Systemic fragility in European total intravenous anesthesia delivery and opportunities for resilient real-time decision support

Abstract

Total intravenous anesthesia (TIVA) is a well-established technique for general anesthesia and is increasingly relevant to Europe’s transition toward more digital, sustainable, and patient-centred healthcare. However, its use remains uneven across centres and still depends heavily on population-based drug models, variable monitoring practices, and repeated clinician adjustment. Differences in training, data access, device compatibility, patient response, and institutional resources can limit safe and consistent personalization. In this Review, we examine the main sources of fragility in current TIVA practice and discuss how interoperable perioperative data, digital patient simulators, patient-adaptive models, multimodal monitoring, and closed-loop decision support could improve safety and resilience. We emphasize that these technologies should support, not replace, the anesthesiologist, and should remain transparent, clinically supervised, and easy to override. We conclude by outlining priorities for European collaboration in data infrastructure, training, validation, regulation, and evaluation of AI-enabled tools.

Ionescu et al. outline a European roadmap for strengthening Total Intravenous Anesthesia Delivery through resilient, human-in-the-loop real-time decision support. Current practice remains uneven and fragile, while interoperable data, validated simulators, adaptive models, robust control, and clear regulation enable safer personalization.

European landscape context

Systemic fragility describes the vulnerability of complex systems whose components are tightly interconnected and mutually dependent¹. In such systems, even a local disturbance can propagate across the network and compromise overall performance. Healthcare systems are a clear example: hospitals, pharmaceutical supply chains, medical personnel, information systems, and emergency-response mechanisms are deeply interdependent, so shortages of staff, medications, intensive-care capacity, or critical equipment can rapidly reduce system-wide responsiveness. The COVID-19 pandemic exposed these structural vulnerabilities across European healthcare systems, including limited surge capacity, workforce strain, and supply-chain disruption, highlighting the need to strengthen long-term health-system resilience and sustainability². Resilience, in contrast, refers to the ability of a system to absorb disturbances, adapt to changing conditions, recover performance, and maintain essential operations during crises^3–5.

Perioperative and critical care services are high-impact areas where healthcare resilience and sustainability can be addressed together. Healthcare accounts for approximately 4–5% of global greenhouse gas emissions, and TIVA offers a lower-carbon alternative to volatile anesthesia by eliminating direct anesthetic gas emissions and reducing dependence on volatile-agent supply chains^6–8. Its compatibility with digital infusion platforms and closed-loop control also supports operational flexibility during resource constraints. However, TIVA adoption remains heterogeneous across Europe, reflecting disparities in technology access, training, and workforce capacity^7,9. European initiatives, such as the European Health Data Space (EHDS) and the EU AI Act https://health.ec.europa.eu/ehealth-digital-health-and-care_en, provide a favorable framework for trustworthy AI, health-data infrastructure, and human-centered perioperative decision support. Yet real-time AI-assisted anesthesia remains limited by scarce physiological data, complex patient dynamics, and insufficient patient-specific computational models. This review paper, therefore, outlines a roadmap for scaling TIVA within Europe’s digital and sustainable healthcare transformation.

The article is assembled by European research groups at the intersection of computer science, control engineering, biomedical engineering, and anesthesiology. It reflects representative expertise and current practices in computer-guided TIVA management in Europe. Supplementary Material, Section 1 provides workforce density data, a Web of Science keyword-based survey across EU countries, and a qualitative word-cloud analysis, highlighting uneven closed-loop anaesthesia research activity and the need for clearer terminology distinguishing TCI, advisory systems, and truly closed-loop computer-assisted TIVA delivery. As summarized in Fig. 1, the European TIVA landscape connects current clinical practice with computer-based decision support and the remaining challenges toward safe closed-loop anesthesia.

Fig. 1. European TIVA landscape and progression toward computer-supported anesthesia.

The schematic links current European clinical practice for total intravenous anesthesia (TIVA) with patient monitoring, infusion delivery, computer-based modelling and control, and real-time decision-making.

