Defining the Smart ICU: Aligning Innovation With What Matters Most

As intensive care units worldwide face mounting pressure from staffing shortages, technological complexity, and the need for improved patient outcomes, Mindray convened a roundtable of leading critical care experts to envision the future of the smart ICU.

With ICUs around the world under growing strain from workforce shortages, expanding technological demands, and the constant pursuit of better patient outcomes, leading critical care experts gathered for a roundtable discussion to reimagine the future of the smart ICU, where advanced data integration, intelligent automation, and human-centred design converge to transform care delivery.

Defining the Smart ICU

The discussion opened with a fundamental question: What constitutes a smart ICU? The group communicated their vision of a smart ICU as an environment enhanced by advanced digital technologies, including artificial intelligence, the Internet of Things, big data analytics, and automation, designed to improve patient monitoring, clinical decision-making, and workflow efficiency.

The participants emphasised that technology alone doesn't define smartness. Prof Jean-Daniel Chiche stated that a smart ICU must be data-driven and patient-centric and must ensure that the human element remains central to care. He emphasised that modern ICUs should be easier to use, not more complicated, comparing the ideal to how today's cars are more technologically advanced yet simpler to drive.

The concept of personalisation emerged as a critical component. The smart ICU should help detect patient deterioration earlier, personalise care, and optimise resource use, while also functioning as a self-learning environment that embeds continuous education into the system.

Dr Adrian Wong compared ICU technology adoption to smart home systems. He observed the contrast between the seamless integration of technology in personal life, where devices communicate effortlessly and anticipate needs, and the fragmented, cumbersome systems in critical care settings. According to him, this disconnect highlights a significant opportunity: if technology can make homes smarter and more efficient, why can't it do the same for intensive care units?

Prof Chiche emphasised that workflow efficiency isn't just about operational metrics; it's about caregiver wellbeing. Healthcare professionals spend excessive time on administrative tasks, data entry, and navigating disconnected systems, leaving less time for direct patient care and contributing to burnout. A truly smart ICU should liberate clinicians from these burdens, allowing them to focus on what they do best: caring for critically ill patients and their families.

Three Priority Challenges for the Smart ICU

The group identified three pressing challenges that must be addressed to realise the vision of a smart ICU:

1. Interconnectivity and Interoperability

Prof Paul Elbers pointed out the need for seamless integration of all devices and systems. Currently, ICUs operate with equipment from multiple manufacturers. They have ventilators from one company, monitors from another, infusion pumps from yet another, but with limited ability to share data effectively. This fragmentation prevents the analysis necessary for advanced clinical decision support.

Prof Chiche and Prof Elbers emphasised that interconnectivity must be bidirectional. It's not enough for devices to send data to an electronic health record (EHR); the EHR must also inform device settings and recommendations. This two-way flow of information is essential for developing intelligent algorithms that can guide clinical decisions.

Dr Wong drew a comparison to Apple's ecosystem, where devices communicate seamlessly within a closed system. While some worried about vendor lock-in, others argued that companies with comprehensive portfolios could create validated, integrated solutions more effectively than fragmented multi-vendor approaches.

2. Predictive Analytics and Early Detection

The ability to predict patient deterioration before it becomes clinically obvious represents a crucial frontier. The group noted that although research has demonstrated the feasibility of predictive algorithms for complications such as hypotension, respiratory failure, and patient-ventilator asynchrony, these tools remain largely absent from clinical practice.

Prof Jean-Michel Constantin spoke about how modern cars can alert drivers to potential mechanical problems before they cause breakdowns. Similarly, ICU monitors should evolve beyond alarming only when a patient is already in crisis to providing early warnings that allow preventive interventions. This shift from reactive to proactive care could fundamentally change outcomes.

However, participants acknowledged that prediction alone isn't sufficient. These systems must be validated, trusted by clinicians, and integrated into clinical workflows in ways that enhance rather than complicate decision-making.

