Superintelligent
Care
Introducing Research at Concurrence
Today, we’re launching a research program at Concurrence that explores how to create systems that learn from every patient and carry that learning into the care of the next. We call this superintelligent care.
Medicine already has a name for the idea of care that learns from itself. In a learning health system, knowledge is generated as a byproduct of delivering care. This remains rare in practice today, and we believe it is for two reasons.
The evidence available to guide care decisions is often insufficient, and many of the events that decide patient outcomes are not documented or considered in their records.
Even where evidence exists, it typically describes populations, while care decisions concern individuals. When it comes to determining whether a treatment will succeed for a specific patient, there is significant variation based on that individual’s biology, history, behavior, and circumstances.
Both are partly a result of limitations on what we are able to observe. Clinicians see patients for brief encounters, lacking visibility into the intervals between visits or the experience of colleagues caring for similar patients. To build a learning health system, we must widen the aperture to capture the right types of evidence.
Our vision of
superintelligent care consists of the following:
We believe superintelligent care will emerge from several components that work together as one system. These components include frontier and specialized models, rich patient representations, calibrated uncertainty, governed action, and a way to learn which actions helped drive patient outcomes. Its success will be measured by how much better patients do with it than without it, and how that advantage grows as the system learns.
Thus, our research agenda is organized around four pillars: context, governance, outcomes, and learning. For each, we describe the problems we think must be solved and what makes them hard.
Context
Seeing the whole patient
Understanding the full picture of a patient’s health is a very hard problem, because a lot of what they experience happens outside of our view. The medical record is a fragmented representation of a patient’s world, containing important facts that support documentation and billing, but failing to provide a complete account of the details that make up a person's life and health.
Our first research problem is to build a “digital twin” of the patient. This is a computational model that draws on their longitudinal history. It updates as new information arrives and it forecasts how their health and care might unfold. The digital twin brings together a patient’s clinical state, their response to prior care, and all of the context that is relevant to their future care. The challenge is capturing and integrating all of the relevant details, and how the remaining unknowns should be represented to construct a view of the patient that looks beyond the medical record.
AI agents participating in care create opportunities to capture more of this context as it emerges. For example, a patient might mention during a conversation that they have not booked a follow-up appointment because they no longer have access to transportation. Additionally, agents enable patient interactions to be captured in rich detail, both during and outside of visits. This produces the rich granular context required for learning, including what was done, why it was done, under what circumstances, and what happened both at the clinic and at home as a result of that interaction.
We build these digital twins in a system called the Health World Model, which brings together data from across a patient’s care and organizes it chronologically. This system houses the history and current state of each patient, providing a structure that allows us to observe the many factors that interact to produce outcomes. Over time, our ultimate goal is to anticipate how a patient’s health and care may unfold, recognize emerging risks, and where the evidence supports it, identify actions that could change their course.
To get there, we plan to train a longitudinal generative foundation model of patient observations and care actions. Alongside it, specialized predictive models, causal estimators, and validated clinical tools will draw on the shared patient representation. How to train these pieces and make them work together, including recognizing when each capability applies, is an open research question.
Patient arcs and discovering archetypes
Fortunately, courses of illness and care are not entirely unique to individuals. We call a segment of an individual patient’s trajectory an arc. There are recognizable patterns that recur across populations, such as the sequence in which symptoms develop or the points at which patients encounter barriers to continuing treatment. We call the abstracted patterns archetypes. They provide a way for learning systems to connect a patient's possible futures to prior experience.
From one patient’s arc to a shared archetype
A segment of one patient’s trajectory.
The pattern abstracted from similar arcs that recur across many patients.
A fuller patient representation should help a system recognize when a patient’s unfolding arc resembles an archetype and identify which features of their history make earlier cases relevant. Comparing similar arcs within an archetype can also reveal where outcomes diverge, providing a starting point for investigating what might have changed them. A major challenge will be determining which archetypes and past arcs are relevant to a given patient.
Governance
Uncertainty and causality
Knowing what you know is especially difficult in healthcare because the medical record is not a neutral representation of patients' lives. It carries the fingerprints of the process that produced it.
Two features of this process are especially important:
Informed presence: Health records only capture people when something brings them into care, meaning that patients who seek care more frequently have more opportunities for their health to be observed and documented than equally healthy patients who visit less.
Confounding by indication: Certain interventions might be directed toward sicker patients, which results in those interventions being associated with worse outcomes, even if they are effective.
In a well-known case, a model trained to predict pneumonia mortality learned that asthma lowered a patient’s risk. Asthmatic patients were reliably sent to intensive care, and because the model couldn’t see that extra care, it credited their better outcomes to the asthma itself.
We treat this as one of the central problems in our research. Part of the solution is to model how the data came to exist, including who was seen and why they were treated, so their influence on what the system learns can be understood and addressed. Some effects will resist identification from the available data without additional assumptions or evidence, such as prospective studies. The system should also be able to recognize when a patient or a care setting is substantially different from anything it has previously seen, and in those cases it should become less confident or hand the decision to a clinician.
Solving the diverse range of problems we are interested in also requires leveraging various methodologies. For example, predicting readmission risk is a very different proposition compared to estimating whether a specific intervention would prevent it. Our research will work on developing dynamic strategies for our system to recognize and trigger different modeling strategies when the evidence is insufficient.
