mostly because of financial and infrastructure barriers [64].

3.6. Perioperative Inflammation after Cardiac Surgery

Cardiac surgery is a field with particular pathophysiological mechanisms, with perioperative inflammation secondary to cardiopulmonary bypass (CPB) being the main trigger of complications that affect patient course and outcome. Despite advances in understanding and managing the systemic inflammatory response to CPB, it remains a challenging, multifactorial reaction associated with significant complications, from mild organ disfunctions to severe multiple organ dysfunction syndrome (MODS), and contributing to an increase in mortality in cardiac ICU settings.

The inflammatory response is triggered by surgical trauma, ischemia-reperfusion injury, blood contact with the CPB circuit, coagulopathy and microcirculatory disfunctions. The release of inflammatory cytokines lead to an increase in vascular permeability, glycocalyx degradation and vasoplegia. Depending on the time of CPB and ischemia time, it may lead to organ injury requiring specific organ-support therapies [65].

3.6.1. Inflammatory Biomarkers

Inflammatory biomarkers have emerged as valuable tools for risk stratification in cardiac surgery. Elevated preoperative C-reactive protein (CRP) levels are associated with an increase in major cardiac events and all-cause mortality in cardiac ICUs, and elevated levels in the postoperative period have been shown to be an independent predictive factor for mortality [66]. Among the inflammatory markers, IL-6, IL-10, IFN- γ, and IL-2 are associated with prolonged mechanical ventilation, acute kidney injury, and 30-day mortality [67].These findings support a dynamic, multimarket approach to the assessment of postoperative inflammation.

Neutrophils and markers of their activation have gained particular interest as early indicators of postoperative inflammation. A 2025 study reported that neutrophil activation occurs intraoperatively during CPB period and can be used as an early marker of inflammatory response, preceding the elevation of CRP and IL-6 levels. The transient immune dysfunction caused by CPB and present in the postoperative period have been shown to contribute to the postoperative coagulopathy and infectious complications [68].

Complete blood count–derived inflammatory indices, particularly the neutrophil-to-lymphocyte ratio (NLR) and systemic immune-inflammation index (SII), have shown prognostic value in cardiac surgery, specifically for postoperative acute kidney injury, prolonged mechanical ventilation and atrial fibrillation. As independent risk factors for these complications, elevated levels of NLR and SII are associated with longer ICU-stays and an increase in mortality rates [69,70,71].

3.6.2. Relevance to the Smart Cardiac ICU

Rather than evaluating specific biomarkers in a static manner with isolated measurements, Smart ICU systems can analyze the trends in values and combine them with hemodynamic and other patient-related relevant data, supporting the concept of dynamic inflammatory profile. The interpretation of specific markers of inflammation require correlations with multiple data – procedure type, CPB and ischemia time, vasopressor requirement, patient`s comorbidities and hemodynamic status. Artificial intelligence can combine these data and generate predictive scores to identify patients at risk for CPB and inflammatory-related complications [65].

In this context, CPB-induced inflammation can be incorporated into a “smart immune-monitoring” framework that can lead to earlier identification of high-risk patients and influence the perioperative care, therefore enhancing the treatment plans and patients outcomes.

Since current biomarkers alone have limited predictive accuracy, the development of multimodal predictive models represents a promising application of Smart ICU technologies [66,67,68].

4. Smart Monitoring of the Inflammatory Response

The heterogeneous nature of the inflammatory response induced by extracorporeal circulation has encouraged the use of artificial intelligence and machine learning (ML) for risk stratification and prediction of postoperative complications. ML algorithms are able to integrate data generated from different devices (pump infusions, ventilators, monitors, invasive monitoring systems) simultaneously, generating dynamic real-time information that can influence clinical decision-making. The capacity of ML to identify complex patterns supports the concept of AI-based patient phenotyping, increasing the capacity of early detection of patients at risk for postoperative adverse outcomes.

Early applications of artificial intelligence in cardiac surgery focused on predicting postoperative systemic inflammatory response syndrome (SIRS). ML models like Random Forest and SHAP have shown good results in predicting and identifying modifiable factors such as anemia and elevated lactate levels as independent risk factors for SIRS [72] NLR and CRP have been reported as independent predictive factors for increased mortality rates associated with severe inflammatory response, and ML models have reportedly predict higher rates of myocardial dysfunction in patients with higher levels of NLR and CRP [73].

A major advantage of machine learning is its ability to identify distinct inflammatory phenotypes after extracorporeal circulation (ECC). Recent data is stratifying patients into different subgroups with specific biomarkers profiles related to specific clinical outcomes - α phenotype - older patients with multiple comorbidities, longer hospital stays, higher mortality; and β and γ phenotypes – more favorable outcomes [74]. This major breakdown in ML findings is confirming that inflammatory response to ECC is heterogenous, therefore suggesting that AI can improve risk stratification from dividing and phenotyping subgroups of patients according to their physiological response to inflammation.

Multimodal phenotyping that integrates intraoperative data has further expanded AI applications. In a large cohort, ML divided distinct phenotypes of patients with a severe impairment in metabolic and immunological response, hemodynamic instability and higher vasopressor requirements and higher rates of organ disfunctions post-CPB [75].