Several participants emphasised that advanced analytics and AI depend fundamentally on high-quality data. Without consistent data acquisition from units with standardised practices and properly annotated events, developing reliable algorithms remains difficult.

Common data quality issues, like arterial line transducers at incorrect levels or inconsistent sensor placement, introduce noise that undermines algorithm development. Therefore, companies investing in AI should also invest in data quality improvement tools, partnerships with wellrun ICUs for data collection, AI-assisted labelling to improve dataset quality and standardised data annotation protocols.

3. Clinical Decision Support Systems

The third priority focused on transforming data into actionable insights through intelligent decision support. Dr Wong emphasised that such systems should synthesise information from monitors, ventilators, infusion pumps, laboratory results, and imaging studies to provide evidence-based recommendations for treatment.

Prof Francesca Rubulotta pointed out that decision support should function as a co-pilot to create a system. Prof Elbers agreed and said the system should work as "a team of professors in your pocket"— expert advisors who suggest possibilities and explain their reasoning but ultimately leave decisions in the hands of the clinical team.

The participants discussed automated treatment systems, particularly for sedation and haemodynamic management. Prof Rubulotta described a robotic anaesthetic system that integrates physiological monitoring with drug infusion, adjusting medications based on patient response, essentially a co-pilot for routine management.

Participants emphasised that automation should handle routine adjustments while clinicians maintain oversight and make critical decisions. For example, a decision support system might recommend a fluid challenge based on integrated haemodynamic data, but the clinician reviews the suggestion and decides whether to proceed.

Prof Elbers highlighted the concept of digital twins. These computational models simulate individual patients, allowing clinicians to test treatment options virtually before implementation. "If I give treatment A, here's what will happen in 15 minutes and one hour. Treatment B produces these different outcomes." This counterfactual reasoning preserves clinical autonomy while providing evidence-based guidance.

Defining Innovations That Matter Most

Making Healthcare Workers Happy

From a practical adoption standpoint, Prof Vincent questioned whether caregivers would be happier with a smart ICU. Prof Chiche suggested that the most impactful investments might be those that directly improve caregiver satisfaction, even if they don't immediately show measurable effects on patient outcomes. Prof Elbers pointed out that reducing administrative burdens and information overload could make the profession more attractive, addressing the critical shortage of ICU nurses while improving the quality of work life for all staff.

The advent of large language models (LLMs) promises to revolutionise clinical documentation. These systems can listen to bedside discussions, extract key information, and generate clinical notes, freeing physicians and nurses from hours of computer work that currently keeps them away from patients.

Prof Chiche raised an important point about the current state of affairs, in which doctors cluster around computers rather than patients, a phenomenon his nurses call "computer medicine." Intelligent EHR systems could reverse this trend, allowing clinicians to focus on observation, examination, and communication.

Leveraging Comprehensive Device Portfolios

For companies with broad product portfolios spanning monitors, ventilators, ultrasound, and infusion systems, Prof Chiche highlighted that the opportunity lies in developing algorithms that leverage device-to-device connectivity. Combining data from ventilators and haemodynamic monitors could provide real-time physiological insights that neither device could offer alone.

One compelling example discussed was integrating patient-ventilator interaction analysis (already validated in research) into clinical monitors. Despite solid evidence that computers can reliably detect asynchrony (which clinicians often miss), this technology hasn't been widely implemented.

Drawing on the Apple App Store analogy, a company with a comprehensive device portfolio could ensure seamless interconnectivity within its ecosystem, create a marketplace for validated algorithms and tools, and establish data standards that facilitate algorithm development.

This approach could accelerate innovation by allowing researchers and smaller companies to develop specialised applications while leveraging established infrastructure. However, it requires the platform provider to balance openness (encouraging innovation) with control (ensuring safety and quality).