The paper first summarizes current TIVA-TCI practice, where population-model-based infusion is manually supervised by the anesthesiologist. It then discusses data access, digital patient simulators, and closed-loop control as foundations for personalization, benchmarking, and safe testing. The final sections address technological bottlenecks, scientific challenges, and opportunities for resilient European TIVA decision support, emphasizing safety supervision, transparent AI-enabled tools, and clinician-centered systems that support rather than replace the anesthesiologist.

Clinical practice for TIVA

TIVA-TCI as current clinical practice

TCI systems administer intravenous anesthetic drugs through microprocessor-controlled syringe drivers. The anesthesiologist enters patient characteristics, selects a plasma or effect-site target concentration, and the pump calculates infusion rates using population-based pharmacokinetic and pharmacodynamic models. Propofol is typically used for hypnosis, and remifentanil for analgesia. TCI pumps and advanced monitoring devices are therefore central to modern TIVA practice¹⁰.

However, standard TCI remains open-loop. Infusion rates are computed from a priori patient information and population models, without automatic adaptation to real-time patient-response data. The anesthesiologist closes the loop manually by interpreting predicted effect-site concentrations, clinical signs, and physiological measurements, and by adjusting targets when needed. TIVA performance, therefore, depends not only on models and devices but also on clinical experience, workload, procedure complexity, and local practice.

Models, monitoring, and clinician supervision

Commercial TCI systems rely on population-based pharmacokinetic models to estimate drug distribution and effect-site pharmacodynamic models to describe the delay between plasma concentration and clinical effect. These models enabled routine clinical implementation, but they cannot fully account for inter-patient variability, comorbidities, or changing intraoperative sensitivity¹¹. Advanced response-surface models of propofol–remifentanil synergy remain mainly confined to research and simulation environments. Further details on patient models used in TCI are provided in Supplementary Material, Section 2.

Advisory systems can support anesthesiologists by visualizing predicted drug effects or suggesting target concentrations, but they do not automatically compensate for model mismatch, unusual drug sensitivity, or sudden surgical disturbances. Closed-loop TIVA represents a further step by adapting infusion automatically from measured depth-of-hypnosis signals, while still requiring robust monitoring, safety supervision, and clinician acceptance^12,13. Such systems should therefore be understood as extensions of supervised clinical decision making, not as replacements for the anesthesiologist. Remimazolam has also been evaluated as an alternative intravenous hypnotic in Japanese multicentre studies¹⁴. For remimazolam, re-sedation after flumazenil reversal highlights the importance of clearance and post-reversal monitoring¹³.

Perioperative monitoring and co-interventions

TIVA is embedded within broader perioperative monitoring and treatment. Routine monitoring includes electrocardiography, heart rate, blood pressure, oxygen saturation, and end-tidal carbon dioxide. In major surgery, additional hemodynamic variables, such as cardiac output, stroke volume, stroke-volume variation, and pulse-pressure variation, may guide cardiovascular management. Because advanced monitoring is not universally available, European digital TIVA systems should remain modular: functional with basic signals, such as mean arterial pressure and heart rate, but extendable when richer hemodynamic data are available.

Fluid therapy, vasoactive drugs, opioid-based analgesia, and neuromuscular blockade are also integral to TIVA practice. For real-time decision support, the key issue is not only which variables are monitored, but whether they are available with sufficient sampling rate, latency, reliability, and interoperability for feedback control or safety supervision. Additional clinical and technical details are provided in Supplementary Material, Section 2.

European fragility in TIVA accessibility and training

TIVA adoption remains uneven across Europe. Survey data indicate substantial variability in frequency of use among anesthesia providers, suggesting inconsistent utilization and variable clinician experience¹⁵. European-level data remain incomplete, particularly for Eastern Europe.

Barriers to broader adoption include limited access to infusion pumps and depth-of-anesthesia monitoring, difficulties with intravenous access, limited detection of tissued cannulas, and the inability to measure real-time Propofol plasma concentrations¹⁶. Equipment cost, logistics, education, and institutional familiarity all contribute to this fragility. Strengthening European TIVA practice, therefore, requires not only improved technology but also harmonized training, transparent certification pathways, and human-centered implementation.

