Critical care is one of the youngest fields in medicine. Simultaneously, the provision of care to those who are critically ill has formed the basis of medicine since ancient times. It is with the advent of technological innovation that the modern age of critical care has emerged. And likely more than any other field in medicine, it will continue to change, as rapid technological and system innovation occurs, and the needs of taking care of the critically ill evolve. Simultaneously, given its dependence on technology and trained personnel, it is the field most susceptible to system-level inequities in its distribution and accessibility, and any imagining as to what lies next for caring for the critically ill needs to consider how it can respond to those escalating inequities across the communities it serves.

This article will describe some ideas for the future of critical care and what preparations need to happen to ensure that the future achieves what is hoped for. This will by no means be an exhaustive or authoritative list of things that will happen, which is an impossible task; rather, there will be much speculation on how the field will evolve over the coming decades, grouped according to major themes. This will also be a somewhat optimistic vision for a future that is largely unknown and may proceed down various diverse pathways. Other articles on this topic have recently been written, for further reading. ^,,,,,,

No discussion of the future of health care can occur without discussing the emerging, and increasingly established, role of artificial intelligence (AI) in health care. The past few years have seen rapid growth in its visibility across different spheres of health care, with dedicated journals, new technologies, and established systems of utilization.

Within critical care at the current juncture, AI is primarily being considered in its role in clinical decision support. As a data-rich environment with tremendous volumes of granular data produced every second, and hundreds of high-acuity decisions made on a daily basis, critical care will likely undergo transformative changes through the advent of large-scale data processing and analysis.

A future likely exists where every available monitor is seamlessly integrated, and, based on models trained on previous patients in individual units or even across entire health systems, risk scores and relevant recommendations for interventions are provided. These systems could integrate imaging data, physiologic monitoring, laboratory data, and even patient data from the precritical illness stage to produce individually tailored recommendations for care. As with a human critical care providers taking a detailed history, examination, and obtaining diagnostic tests, emerging AI models will help integrate increasingly large datasets that are applicable for every individual patient.

Although there are currently a wide range of traditional scores that exist to predict various outcomes for the critically ill, constant refinement and integration of far greater volumes of data can produce much more robust tools for predicting patient trajectories. Knowing that an individual patient will develop a pneumothorax on that ventilator setting, or that another patient in shock will benefit from hydrocortisone, could be examples of outputs of AI models defining care interventions. A future critical care unit where these recommendations are provided to clinicians in real time to help shape subsequent decisions, conversations with family, and patient flow through units, has already been envisioned, and is likely to become more commonplace over the coming decades. ^, Patients will likely have a baseline level of pre-existing data regarding their health, whether that relates to their genetic information, detailed-omic analysis of various risk profiles, or other metrics of overall status. Integrating these risk profiles with the acute problem that initiated the critical illness would be part of the diagnostic pathway. Like all AI models, these will only ever serve as a complement to human decision making, particularly in the care of the critically ill, where human interactions and intuition will likely remain indispensable. These human interactions, core to the practice of caring for the critically ill since its inception, will only be augmented by these data-driven decision-support tools, allowing for more time spent in conversation and at the bedside than at computers.

Even beyond clinical decision support for providers, assisting with family decision making during a goals-of-care conversation could benefit from AI. Risk prediction is one of the core strengths of AI systems. Families could be privy to updated likelihoods of specific outcomes of individual importance– ability to live independently, cognitive function, whether they will be able to get back to work in the next few months– and use that information, combined with individual expressions of preference stated precritical illness, to help inform surrogate decision making. These incorporated stated preferences could even be through integrating the reams of data available from individuals’ online presence, emails, and other documentation that exists. One of the fundamental challenges of critical care is in understanding the preferences of incapacitated patients, and a future world could integrate baseline patient preferences through large-scale language models incorporating individual patient data.

Fundamental issues with AI systems must be considered as this future state is built where it fairly and accurately adds utility to taking care of critically ill patients. The well-described biases in the development of many of the AI models currently in use should be directly addressed, for fear of propagating ongoing biases in clinical medicine. As decisions of substantial import will be made based on data from these AI models, communities– both from providers and patients– will continue to push for reliable and accurate models that constantly improve. Ensuring that any data used is kept appropriately private and not used for purposes for which it was not designed will be crucial to ensure trust is maintained. There is also a known concern that over-reliance on AI for clinical decision making gradually chips away at clinician knowledge and decision-making ability over time. Frameworks that protect patient data, ensure biases are minimized, and prioritize people in decision making and training will have to be established.

Critical care is reasonably protected, compared with other fields in health care, from job losses caused by AI. The technical and interpersonal nature of the specialty implies that human activities will always be necessary. There is no world where decisions about goals of care can be explicitly reduced to mere regurgitations of statistics, or counseling about critical illness trajectory will be replaced exclusively by a model trained on previous patients. The complementary nature of AI to assist in ways that are currently lacking in the field, namely the integration of reams of granular data with the humanity of modern medicine, shows great promise for the future of the field.