Continuous Monitoring and Integrating Clinical Guidelines into Smart Systems

Prof Constantin steered the discussion on precise and innovative monitoring approaches:

Continuous Echocardiography: The group envisioned wearable or patch-based ultrasound devices that could provide continuous cardiac imaging with automated analysis. As ultrasound technology becomes smaller and more sophisticated, continuous visualisation of cardiac function, combined with other physiological data, could revolutionise haemodynamic management.

Wireless Monitoring: Dr Wong expressed frustration with the tangle of wires connecting patients to monitoring equipment, which complicates mobilisation, transport, and basic care. While acknowledging regulatory and technical challenges, particularly regarding signal interference and patient safety, Prof Elbers argued that wireless monitoring should be standard rather than aspirational. The technology exists; what's needed is industry commitment and regulatory frameworks that enable safe implementation.

Computer Vision and Movement Tracking: Cameras and sensors that analyse patient and caregiver movements show promise for detecting subtle signs of deterioration. Prof Chiche and Dr Wong both discussed the application of this technology. Research has demonstrated that increased caregiver activity around a bed, even when vital signs appear stable, often precedes adverse events, capturing the experienced nurse's intuition that "something isn't right."

Prof Constantin proposed integrating clinical guidelines and emergency checklists directly into monitoring systems. Drawing inspiration from aviation, where pilots access emergency checklists on their screens during critical situations, the group discussed whether similar functionality would benefit ICU clinicians.

Dr Wong elaborated on the concept, envisioning monitors that could display relevant protocols based on detected abnormalities. When a patient goes into cardiac arrest or anaphylactic shock, the system would automatically present the appropriate emergency checklist alongside vital signs, ensuring clinicians don't overlook critical steps during high-stress situations.

However, this seemingly straightforward idea sparked important debate. Prof Rubulotta questioned whether protocoldriven care contradicts the personalised medicine philosophy that had dominated earlier discussions. If the smart ICU's purpose is to individualise treatment based on each patient's unique physiology, does standardised protocol guidance undermine that goal?

The resolution came through recognising context-dependence. For truly emergent situations, such as cardiac arrest, where evidence-based protocols apply universally, standardised checklists add value. But for complex clinical decisions that require nuanced judgment, the system should provide personalised recommendations based on each patient's data rather than generic protocols. Hence, a smart ICU should know when to apply standardised guidelines and when to offer personalised decision support.

Integrating a Mobile App into the Smart ICU Ecosystem: Prof Rubulotta discussed the potential of a mobile application that could serve as a seamless extension of the Smart ICU ecosystem, enabling clinicians to access the same real-time data environment available at the bedside. By continuously drawing information from patient monitors, ventilators, and electronic health records, the app could deliver personalised alerts, highlight deviations in clinical trends, and present key data directly on a clinician’s device. During rounds, staff could use the app to enter structured notes, capture voice-to-text observations, and tag clinical events, with all inputs synchronising instantly to the central ICU platform.

Because the application would interface directly with the Smart ICU’s big-data architecture, it could also provide predictive risk scores, early-warning indicators, and decision-support recommendations generated from continuously updated analytics. In doing so, the app becomes both a mobile command centre and a streamlined documentation tool—integrating bedside assessments, objective data, and algorithm-driven insights into a unified workflow that enhances coordination, accuracy, and situational awareness across the care team.

The app could additionally function as a rapid-response communication platform, allowing clinicians to activate teams such as extracorporeal membrane oxygenation (ECMO), medical emergency teams (MET), cardiac arrest teams, or other urgent response groups with a single action.

Alternatively, the Smart ICU design could incorporate a lightweight mobile interface that provides clinicians with continuous access to real-time vital signs, alerts, and trend data while enabling quick documentation at the point of care. Features such as bedside note capture, voice recordings, or secure clinical images could sync immediately with the patient record. Integrated device recognition, using QR codes or NFC technology, could support accurate patient or equipment identification, further reducing documentation errors. In a data-intensive intensive care environment, such an app would serve as a true mobile extension of the Smart ICU, offering rapid access to information, supporting timely decision-making, and keeping clinicians connected to critical updates even when they are away from the workstation.