Data infrastructure for personalized and resilient TIVA decision support

Data access is not a secondary requirement for computer-guided TIVA, but one of the core infrastructures needed for personalization, validation, benchmarking, and regulatory evaluation. Real-time decision support depends on models that can be trained, updated and tested using clinically representative perioperative data. For TIVA, such data must go beyond generic vital signs and include synchronized physiological time series, drug-administration records, surgical-event annotations, co-interventions, patient covariates, and clinically meaningful outcomes. Without these elements, algorithms may perform well in simulation or single-center studies but remain difficult to generalize across European healthcare settings.

Public datasets for perioperative and critical-care research

Several public perioperative and critical-care datasets now support physiological modeling, outcome prediction, and algorithm benchmarking. However, no current dataset fully captures the information required for personalized closed-loop TIVA. The main gaps concern synchronized drug delivery, surgical stimulation, high-resolution physiological monitoring, intervention records, waveform availability, and outcome annotation. Operating-room datasets are most directly relevant for TIVA automation, whereas ICU datasets provide scale and longitudinal information but are less aligned with intraoperative drug titration. PhysioNet is an open platform hosting various physiological and clinical collections with resolutions from milliseconds to minutes¹⁷. Table 1 summarizes the key characteristics of available datasets, with additional details provided in Supplementary Material, Section 3.

Table 1.

Key characteristics of public datasets

Dataset	Cohort size	Resolution	Waveform	Centers	Data types/access
VitalDB51	6388 subjects	62.5–500 Hz/ 1–7 s	Yes	Single (SNUH)	Waveforms, demographics, labs, and procedures/simple Python API.
INSPIRE52	131,109 subjects	5 min	No	Single	Vitals, labs, meds, and outcomes/PhysioNet access.
MOVER53	83,648 surgeries	1 min	Yes	Single	Vitals, waveforms/registration required.
PulseDB54	5,361 subjects	10 s segments	–	Multi	ECG, PPG, ABP, and demographics.
Global	Physiological signals, clinical records
HiRID55	~ 34,000 ICU stays	2 min	–	Single (Bern)	Monitors, labs, therapies, and observations.
AmsterdamUMCdb56	23,106 admissions	≤1 min	–	Single (Amsterdam)	Vitals, labs, scores, meds, and procedures.
eICU-CRD57	200,859 admissions	1 min-event	–	Multi (U.S.)	Vitals, labs, APACHE, and care plans/ controlled access.
MIMIC-III58	58,976 hospital admissions	~1 h	Yes	Single (BIDMC)	Waveforms, vitals, labs, meds, notes, and outcomes/controlled access.

An important distinction remains between unrestricted open access and controlled access. Some datasets, such as MIMIC-III and the eICU Collaborative Research Database, require user training and data-use agreements before access is granted. These safeguards are necessary for patient privacy, but they also affect the speed and reproducibility of algorithm benchmarking across research groups. In Europe, such governance issues are especially important because future TIVA decision-support systems must comply with strict requirements for data protection, interoperability, and accountability.

Data-driven model updates for personalized TIVA

Data access alone does not enable personalized TIVA. Patient models must also be updated from sparse, noisy, and clinically disturbed data. While pharmacokinetic behavior can often be approximated using established population models, pharmacodynamic response remains highly variable across and within patients, especially during induction, maintenance, and emergence.

Model updating should therefore respect the clinical structure of anesthesia. Induction may provide a relatively disturbance-free window for estimating patient sensitivity; maintenance is affected by surgical stimulation, hemodynamic interventions, and other intraoperative events; and emergence corresponds to decreasing drug concentrations and the return of physiological reflexes. These procedure-dependent effects can be represented through digitalized disturbance profiles^18,19. Embedding adaptation within this phase-specific structure is essential for patient-specific updates that remain clinically interpretable, safe, and compatible with real perioperative workflows, as illustrated in Fig. 2. Technical details are provided in Supplementary Material, Section 3.

Fig. 2. Anesthesia control workflow under conventional and digitalized surgery scenarios.

Upper panel: Schematic representation of the conventional workflow, showing the bispectral index (BIS, blue), propofol infusion rate (red), and remifentanil infusion rate (purple) during dose-effect identification, controller tuning, initiation of automatic control, and surgery. The vertical dashed lines indicate key workflow transitions. Lower panel: Proposed workflow when the surgical procedure is digitalized, comparing the true surgical disturbance (yellow) with the available a priori digital representation (green).