From the original model of ventilatory support for polio, to the current age where nearly every organ has artificial support technologies during acute illness, technology will continue to be the primary driver of change in the field of critical care. Predictions in this space for the model of critical care in 30 years are difficult, given the rapid advancements in bioengineering. Extracorporeal membrane oxygenation (ECMO), for example, has scaled out over the past 2 decades to many regions of the world, and further advancements in how ECMO is used, in terms of reducing its invasiveness, expanding its eligible population, and reducing risk of complications, will likely make it become much more frequent in its application. Having ECMO act as a destination therapy will allow for discharge of patients with miniature or implantable devices with remote monitoring technologies embedded or closed-loop systems integrated, possibly even supplanting endotracheal intubation as the primary mode of support for acute respiratory failure. Other organ support technologies, including liver, kidney, heart, and pancreas, will continue to evolve and expand, and critical care will likely be able to extend organ function while awaiting recovery, transplant, or long-term device-based replacement, which will become far more commonplace.

Monitoring of patients will move beyond the current crude metrics of physiology to more useful information. Real-time monitoring of inflammation through biosensors or other strategies will become commonplace ^, and allow for closed-loop or real-time responsiveness to changes in individual inflammatory profiles as new vital signs are integrated into decision making. The augment of saturation probe monitoring for oxygen transformed critical care over the past 40 years, and subsequent ability to monitor inflammatory profiles may trigger similar understandings.

The ability to monitor neurologic function remains a space in critical care where current technologies are lacking. Abilities are evolving, however, but their accuracy and validity in predicting outcomes remain unclear. A large swath of technological innovation, from implantable electrodes to novel electroencephalogram (EEG) technologies, to imaging techniques, will help evolve understanding, not just of physiologic metrics such as flow or ischemia, but more involved approaches at monitoring brain function. Quantitative metrics of cognition are clearly difficult to establish, but with the advent of brain-machine interfaces, strategies for monitoring and communicating with previously thought-to-be incapacitated patients will expand. A future where every patient with a severe neurologic injury has an implanted brain-computer interface for both direct neurologic monitoring and direct patient communication for subsequent conversations about goals of care, could allow for much more nuanced conversations reflecting true patient needs. The striking finding of cognition in a substantial proportion of vegetative patients will, by necessity, introduce complexities in how physicians communicate with families and patients, and require practitioners to reflect on how care is provided to brain-injured patients.

Fundamental to all of these ideas is ensuring that the tenets of evidence-based medicine are followed. New technologies will always be enticing for a field that is prone to early adoption of new innovations, but making sure that any innovations improve outcomes that patients care about will remain a challenge. Implementation of new devices or interventions without formal evaluation of their role often leads to difficulties in de-implementation. Hence, new models for evaluation are required to respond to the exponential growth in available technologies.

The ideas of learning health systems are particularly germane to care of the critically ill. Data-rich and with thousands of decisions and interventions performed on patients, there is a tremendous amount of improvement that can happen. Building systems where each subsequent patient gets better care than the previous one due to ongoing learnings should be the goal for systems, and fundamental to those ideas is embracing randomization to reflect the highest standards at establishing causality. The coronavirus disease 2019 (COVID-19) pandemic highlighted how systems can work together to implement large-scale platform trials and rapidly accrue useful knowledge. Integrating randomization as the core of these learning health systems through large-scale platform trials that can test numerous interventions simultaneously, with integrated response-adaptive randomization, will allow for maximal efficiency in knowledge generation and implementation. A system where all patients in every ICU are randomized at least once during their journey with critical illness must be the goal of any systems-level thinking regarding the future of critical care. Defining the important outcomes to patients and health systems, inclusive of cost and externalities such as climate impacts, will become mainstreamed in these learning systems.

There is a tremendous amount to learn about how to best manage these patients, and only the surface has been scratched thus far. A future where embracing that humility about how little is known about optimal management, and thereby embracing randomization, can form the core of a learning health system. AI systems can play a role here at integrating with clinical trial design, but they will never replace the importance of randomization to truly define causality.

The growth of concepts of precision medicine, where biologic signatures lead to specific therapeutic interventions, need to be tested in a similar fashion through randomized evaluations. Although personalized care in critical care has always been provided based on physiologic signatures– using an array of physiologic parameters to decide if a patient should get a fluid bolus, for example,– moving toward additional parameters to further personalize therapies may have added value. Of course, ensuring that this is through a randomized approach, where patient outcomes are actually improved, should be the goal. Understanding that there are heterogeneous treatment effects for all of the interventions used in the critically ill, and knowing which parameter defines that heterogeneity, will increasingly be deployed within randomized trials and can inform the future of learning health systems in critical care. Although there is much attention on precision medicine in critical care, it will need to be rigorously evaluated to ensure that it adds value to the future.