A further application of AI involves real-time integration of extracorporeal circulation data by monitoring different parameters – perfusion data, flows, pressures, oxygen delivery and consumption, hemoglobin and hematocrit levels, lactate levels, allowing risk assessment for the intraoperative and early postoperative period, with extending these applications to other mechanical circulatory support devices [76].

However, current models remain limited by retrospective validation (most performant models - XGBoost AUC > 0,92, GBM AUC > 0,85 validated on retrospective cohorts), variable standardization, and the lack of prospective real-time integration of inflammatory biomarkers with CEC data. Further multicenter validation and interoperable platforms are required before AI becomes a routine component of Smart ICU systems [72,73,74,75,76,77].

Figure 1 illustrates the proposed Smart Cardiac ICU data flow, in which perioperative and procedural data, immunomonitoring, predictive algorithms, and clinical decision support systems are integrated within a single digital infrastructure.

Table 3 contrasts conventional ICU tools with Smart Cardiac ICU analytics.

5. Implementation Barriers, Governance, and Future Directions

5.1. From concept to clinical practice

Implementation of a concept into clinical practice requires several conditions: standardization of data, external validation, model explainability, efficient training of healthcare workers, cybersecurity and cost efficiency. Table 4 summarizes the principal barriers currently encountered, and required solutions..

5.2. Governance and Data Ethics

Governance must address confidentiality concerns, risk of bias and traceability of decisions. It also requires transparent and sustainable development, accountability for the treating physician and monitoring of the results followed by implementation.

AI tools will not replace clinical judgment, therefore responsibility remains in the hands of the clinicians. Continuous monitoring of performance after implementing Smart-ICU designs will be required [79]. Future research should focus on implementation studies and explore and monitor patient safety and system efficiency after implementation in clinical practice of AI-tools [80].

5.3. Future Directions

The future of Smart ICU lies in the transition from predictive to actionable systems that can be integrated into clinical practice. Current technologies are focusing on AI-tools that address mechanical ventilation – measuring ventilatory parameters and predicting mechanical ventilation associated adverse effects, alarm prioritization – reducing the alarm fatigue and improving the reaction-time to emergency situations; and postoperative risk assessment – scoring systems and predictive factors that can influence the postoperative outcome of patients. The FIRST-ICU model, combining graph neural networks and LSTM that can process data sequentially, plays a crucial role for the future direction of AI into clinical practice, supporting AI-augmentation of decision-making process [81]

A key future direction is prospective, multicenter validation and standardization of AI assessment tools. Current literature is lacking prospective studies and is limited by the lack of evidence regarding clinical impact of AI implementation in ICUs. A meta-analysis of 54 ML models for predicting postoperative cardiac complications in 927,113 patients found better results over traditional scores, but is limited by inconsistency in calibration methods and clinical integration [82].

Risk stratification and predictive scores should involve multimodal data integration, combining several parameters – physiological monitorization, laboratory data and biomarkers and imaging. Hamza et al. proposed a neuroprotection framework in cardiac surgery integrating imaging, inflammatory and metabolic biomarkers, neuromonitoring (TCD, NIRS, EEG/BIS), and ML models for predicting postoperative delirium and cognitive dysfunction [41]. In cardiac surgical ICU, implementing ML models would allow a better understanding in inflammatory trends and complications, enabling “smart immune-monitoring”. Early identification of inflammatory phenotypes could support personalized perioperative immunomodulation strategies in the future.

Lastly, the future of Smart ICU will also be shaped by an interdisciplinary collaboration between clinicians, informaticians, statisticians, in order to achieve safe, transparent and sustainable Smart ICU development and implementation into clinical practice [79].

6. Limitations

This narrative review has several limitations. First, the literature selection was not exhaustive; rather than a predefined, reproducible search protocol, studies were identified and included on the basis of their conceptual relevance to the Smart Cardiac ICU framework, which introduces the possibility of selection bias and means that some pertinent reports may not have been captured. Second, no formal assessment of risk of bias or methodological quality was performed for the included studies, so the individual findings summarised here should be interpreted with caution and are not weighted according to their evidentiary strength. Third, the objective of this review is deliberately conceptual and translational rather than confirmatory: it aims to integrate heterogeneous evidence into a coherent architectural and mechanistic model and to identify translational opportunities for perioperative inflammation monitoring, not to provide pooled effect estimates or definitive clinical recommendations. Accordingly, the conclusions should be regarded as hypothesis-generating and require prospective, multicentre validation before clinical implementation.

7. Conclusions

The Smart Cardiac ICU has the potential to transform intensive care units from device-centered units into an integrated ecosystem, combining continuous monitoring, digital integration, predictive analytics, decision support, and tele-medicine. While the benefits are promising – early detection of specific changes in clinical behaviors and alterations, enhanced workflow and patient-centered care, real-time data analyzing and prediction of critical conditions; they require clinical validation by prospective future research.

This review links the Smart Cardiac ICU concept to perioperative inflammation in cardiac surgery with extracorporeal circulation and AI-based inflammatory phenotyping. We propose the cardiac surgical ICU as a model for data-driven monitoring of inflammatory and immunological changes, integrating the kinetics of specific inflammatory biomarkers among procedural data, to support the decision-making process. Post-CPB inflammatory syndrome may become a demonstrative case for data-driven critical care, and offer future directions for Smart-ICU models.