Digital patient simulators for TIVA

Digital simulators are essential for computer-guided TIVA because they provide a safe intermediate layer between clinical data, model development, and bedside deployment. In anesthesia, prospective testing of new control algorithms is constrained by safety, regulatory, and ethical requirements; simulators, therefore, allow drug-delivery strategies, closed-loop controllers, and patient variability to be explored before clinical implementation. For resilient European TIVA decision support, they serve four complementary roles: controller development, safety testing, clinician training, and regulatory evidence generation.

Only a limited number of open-source simulators have addressed multidrug or multivariable anesthesia control. Key examples include the Matlab/Simulink-based simulator by Ionescu et al.¹⁸, the Python Anesthesia Simulator by Aubouin-Pairault et al.²⁰, and the AReS simulator²¹. Together, these platforms extend TIVA simulation from hypnosis-only control toward multidrug and multivariable management, including hypnotic, analgesic, neuromuscular, and hemodynamic components.

Despite this progress, current platforms remain patient simulators rather than true digital patient simulators or digital twins. They capture representative population-level dynamics and allow reproducible controller testing, but they do not yet provide continuously updated, patient-specific computational representations of the individual undergoing TIVA. Validation also remains limited, with available evaluations based mainly on artificial, retrospective, or small clinical datasets. Additional details on model structures, disturbance handling, and validation limitations are provided in Supplementary Material, Section 4.

For future European TIVA systems, the key opportunity is to move from generic simulators toward clinically validated, interoperable, and updatable digital patient simulator frameworks. Such frameworks should integrate real-time patient data, perioperative events, surgical disturbance profiles, uncertainty-aware model adaptation, and standardized safety scenarios. This would allow simulators to support not only controller design, but also clinician training, safety testing, regulatory evaluation, and human-in-the-loop decision support.

Computer-guided optimal closed-loop control of TIVA systems

Solutions that move beyond conventional TCI have been widely investigated in the control-engineering literature. For depth-of-hypnosis control, both in-silico and experimental studies have considered proportional-integral-derivative (PID) control, model-based feedforward control, fuzzy control, and model predictive control (MPC)²². Clinical studies have demonstrated the feasibility of closed-loop propofol control in both paediatric and adult patients¹². These SISO approaches can improve the consistency of hypnosis regulation and reduce dependence on repeated manual target adjustments. However, they address only one component of TIVA and therefore do not fully capture the clinical complexity of anesthesia delivery. Their translation also depends on robust monitoring, safety supervision, and controller behavior that remains understandable and acceptable to the anesthesiologist.

Despite these promising results, hypnosis-only control does not fully reflect the pharmacological structure of TIVA. Remifentanil administration plays an important role in controlling the response to surgical stimulation, and the interaction between hypnotic and opioid drugs must be considered when designing clinically meaningful control strategies. This has motivated MISO control structures, where propofol and remifentanil are jointly manipulated while the depth of hypnosis remains the main feedback signal. Recent approaches have explored PID-based and MPC-based co-administration strategies to manage the opioid–hypnotic balance within clinically acceptable ranges²³. Clinical and experimental studies have further demonstrated the feasibility of simultaneous propofol–remifentanil administration under closed-loop supervision^12,24. Automated delivery of propofol, remifentanil, and rocuronium has also been evaluated in a randomized clinical trial²⁵. MISO control, therefore, represents an important step toward more realistic TIVA support, but it remains limited by the absence of a mature, widely accepted feedback signal for analgesia.