The Adoption Challenge

While technological feasibility was rarely questioned, the group spent considerable time on adoption barriers. Prof Constantin raised the issue of adoption and how ICU workers will adapt to the changes brought about by a smart ICU. Prof Chiche estimated that digital transformation success depends 10% on developing the model, 20% on making it robust and reliable, and 70% on adoption, transforming human behaviour and workflow.

Several factors influence adoption:

Explanation: Decision support systems must explain their reasoning. Dr Chiche raised the need for a "why" button that explains the physiological principles and data patterns underlying recommendations, and how this could help build trust and provide ongoing education.

Autonomy: Systems that mandate specific actions will face resistance. Co-pilot models that offer options and support clinical judgment are more likely to succeed than autopilot approaches that remove decision-making authority.

Demonstrable Value: Early adopters need to see clear benefits, whether reduced workload, improved outcomes, or enhanced job satisfaction. One strategy might be to prioritise innovations that make caregivers' lives better, even if patient outcome benefits take longer to demonstrate.

Cultural Differences: The discussion revealed varying perspectives on technology adoption across countries and healthcare systems. For example, Prof You Shang noted that some regions readily embrace cameras and automation, while others face significant privacy and regulatory constraints. Successful innovations must navigate these diverse contexts.

Economic Considerations and Access

Prof Vincent raised a question about equity: Will smart ICU technology create a two-tier system where wealthy institutions provide superior care while resourcelimited settings fall further behind?

The group offered an optimistic counterargument. Smart ICU systems should reduce costs by improving efficiency, reducing length of stay, and enabling fewer staff to care for more patients effectively. Prof Chiche highlighted that data from Brazil showed that public hospitals using performance feedback systems increased efficiency enough to treat significantly more patients with the same resources.

Moreover, resource-limited settings often adopt technology creatively, bypassing traditional barriers. Dr Wong described teaching ultrasound to African colleagues who used WhatsApp and Google Drive to share cases and receive guidance: simple, effective approaches that wealthier institutions might reject due to regulatory complexity.

As technology costs decrease and the benefits of efficiency become clearer, smart ICU capabilities should become more accessible, not less. However, this requires industry commitment to developing scalable, affordable solutions rather than premium products for elite institutions only.

While the roundtable focused primarily on promising innovations, participants acknowledged that not every technological advance deserves pursuit. Some innovations seem technically feasible and reasonably accurate but are ultimately not valuable enough to justify costs for the small number of patients who might benefit.

The lesson: technical capability alone doesn't ensure success. Innovations must address genuine clinical needs, integrate smoothly into workflows, and provide value proportionate to their costs. The industry should engage clinicians early in development to identify which problems truly need to be solved.

Continuous Learning Environment

Perhaps the most compelling vision was of the smart ICU as a continuous-learning environment, not just for AI systems but for the humans who work there. Prof Constantin noted that ICU caregivers often spend only a few years in intensive care units before moving to other roles, making ongoing education essential.

Dr Wong also pointed out that if smart ICU systems can make caregivers feel more valued, competent, and continuously developing, rather than overwhelmed and inadequately prepared, they address one of healthcare's most pressing workforce challenges. Technology that facilitates learning without interrupting workflow becomes not just a clinical tool but a retention strategy. Most colleagues who go into healthcare want to do the best job possible. Whatever can be done to help them develop and make their learning easier should be encouraged. Making it easy for people to do the right thing, both clinically and educationally, defines good workflow design.

Smart systems should teach while they assist, explaining physiological principles, highlighting relevant evidence, and gradually building expertise. Rather than replacing human intelligence, AI should augment it, making good clinicians better and helping less experienced caregivers learn faster

This learning extends to the system itself. With proper data feedback loops, smart ICUs should continuously improve their algorithms, adapting to local practice patterns while incorporating new evidence. Hence, the goal should not be static perfection but ongoing evolution.

