Digital Twins and ICU Care

• In ICU
• Thu, 19 Mar 2026

In an interview with ISICEM News, Nathalie Layios, Thomas Desaive, and Vincent Uyttendaele of the University of Liège shared insights on how digital twins could transform ICU care. 

The concept of a digital twin, defined as a continuously updated computational replica of a patient or physiological system, is rapidly transitioning from engineering into intensive care medicine, where it offers a way to model patient physiology in real time and support high-stakes clinical decision-making. Unlike static prediction tools, digital twins evolve alongside the patient, integrating ongoing clinical data (e.g. vital signs, laboratory values, and sensor outputs) to simulate responses to treatment and anticipate changes in clinical trajectory.

In practical terms, a digital twin is a dynamic, personalised model grounded in physiology, physics, or artificial intelligence. It can monitor, predict, and optimise patient states, even when data are initially limited. Early in a patient’s ICU stay, partially personalised models can still guide decision-making by narrowing plausible physiological states and simulating likely responses. As more data become available, the model’s accuracy and predictive capability improve, making it a progressively more reliable tool.

Currently, digital twin applications are uneven across ICU domains. Metabolic models are the most advanced and are already used for personalised glycaemic control in stress-induced hyperglycaemia, representing a mature example of safe clinical application and in silico testing. Respiratory models, particularly for optimising ventilator settings in acute respiratory distress syndrome (ARDS), are the next most developed, supported by robust physiological understanding and continuous ventilator data. Cardiovascular models, while sophisticated, are limited by the lack of routinely high-quality haemodynamic data, restricting their bedside use.

The reliability of digital twins in rapidly changing ICU environments depends on continuous data input, real-time model updating, and robust physiological frameworks. Key technical challenges include ensuring model identifiability (i.e. sufficient data to constrain model parameters) and quantifying uncertainty so clinicians understand the confidence of predictions. Transparency around uncertainty is essential to avoid overconfident and potentially harmful recommendations.

In terms of clinical targets, respiratory failure (especially ARDS) is a realistic near-term application, whereas sepsis remains more complex due to its multi-system nature involving cardiovascular, immune, metabolic, and microvascular interactions. Progress in sepsis is more likely to come from targeted sub-models (e.g. fluid responsiveness or vasopressor management) rather than attempting to model the entire syndrome.

A major advantage of digital twins is their ability to predict deterioration before it becomes clinically apparent. By detecting subtle changes in physiological dynamics and linking them to mechanistic explanations, digital twins can provide earlier, more interpretable, and actionable warnings than traditional early warning scores.

The use of in silico testing, simulating treatment strategies on virtual patients, is already well established in glucose control, where it supports the design and validation of insulin protocols. In other ICU domains, this approach remains largely in the research phase.

Key barriers to implementation include access to high-quality, high-frequency clinical data, rigorous validation across diverse patient populations, and the need to build clinician trust. Trust depends on transparency, interpretability, and clear communication of uncertainty. Additionally, successful deployment requires integration into clinical workflows; tools must deliver timely, practical recommendations rather than remain theoretically sophisticated but impractical.

Ultimately, while the long-term vision is a comprehensive whole-patient digital twin, near-term success will depend on focusing on a small number of high-value clinical decisions. Targeted, efficient, and interpretable models that solve specific problems are more feasible and impactful than attempting to replicate the full complexity of human physiology prematurely. Sustained collaboration between engineers, clinicians, and frontline staff is essential to ensure these systems are both technically robust and clinically usable.

Source: ISICEM
Image Credit: iStock