A further step is the development of MIMO closed-loop architectures, where different feedback signals are used to regulate hypnosis, analgesia, and hemodynamic variables. Such systems are clinically attractive because they better reflect the multivariable nature of anesthesia, including drug–drug interactions, surgical stimulation, cardiovascular responses, and co-interventions. However, their implementation remains constrained by sensing limitations, model uncertainty, safety-certification requirements, and the need for transparent human-in-the-loop supervision. Reliable analgesia monitoring remains a major limitation, and closed-loop Remifentanil regulation has therefore been less extensively investigated than Propofol control. Some studies have used surrogate indicators derived from electromyography, heart rate, blood pressure, or composite nociception scores to guide opioid administration, but these signals remain less mature than depth-of-hypnosis indices. Automated delivery of propofol, remifentanil, and rocuronium has also been clinically evaluated²⁵. More recent clinical work has shown the feasibility of combining automated hypnosis, analgesia, fluid, and hemodynamic management, indicating a possible pathway toward integrated multivariable anesthesia control²⁶. From a European resilience perspective, the challenge is therefore not only to design more advanced controllers, but to embed them within interoperable, validated, and clinician-supervised decision-support systems.

Technological bottlenecks

Human-in-the-loop implementation

Closed-loop anesthesia has shown feasibility and safety in randomized studies, meta-analyses, and engineering prototypes, with reported benefits in maintaining physiological variables within target ranges, reducing drug use, and improving recovery-related endpoints²⁷. However, these systems remain far from routine clinical deployment. Translation is limited not only by controller performance but also by practical implementation barriers, including workflow integration, clinician trust, device interoperability, manufacturer involvement, and the ability to supervise and override automation in real time.

For clinical adoption, closed-loop TIVA must remain interpretable, supervisable, and interruptible. If the anesthesiologist cannot understand the controller’s actions, anticipate its response, or safely override its behavior, the system may be perceived as an additional source of risk rather than as decision support. A practical direction is therefore shared-control anesthesia, where automated algorithms support titration and safety supervision while the anesthesiologist remains responsible for patient-specific and procedure-specific decisions. Such systems should be evaluated not only against physiological disturbances, but also under realistic workflow conditions, including manual boluses, target changes, signal loss, alarm handling, and temporary controller override.

Sensing and monitoring limitations

A major implementation bottleneck for integrated TIVA control is the limited availability of robust and clinically accepted feedback signals. Depth-of-hypnosis indicators are routinely available, but analgesia assessment remains less mature. Several indices have been proposed, including the Analgesia Nociception Index, the Nociception Level index, and EEG-derived nociception-related measures²⁸. EEG bicoherence may also respond to noxious stimulation during general anaesthesia²⁹. Yet their clinical interpretation, causality, and robustness across surgical contexts remain insufficiently established for routine closed-loop opioid control. Recent evidence supports this cautious interpretation: a network meta-analysis found limited outcome benefit across nociception monitors, with only pupillometry reducing intraoperative opioid use, while Linassi et al. showed that qCON/qNOX indices can be strongly influenced by EMG activity and are not fully independent monitoring signals^30,31.

This limitation explains why MISO strategies are often more clinically acceptable than fully MIMO approaches: clinicians may trust familiar hemodynamic signs and pharmacological reasoning more than non-validated analgesia indices. Recent work comparing commercial nociception monitors with a prototype device further illustrates the limited data richness and identifiability challenges associated with nociception monitoring during general anesthesia³². From an implementation perspective, these limitations affect not only controller design, but also clinician acceptance, alarm interpretation, and the safe use of Remifentanil feedback control³³.

Instrumentation suitability for real-time closed-loop optimization

Practical instrumentation constraints also limit translation. Infusion pumps introduce delays, saturation effects, and communication constraints, and their dynamic behavior must be characterized before safe real-time optimization. Beyond individual devices, TIVA decision support also requires interoperability between pumps, monitors, controllers, and hospital information systems to ensure reliable, time-traceable exchange of drug delivery, patient-response signals, alarms, and clinical context.

Monitoring hardware imposes similar constraints. Depth-of-anesthesia monitors, such as BIS and NeuroSENSE provide processed EEG-derived indices at sampling intervals of seconds. Propofol produces characteristic EEG signatures during loss and recovery of consciousness³⁴. These signals are suitable for feedback control, but their update rate, proprietary processing, latency, and artifact sensitivity must be considered in controller design and safety supervision. Nociception monitors, including skin-impedance-based devices, such as Anspec-PRO and multivariate indices, such as Medasense NOL or Med-Storm skin conductance, may provide complementary information on responses to surgical stimulation, especially when combined with depth-of-hypnosis and hemodynamic variables.