The vision of the smart ICU as a "selflearning environment" took concrete form in several proposals:

Contextual Education: Prof Shang noted that when the system detects an abnormality or recommends an intervention, it should provide explanations at varying levels of detail. A trainee might want detailed physiological reasoning; an experienced clinician might just need a reminder of recent guideline changes.

Interactive Guidance: Ultrasound systems already demonstrate this principle. Newer devices guide operators through proper probe positioning and measurement techniques, teaching as they perform clinical examinations. Similar functionality could extend across all monitoring and therapeutic systems.

Post-Event Debriefing: Prof Rubulotta mentioned that after managing a critical event like cardiac arrest, the system could generate an automated report summarising what happened, comparing actual management to guidelines, and identifying learning opportunities—turning every crisis into an educational experience.

Badge-Linked Learning: Prof Chiche offered an innovative suggestion involving linking educational content to individual clinician identities. If a system provides an explanation that a clinician doesn't have time to review during patient care, they could tag themselves to receive that information later for self-study. This way, the system learns the knowledge gaps and interests, providing personalised continuing education.

Medication Safety and Reducing Human Error

Prof Chiche also raised concerns about medication safety. The integration of infusion pumps with monitoring systems and electronic health records presents significant opportunities to prevent medication errors, one of the most common sources of harm in intensive care.

Specific recommendations included:

Catalogue Reconciliation: Ensuring that infusion pump drug libraries align perfectly with EHR medication databases, eliminating discrepancies that can lead to dosing errors.

Barcode Integration: Implementing barcode scanning at the point of care to verify that the right medication, at the right concentration, is being administered to the right patient.

Automated Fluid Balance: If all infusions are connected to the monitoring system, calculating accurate fluid balance becomes trivial; yet many ICUs still perform this calculation manually, a time-consuming, error-prone process.

Drawing another parallel to surgical robotics, the discussion turned to how smart systems might actively prevent errors. Prof Constantin discussed liver resection surgery and how robotic systems could potentially warn surgeons when they're about to cut in anatomically unusual locations: 

This system, capable of detecting deviations from best practices or the clinician's typical patterns, could catch errors before they cause harm. The key is framing these interventions as helpful reminders rather than autocratic mandates, preserving clinical autonomy while providing a safety net.

Prof Rubulotta also discussed that the multi-professional nature of ICU care adds complexity. ICU teams involve physicians, intensivists, nurses, pharmacists, dieticians, and others. How do you provide decision support to a team rather than an individual? The answer may lie in decision support systems that facilitate team communication rather than directing individual actions, highlighting discrepancies or concerns that the team should discuss rather than dictating specific interventions.

Family Engagement and Patient-Centred Design

Prof Rubulotta highlighted an important gap: the smart ICU collects extensive patient data but rarely allows patients and families to input information or interact meaningfully with the system.

Research on medical emergency teams suggests that family and patient concerns often predict deterioration before clinicians recognise problems. A truly comprehensive smart ICU might include mechanisms for patients to signal distress or families to raise concerns, integrating these subjective inputs with objective monitoring data.

Medical equipment can be intimidating and mystifying for families. What is that machine? What do these numbers mean? Is my loved one getting better or worse?

Smart systems could include familyfacing interfaces that explain, in accessible language, what different devices do and what monitoring parameters indicate. Rather than generic pamphlets about ventilators and dialysis machines, families could interact with screens that provide personalised information about their specific family member's condition and treatment.

Both Prof Constantin and Prof Rubulotta noted that many ICU patients are increasingly awake and aware, even while on ECMO support. These patients have questions too: What's happening to me? What's my prognosis? What can I do to help my recovery? Patient-facing interfaces could address these information needs while respecting clinical boundaries.