Earning the right to act
For over 18 months, we’ve been building a simulation engine for clinical AI agents and learning how to make performance in simulation transfer to real patient interactions. We use these environments to surface failures prior to deployment, then draw on experience in care to improve what our simulations test.
This gives us a starting point for studying when agents can safely take on greater responsibility. Physicians earn specific clinical privileges through demonstrated competence, and we think agents should earn permission to act on the same basis.
One question is whether an agent can recognize when a patient needs a capability it has permission to use. An agent conducting routine follow-up might encounter symptoms that warrant a structured triage assessment. The assessment itself may be well validated, but the agent still has to recognize when to initiate it. Missing that moment can cause a failure even if every subsequent step would have been correct. This is even more significant when it involves making a decision within a regulated pathway – an area of active research for us.
We want to understand what evidence is sufficient to authorize these decisions and how reliably simulation can provide it. Validation in actual care settings will remain necessary, and clinicians will approve any changes to clinical protocols and agent capabilities that follow from new findings.
Outcomes
We believe there is an enormous opportunity for agents to improve clinical and operational outcomes. An encouraging precedent has been established in the field of oncology, where weekly digital symptom reporting with automated triage extended median overall survival from 26 months to 31 months compared to standard care in a randomized trial.
A critical question to ask is what we consider to be a “good” outcome. A system that learns needs to know what it is trying to improve. Clinicians define what good care achieves, such as what counts as remission or recovery, how it is measured, and which options are acceptable at all. Patients decide what matters most to them when acceptable options compete. Organizations set the constraints of cost and capacity, but those constraints are never the target.
Some decisions are more dependent on preference than others. John Wennberg made the distinction between effective care, where benefits are established and tradeoffs are limited (e.g. treating bacterial pneumonia), from preference-sensitive care, where the choice depends on what a patient values (e.g. setting a weight loss target and deciding which side effects are tolerable). A patient’s priorities can also change, such as when the treatment for late-stage illness moves from cure to comfort.
So the system should estimate what each option is likely to do for each outcome that matters, from symptom relief to side effects, burden, and cost, and leave the weighing to the people entitled to do it. A forecast showing that most trajectories reach the target, but a third involve significant nausea for two months, lets patients and clinicians choose with the tradeoff in view. Additionally, for every partner, including life sciences companies that need to understand medication adherence and real-world effectiveness, the principle is the same:
Measure rigorously, and optimize for the patient.
Learning
Finally, learning should allow what happens in one patient’s care to improve our understanding of others. This requires determining how far a finding applies beyond the cases in which it was discovered. Our learning research will study how to build this knowledge from patient trajectories and revise it as new evidence arrives.
We plan to develop AI Discovery Agents that examine trajectories alongside our team at a scale that would be difficult to do manually. For example, they might identify patients whose medication adherence declined after the first month but recovered after structured reminders. These findings give us hypotheses to investigate about what could change a patient’s course.
Findings that survive evaluation would become available for use in subsequent care, subject to clinical approval. Each new trajectory would provide further evidence. If a patient responds differently than expected, the system should reconsider whether the finding applies to them or whether its explanation is incomplete.
Simulation can help design these investigations. When two explanations fit the same histories, we could simulate interventions to identify a test that would distinguish them. Simulated outcomes follow the model’s assumptions, so they cannot independently establish that those assumptions are correct. The result would still need to be observed in actual care.
Combining a patient’s longitudinal history with relationships supported by evidence would give us a model of how their health and care may unfold. We want to use this model to estimate how they might respond to a care plan, which outreach they are likely to answer, and when additional support could help. Whether these estimates remain reliable as a patient’s circumstances change is a central research question.
Our research will focus on:
Several factors can contribute to the same outcome. We need methods for determining which relationships account for a patient’s course and when the evidence warrants a different explanation.
We will test whether findings hold in populations beyond those in which they were discovered. Failures to transfer may reveal circumstances that the original explanation overlooked.
The system’s decisions affect the evidence it later learns from. If it incorrectly concludes that a patient is unlikely to respond to outreach, it might contact them less often and interpret their lack of engagement as confirmation. We need ways to detect and interrupt these errors.
We believe care could improve faster if institutions could learn from one another’s experience. We are beginning research into how we can extract clinically useful patterns and findings in a privacy-preserving manner. The aim is to create a network in which insights travel between consenting institutions while each one maintains stewardship and privacy of its patients' information.
Where we are
We are building the shared data foundation and launching the research program described here. Our agents are deployed in care today and have supported over 10 million patient interactions, deliberately within well-established clinical protocols where performance can be evaluated against existing standards.
Our first research priorities are patient representation and privacy-preserving archetypes. This includes testing whether context captured through care improves predictions beyond the clinical record alone, how much of that value survives privacy protection, and the extent to which it transfers to another setting.
The frontier lies beyond these first steps. Consider a goal as ordinary as helping a patient eat well. It seems simple enough, yet whether a particular person changes their diet depends on a tangle of biology, habit, household economics, and small daily frictions. A system that could take in that tangle, recognize the patterns within it, foresee which plan suits which person, and learn from what follows could turn common advice into a personalized intervention that drives outcomes.
Whole regions of medicine sit in this category: outcomes everyone wants and no one can yet reliably produce.
Problems like these make superintelligent care worth pursuing. We are working toward it, and we will share what we discover along the way.
If you are a researcher, clinician, or partner who wants to extend or challenge these ideas, we would like to hear from you.