Hemodynamic variables remain central for both safety and optimization. Multiparameter monitors provide non-invasive blood pressure, heart rate, pulse oximetry, and end-tidal gases, while advanced systems provide stroke volume and cardiac output. From a control perspective, sampling interval, latency, artifact sensitivity, and interoperability determine whether these signals can serve as primary controlled variables or only as supervisory constraints. Similarly, syringe pumps and TCI systems act as the control actuators, and their internal models, update periods, start-up delays, and communication interfaces directly influence achievable controller performance. A non-exhaustive overview of monitoring and infusion devices used in TIVA research and clinical studies is provided in Supplementary Material, Section 5.

Scientific challenges

Although anesthesia control can be represented through compartmental drug dynamics, clinical TIVA remains a difficult modeling and control problem. The main scientific challenges arise from model uncertainty, incomplete observability, drug interactions, changing surgical stimulation, and physiological dynamics evolving over multiple time scales. These challenges are distinct from the technological bottlenecks described above: they concern what must still be solved at the level of modeling, estimation, control design, and validation before resilient real-time decision support can be generalized across patients and clinical settings.

The pharmacological models used in TIVA are widely accepted for representing intravenous drug effects, but they remain population-based and cannot fully capture inter- and intra-patient variability during changing surgical and physiological conditions^11,35. The pharmacodynamic layer is particularly uncertain, especially for patient-specific drug sensitivity, hypnotic–analgesic interactions, and time-varying responses during induction, maintenance, and emergence.

A central open question is therefore which parameters can be reliably personalized in real time, and under which clinical conditions such adaptation remains identifiable and safe.

A second challenge is observability. Key TIVA states, including analgesia, surgical stimulation, and patient-specific drug sensitivity, cannot be measured directly. Although nociception indices, pupillometry, hemodynamic variables, and soft-sensing methods may improve state estimation, their robustness and suitability for feedback control remain insufficiently validated.

A third challenge is defining clinically meaningful control objectives. Future controllers should move beyond single-index tracking, such as BIS, and incorporate trade-offs between hypnosis, analgesia, hemodynamic stability, drug exposure, safety margins, and recovery.

Control objectives also change during the procedure. Induction requires rapid achievement of adequate hypnosis while avoiding excessive overshoot or prolonged deep sedation. Maintenance requires stable regulation despite surgical disturbances, physiological variability, and clinical co-interventions. Emergence requires timely recovery while maintaining patient safety. This motivates phase-specific, adaptive, and supervisory control strategies rather than a single controller applied throughout the entire procedure²³.

TIVA is inherently multivariable, involving interacting control of hypnosis, analgesia, neuromuscular blockade, hemodynamics, and drug–drug effects. Integrated multi-drug dosing control is therefore attractive, but remains limited by unreliable analgesia feedback, patient-specific model uncertainty, constraint handling, and the need for explicit safety guarantees.

Translation from engineering prototypes to clinical systems also requires stronger validation frameworks. Closed-loop approaches have demonstrated efficiency for depth-of-hypnosis control, but prospective validation under realistic clinical variability remains limited. Computer-based and data-driven methods may support personalization and workload reduction, but they depend on large, well-annotated, and clinically representative datasets, which remain difficult to obtain in perioperative care³⁶. Reinforcement learning and other learning-based methods may offer new opportunities, but their use in safety-critical anesthesia requires rigorous validation, interpretability, uncertainty quantification, and safeguards before clinical deployment.

In control-engineering terms, substantial progress has been made in controller analysis and synthesis for drug titration²⁷. Nevertheless, several open problems remain:

• Control design still relies largely on compartmental models, and it remains unclear which parameters should be personalized in real time.

• The selection of reliable feedback measurements remains unresolved, especially for analgesia and patient-specific drug sensitivity.

• Control during poor observability or degraded signal quality remains insufficiently addressed, although soft-sensing and Kalman-filter approaches provide promising directions³⁷.

• Multicenter clinical trials under realistic clinical variability remain limited.

• Most approaches assume continuous infusion, whereas clinical practice also involves boluses, manual interventions, and temporary overrides.

• Patient history, comorbidities, medication use, and surgical context are not yet sufficiently integrated into control algorithms.