Prof Rubulotta also spoke about recent innovations, including virtual reality systems that allow geographically distant family members to "visit" ICU patients remotely. While technologically impressive, the group questioned whether this level of sophistication should be a priority.

More practical was the concept of automated patient diaries, a labour-intensive process in which nurses document the patient's journey through critical illness, which are later provided to patients to help them understand what happened during periods of sedation or delirium.

Dr Wong pointed out that LLMs could automate much of this process, extracting key events from medical records and generating narratives appropriate for patient and family consumption. The system might create different versions for different audiences: a detailed summary for referring physicians, a family-appropriate narrative omitting technical details, and a coroner's report format when needed.

An important insight was that, while smart ICUs capture enormous amounts of patient data, they rarely allow patients or families to enter information into the system.

Research on medical emergency teams shows that family and patient concerns often predict deterioration before clinicians recognise problems. How might systems capture and act on these subjective warnings?

Suggestions ranged from simple call buttons that alert the clinical team when families are worried, to more sophisticated "ambient listening" that could detect distress in patient or family voices and flag it for clinician attention (with appropriate privacy protections and consent).

The goal isn't replacing human interaction. Participants emphasised that awake patients and present families can and should communicate directly with nurses and doctors. Rather, it's ensuring that when direct communication doesn't happen or concerns aren't adequately addressed, there's a backup mechanism to escalate worries that might have clinical significance.

The Dream Interface

Prof Chiche said that his dream is a monitoring system where basically you have nothing on the screen: just a green dot telling you everything is fine. When something isn't fine, you see a spider diagram where different dimensions represent different organ systems or physiological domains. You see one dimension turning orange —maybe haemodynamics or ventilation —and you don't understand why. So you click on that dimension and get the detailed waveforms with insightful information. You can then ask the system, 'Tell me why this changed from green to orange,' and it will explain its reasoning. This vision embodies several principles:

1. Default simplicity: Don't overwhelm clinicians with data when everything is stable
2. Prioritised complexity: Surface detailed information only when needed
3. Explainable AI: Always provide reasoning behind alerts and recommendations
4. Interactive depth: Let users choose their level of engagement with the system

The Giant Screen Debate: Information Display and Design

The roundtable meeting raised another question: Could ICU rooms of the future feature vast displays with different sections tailored for clinicians versus families?

The appeal lies in information accessibility and customisability: important data are prominently displayed, with less critical information available without cluttering the view.

The practical applications seemed clear:

1. Emergency situations: During resuscitation, the most critical parameters and emergency protocols displayed prominently
2. Checklist integration: Procedural checklists appearing when needed (intubation checklist when preparing to intubate, trauma protocol when a trauma patient arrives)
3. Imaging integration: X-rays and CT scans displayed alongside physiologic data for integrated interpretation
4. Team coordination: All team members able to see the same information simultaneously, facilitating communication

However, concerns emerged about sustainability and patient experience. Does every ICU bed need a massive screen? What about the patient's perspective: do they want to lie beneath a wall of numbers and waveforms? Perhaps sophisticated displays make sense in resuscitation bays and procedure rooms, but not in standard patient rooms, where simpler interfaces and actual windows might better support healing.

The environmental impact also merited consideration. Healthcare's carbon footprint is substantial, and ICU care is particularly resource intensive. While no one suggested forgoing beneficial technology for environmental reasons, the principle of appropriate rather than maximal technology deployment reflects responsible stewardship.

The Cost Element: Investment, Value, and Return

The superficial observation that ICUs have become more expensive as they've become more technologically sophisticated is undeniable. The monitoring systems, therapeutic devices, and support technologies in modern ICUs dwarf the capabilities and costs of units from even two decades ago.

However, several counterarguments emerged:

Value vs. Price: Like smartphones, ICU technology may cost more in absolute terms but delivers more value. The question isn't whether costs have risen but whether value has risen faster.