• Most systems still focus on hypnosis control rather than integrated MIMO control of hypnosis, analgesia, hemodynamics, and drug–drug interactions.

• Formal safety guarantees, including constraint satisfaction under model uncertainty, sensor degradation, and clinician override, remain insufficiently developed for routine clinical translation.

Opportunities to strengthen European TIVA clinical practice

Closing the loop

The challenges identified in multi-drug to multi-effect closed loop control frameworks motivate a set of practical recommendations for translating these solutions into clinically viable decision-making support systems.

Adding features in patient simulators: patient simulators should support goal-directed fluid therapy using physiological indicators, such as MAP, SVV, and lactate, while implementing phase-specific strategies consistent with the R.O.S.E. framework. Digitalized clinical decision rules for fluid selection and safety constraints to prevent fluid overload should be integrated, together with predictive models of fluid responsiveness and organ injury. These capabilities would enable advanced real-time decision support and optimization of drug and fluid management during TIVA.

Surgical protocol digitalization: digitization of surgical protocols is essential for translating clinical procedures into structured inputs that can be integrated into control-oriented anesthesia simulators¹⁸. By mapping surgical events to disturbance profiles in the BIS, it becomes possible to test, validate, and optimize drug delivery strategies under realistic conditions¹⁸. To achieve this, representative surgical procedures are grouped into categories based on their complexity and physiological impact, and each category is represented by parameterized disturbance profiles characterized by defined amplitude, duration, and temporal structure. These profiles combine step, ramp, and pulse patterns to emulate intraoperative events and their corresponding effects on BIS, as given in Supplementary Material, Section 6.

Standardization of alarms and safety loops

Patient safety remains the primary prerequisite for the clinical adoption of closed-loop anesthesia systems. Current alarm systems are limited by alarm fatigue, signal artifacts, and fixed thresholds that often fail to account for patient-specific physiology and surgical context. Consequently, safety mechanisms have been extensively investigated, including rule-based constraints on drug infusion rates, drug concentrations, depth of hypnosis, and hemodynamic variables, as well as automatic interventions to prevent awareness and physiological instability³⁸. Formalized safety frameworks based on therapeutic drug windows, blood pressure constraints, and reference governor techniques have further demonstrated the feasibility of maintaining safe operation while preserving clinician authority whenever patient responses deviate substantially from expected behavior³⁸. Because closed-loop performance depends critically on monitoring quality, existing systems typically revert to manual control when signal reliability is compromised, while recent observer- and Kalman-filter-based approaches seek to improve robustness against measurement degradation³⁷. Looking forward, personalized safety architectures integrating electronic health records, comorbidities, surgical protocols, and predictive machine-learning alarms could enable risk-aware optimization and real-time recommendations, allowing human-in-the-loop supervision to remain central to safe and adaptive anesthesia management.

Role of artificial intelligence in real-time TIVA decision making

AI-enabled methods may support real-time TIVA decision making through hypnosis prediction, drug dosing, PK/PD model updating, adverse-event detection, alarm prioritization, and multimodal state estimation. However, in anesthesia these tools must be treated as decision-support technologies rather than autonomous systems, with the anesthesiologist retaining clinical authority and responsibility³⁹. Existing examples, such as Assisted Fluid Management systems, illustrate how algorithmic recommendations can support clinicians while preserving human oversight⁴⁰. Machine learning and reinforcement learning have shown promise for hypnosis prediction, drug dosing, PK/PD modeling, and adverse-event detection, but most applications remain at the proof-of-concept stage and require further validation before clinical deployment⁴¹.

For TIVA, the main value of AI is not replacing control theory or clinical judgment, but strengthening estimation, personalization, and safety supervision under uncertainty. Trustworthy deployment requires reliability, transparency, explainability, and auditable performance, motivating interpretable control approaches and open-source development practices where possible⁴². Ethical AI principles, including fairness, accountability, transparency, safety, and privacy, remain essential for healthcare applications⁴³. They are particularly relevant in TIVA because algorithms may be trained on incomplete or biased perioperative data, may perform differently across patient groups or institutions, and may influence drug titration during safety-critical care^44,45.