Opportunity Costs of Not Innovating: Without technological advancement, ICUs might need even more staff to provide inferior care. Technology can substitute for scarce labour, particularly in regions with severe nursing shortages.

Length of Stay Reduction: Evidence suggests that well-integrated systems with good performance feedback can reduce ICU length of stay by significant margins. Prof Chiche cited one example: a 0.8-day reduction in ICU stay and a 4-day reduction in total hospitalisation. With the same bed capacity, units could treat 17% more patients, a dramatic efficiency gain.

Complication Prevention: If smart systems reduce hospital-acquired infections, medication errors, or other complications, the avoided costs of treating these complications may exceed the technology investment.

For medical device manufacturers, the challenge is demonstrating value that justifies premium pricing. If a smart monitoring system costs substantially more than a conventional monitor, ICUs will demand evidence of benefits, reduced complications, shorter stays, better outcomes, or improved staff satisfaction and retention.

The group suggested that rather than focusing narrowly on costs, companies should emphasise value creation. Show that the system prevents complications worth thousands of dollars per patient. Demonstrate that it improves nurse retention, saving recruitment and training costs. Prove that it enables higher patient throughput without quality degradation.

While less glamorous than AI and decision support, cybersecurity and IT infrastructure represent critical challenges that can make or break smart ICU implementations.

Prof Chiche observed that the budgets and capabilities of hospital IT departments are usually underestimated compared to what is expected of them. Implementing sophisticated interconnected systems requires extensive IT support. Hospitals considering such investments must budget for the IT infrastructure and personnel to support them.

Healthcare has become a prime target for cyberattacks, with ransomware attacks disabling hospital systems and compromising patient data. The more interconnected and dependent on digital systems ICUs become, the more vulnerable they are to such attacks.

Device manufacturers must design secure systems, but hospital IT departments must implement security protocols, manage access controls, and respond to threats. The division of responsibility isn't always clear, and failures can have catastrophic consequences.

While the roundtable didn't claim expertise in cybersecurity specifics, they acknowledged it as a critical consideration that cannot be an afterthought. Any company developing interconnected smart ICU systems must prioritise a secure architecture from the outset.

Validation, Regulation, and Trust: The Path to Adoption

A major theme addressed the practical realities of bringing smart ICU technologies from concept to bedside implementation.

Prof Rubulotta pointed out that before clinicians will trust AI-driven recommendations, they need evidence, preferably from rigorous clinical trials. These trials could compare AI-augmented care to usual care on mortality, complications, and length of stay; measure time savings and workflow efficiency for clinicians; assess user satisfaction and trust among healthcare workers; evaluate cost-effectiveness from a health system perspective; and test generalisability across different institutions and patient populations.

Dr Wong emphasised that clinician conviction was probably the most important component of any implementation. Even with robust evidence, sceptics can argue that their unit is special, their patients are different, and their practice is already optimal. Conversely, when clinicians become believers in a technology's value, they find ways to overcome obstacles, arguing for budget allocation, adapting workflows, and training colleagues. Trust is the critical factor that determines whether innovations languish unused or become integral to practice.

How do you build trust? The roundtable offered several strategies:

Transparency: Explain how systems reach conclusions; never provide recommendations without accessible reasoning.

Education: Help clinicians understand the principles underlying AI recommendations, making the technology less like a "black box" and more like a consultant

Gradual Implementation: Start with simple, obviously beneficial features (like automated documentation) before introducing more complex decision support

User Involvement: Include end users in development and testing; systems designed with clinicians rather than for them gain acceptance more easily

Performance Feedback: Show clinicians how the system improves their outcomes, patient safety, or workflow efficiency

Peer Endorsement: Leverage opinion leaders and early adopters to encourage broader acceptance

Strategic Guidance for Innovating the Smart ICU

As the expert roundtable entered its concluding session, the focus shifted from identifying challenges and innovations to practical implementation strategies. This final discussion explored what medical technology companies should and should not do as they develop smart ICU solutions.

Overly Complex Parameters

Prof Chiche highlighted the fact that companies frequently develop new monitoring parameters or measurements, then ask clinicians: "We've developed this technology. Can you find a use for it?" This approach, in which technology seeks a problem rather than solving an identified need, rarely succeeds. The advice to manufacturers was unequivocal: Focus on solving existing, well-defined problems rather than creating new parameters that clinicians must then figure out how to incorporate into care.

Prediction of Specific Syndromes vs General Deterioration

The discussion of predictive analytics revealed an important distinction. Prof Elbers highlighted that while many systems attempt to predict specific conditions (sepsis, ARDS, pulmonary embolism), the definitions of these syndromes are often debatable and difficult to operationalise from data.

More valuable would be the prediction of general patient deterioration requiring immediate attention, regardless of the specific diagnosis. ICU syndromes are "broad and debatable," making them poor targets for prediction algorithms. What matters isn't labelling the problem as "sepsis" versus "bleeding" but recognising that action is needed.

This insight applies particularly to early warning systems on general hospital wards, where studies show that systems predicting "deterioration" work as well or better than those specifically targeting sepsis, despite the latter's diagnostic specificity. As Prof Rubulotta explained, the greater need is to predict the need for action, not to assign diagnostic labels.

The panel reached consensus that prediction should drive therapeutic action, not diagnostic labels. The value lies not in diagnostic precision but in creating awareness that prompts appropriate clinical response. Therefore, prediction must identify deterioration first; specific therapeutic guidance can follow once the clinical picture clarifies.

A Framework for Smart ICU Development

As the roundtable concluded, several overarching principles emerged that should guide medical technology companies in their innovation efforts:

1. Solve Real Problems, Don't Create New Ones: Don't develop technology and then search for applications. Identify the challenges clinicians actually face and design solutions explicitly addressing those problems.
2. Integrate Rather Than Fragment: In an environment already suffering from information overload and disconnected systems, don't add another standalone device or application. Build platforms that unify data and make information more accessible, not less.
3. Make Things Simpler, Not More Complex: More technology doesn't mean more complexity. The best systems handle complexity internally while presenting simple, intuitive interfaces externally. The goal is to make clinicians' jobs easier, not harder.
4. Preserve Human Autonomy and Judgment: Build co-pilots, not autopilots. Provide recommendations and insights, but leave decisions to human judgment. Systems that dictate actions will face resistance; systems that augment human capabilities will be embraced.
5. Explain Your Reasoning: Never provide recommendations without accessible explanations. Clinicians need to understand why the system suggests a particular action, both to evaluate the appropriateness of the recommendation and to learn from the interaction.
6. Design for the Team, Not Just the Individual: ICU care is inherently collaborative. Technologies that facilitate team communication and shared situational awareness will prove more valuable than those optimising individual tasks.
7. Invest in Data Quality: Sophisticated algorithms are worthless if trained on poor data. Partner with excellent institutions, develop robust data collection protocols, and use AI to clean and annotate datasets before using them for model development.
8. Prioritise What Matters Most: Do not try to do everything at once. Focus on the high-impact opportunities: interconnectivity, decision support, predictive analytics, workflow optimisation, and education integration. Leave novelties like voice control and comprehensive surveillance for later, or never.

Implementing the vision of a smart ICU will take time. The priority must be on foundational elements before pursuing more speculative additions.

The question isn't whether smart ICUs will happen—the consensus was that they will, and sooner than many expect. The question is how healthcare systems, technology companies, and clinicians will adapt to work effectively with these intelligent systems while preserving the irreplaceable human elements of empathy, judgment, and care.

The future of intensive care will be smarter, but also more human, more educational, more satisfying for caregivers, and ultimately better for the patients and families who depend on these life-saving units in their most vulnerable moments.