The European regulatory context further strengthens the need for transparent and clinically accountable AI-assisted TIVA systems. Future systems will need to be aligned with data-protection and medical-software requirements, including GDPR, the Medical Device Regulation, the European Health Data Space, and the EU AI Act. Regulatory oversight is especially important as medical AI systems become adaptive or capable of continuous learning⁴⁶. Traditional distinctions between locked and adaptive algorithms may not adequately address concept drift, changing patient populations, and post-deployment model updates⁴⁷. Consequently, future certification frameworks for AI-assisted TIVA should combine external validation, continuous performance monitoring, auditability, traceability, transparency, and human oversight.

Several unresolved questions are particularly important for adaptive TIVA decision support. First, informed consent becomes more complex when an adaptive system used at the point of care may differ from the version originally validated in trials. Second, liability allocation remains unclear when recommendations arise from interactions between a controller, clinician, monitoring devices, hospital information systems, and institutional protocols. Third, opaque “black-box” models may challenge clinical accountability and patient communication, especially when recommendations cannot be explained in pharmacological or physiological terms^44,48. These concerns reinforce the need for AI systems that are clinically interpretable, prospectively validated, and designed around human-in-the-loop supervision.

Overall, AI in real-time TIVA should be positioned as an enabling layer for resilient decision support rather than as a stand-alone solution. Its role is to improve model adaptation, signal interpretation, alarm prioritization, safety supervision, and individualized recommendations, while preserving clinician authority. Safe European deployment will require representative datasets, multicenter validation, transparent software governance, post-market surveillance, and continuous monitoring of performance across institutions and patient populations^49,50.

Key message

Roadmap for resilient European TIVA decision support.

Resilient TIVA decision support in Europe requires: (1) harmonized terminology for TCI, advisory systems, decision support, shared control, and closed-loop TIVA; (2) interoperable perioperative datasets with synchronized drug delivery, monitoring signals, surgical events, co-interventions, patient covariates, and outcomes; (3) validated digital patient simulators for benchmarking, safety testing, training, and regulatory evidence generation; (4) human-in-the-loop shared-control systems with transparent recommendations, safety supervision, and clear override mechanisms; (5) standardized evaluation of analgesia and nociception monitoring for feedback control; (6) transparent regulatory pathways for adaptive medical software, including auditability, traceability, post-market surveillance, and liability allocation; and (7) multicenter clinical validation before broad deployment.

Summary

Future TIVA systems should not be framed as replacements for the anesthesiologist, but as resilient human-in-the-loop decision-support systems that strengthen clinical judgment under uncertainty. The current TIVA-TCI paradigm remains vulnerable to population-model mismatch, incomplete observability, surgical disturbances, variable training, and unequal access to technology across Europe. Addressing these fragilities requires shared-control architectures in which automated algorithms support patient-specific adaptation, disturbance rejection, safety supervision, and multi-drug optimization, while responsibility, contextual interpretation, and final clinical authority remain with the anesthesiologist.

The central message of this review paper is that resilient TIVA delivery cannot be achieved through isolated advances in pumps, monitors, algorithms, or AI alone. It requires a coordinated European roadmap built on interoperable perioperative datasets, clinically validated digital patient simulators, uncertainty-aware PK/PD modeling, multimodal monitoring, formal safety constraints, multicenter validation, and transparent regulatory pathways for adaptive medical software. By aligning engineering innovation with clinical workflow, auditability, and human oversight, Europe can move from fragmented TIVA practice toward safer, more consistent, and more personalized perioperative care.

Author contributions

Conceptualization: C.I. and D.C. Methodology: E.Y., M.S., A.V., I.B., C.M., M.F., N.L., B.P., K.M., E.D., T.M., L.M., R.S. and M.P. Formal analysis: A.V., I.B., C.M., M.F. and N.L. Investigation: E.D., T.M., L.M., R.S. and M.P. Supervision: C.I. and D.C. Writing-original draft: C.I., D.C. and E.Y. Writing-review and editing: all authors.

Peer review

Peer review information

Communications Medicine thanks Alfredo Abad-Gurumeta and the other anonymous reviewer(s) for their contribution to the peer review of this work.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains Supplementary material available at 10.1038/s43856-026-01833-x.