At Visiba, we believe safe AI triage requires calibrated confidence: the ability to represent uncertainty, update it as new information emerges, and use it to determine the safest next action.
Just as importantly, we believe that clinical AI must be auditable. It should be possible to understand not only what outcome a system reached, but why it reached it – and to reconstruct how the available evidence led to that particular recommendation.
At first glance, triage may look like a classification problem: a patient describes their symptoms, and the system assigns a level of care.
In reality, patients rarely present with complete information. Symptoms may be vague, observations unavailable and the condition still evolving. When someone reports breathlessness, dizziness or chest discomfort, the system must consider:
The appropriate response is not always to make a prediction. It may be to ask another question, obtain an observation or arrange clinical review.
This makes triage fundamentally different from a conventional classification task. The system is not simply mapping an input to an output. It is reasoning over incomplete evidence and deciding what should happen next.
Intelligence is not simply the ability to produce the right answer. It is also the ability to recognise how strongly the evidence supports that answer—and when more information is needed before acting.
An AI triage system might be accurate in 95% of cases. But if it presents every assessment with complete confidence, it provides no way to identify the 5% in which it is wrong. Common presentations and unfamiliar combinations of symptoms are communicated with the same certainty, despite carrying very different risks of error.
A clinician faced with an ambiguous presentation may ask further questions, seek advice or take a more cautious course. An AI system that cannot recognise its uncertainty may instead give a definitive recommendation in the same fluent language it uses when the evidence is strong.
This is the difference between accuracy and calibration. Accuracy measures how often a system is correct. Calibration measures whether its confidence reflects how often it is correct. If a well-calibrated system assigns 70% confidence to a group of assessments, approximately 70% should prove correct.
Calibration does not mean always being right—or constantly hedging. It means being confident when the evidence justifies it, uncertain when it does not, and reliably distinguishing between the two.
Frontier language and foundation models are becoming increasingly capable, but their expressed confidence is not necessarily a reliable indication of whether they are correct. Fluency can create the appearance of certainty without a genuinely calibrated estimate of uncertainty.
For triage, reaching a conclusion is not enough. A safe system must know how much confidence that conclusion deserves.
Some uncertainty is inherent in medicine. Patients with similar symptoms may have different diagnoses, and the course of an illness cannot always be predicted. More information will not remove this uncertainty entirely.
Other uncertainty arises because information is missing or a presentation is unfamiliar. A history may be unclear, an important observation unavailable or the case outside the data on which the model was developed.
The distinction matters because it should change what the system does. If uncertainty can be reduced, it might ask about the onset of pain, request a temperature or clarify a relevant risk factor. If it cannot be reduced safely within the interaction, the patient may need escalation.
In this context, “I do not know” is not a failure. It is clinically useful information—provided it leads to the right action.
Large language models have transformed what is possible in healthcare AI. They can interpret natural language, conduct fluid conversations and translate patients’ descriptions into clinically useful information.
A patient may say:
“I’ve felt strange for a few days and I’m much more tired than usual.”
An LLM can explore what “strange” means and identify relevant details. But the apparent confidence of its language should not be confused with calibrated clinical confidence.
LLMs generate plausible responses by predicting language. They do not necessarily maintain an explicit, coherent probability distribution over their clinical beliefs. An answer may therefore sound authoritative even when the evidence is weak. Asking the model how confident it is does not solve the problem: a percentage expressed fluently is not necessarily a validated probability.
The danger is not only that an LLM may be wrong. It is that it may be confidently wrong in a way that encourages reliance on its recommendation.
This also creates an auditability challenge. If the reasoning underlying an outcome is not explicitly represented, it can be difficult to reconstruct precisely why the system arrived at one recommendation rather than another. A fluent explanation generated after the fact is not necessarily the same thing as an underlying, reproducible reasoning process.
Uncertainty is useful only if it changes behaviour.
When confidence is low, a triage system might:
The acceptable level of uncertainty depends on both the recommendation and the consequences of getting it wrong. The confidence required to offer self-care advice for a minor problem is different from that required to exclude sepsis, stroke or acute coronary syndrome. In high-risk triage, even a relatively unlikely serious condition may justify escalation.
Calibration must therefore be assessed in the context of the clinical decision, not as an isolated technical measure.
But there is another important requirement: the basis for that decision should be inspectable and explainable.
If a system recommends urgent assessment, clinicians and developers should ideally be able to determine what drove that recommendation. Was it the patient's age? A particular symptom? The duration of symptoms? A risk factor? A combination of several findings? Or the absence of information that would otherwise have reduced uncertainty?
That distinction matters because an auditable system can be investigated, challenged and improved.
Structured probabilistic models, including Bayesian networks, explicitly represent relationships between symptoms, risk factors, diseases and outcomes.
They can begin with prior probabilities, incorporate new evidence and update the likelihood of different outcomes. This reflects how triage unfolds: each answer should change what the system believes and influence what it asks or recommends next.
One of the most significant strengths of this approach is explainability.
A Bayesian Network does not simply produce an outcome; it provides a structured representation of the factors contributing to that outcome. Each piece of evidence can be traced through the network, showing how it changes the probability of relevant conditions and, ultimately, how those changes influence the recommended action.
This creates the possibility of reconstructing the decision pathway step by step.
For example, if a patient is initially assessed as suitable for routine care, a reviewer can examine the underlying probabilities and evidence supporting that assessment. If the patient subsequently reports a new symptom or risk factor, the network can be examined again to see how that new information changed the relevant probabilities and whether it crossed a decision threshold for escalation.
The important point is that the outcome does not have to be treated as a black box. It is possible to understand exactly how the system got there.
This is particularly valuable in healthcare, where an outcome may need to be reviewed after the event. A clinician, safety team or developer may want to ask:
With a structured probabilistic model, these questions can be investigated against the model's explicit structure and inputs rather than relying solely on an explanation generated by the AI.
This also creates a powerful mechanism for clinical governance. If a recommendation appears surprising, the reasoning pathway can be inspected. If a relationship in the model is clinically inappropriate, it can be identified and reviewed. If new evidence changes the expected relationship between symptoms and outcomes, the relevant part of the model can be updated and subsequently validated.
Auditability therefore becomes more than a technical feature. It becomes part of the safety architecture.
It also supports reproducibility. Given the same model, assumptions and evidence, the pathway to the resulting probabilities and decision can be reconstructed. That makes it possible to distinguish between a genuine change in the patient's information influencing a deterministic inference model and a change caused by an opaque system behaving differently in a non-deterministic way.
However, probabilistic models are not automatically well calibrated or clinically correct. Their performance depends on their data, assumptions and validation in the population where they will be used. Calibration may also deteriorate as populations and clinical practice change. It must be measured, monitored and maintained.
The advantage is not that a Bayesian Network is inherently infallible. It is that its assumptions, relationships and decision pathways can be made explicit enough to be examined.
The ability to inspect how an outcome was reached is also important after deployment.
Suppose a particular triage recommendation repeatedly results in clinical review, and subsequent assessments show that the original level of urgency was too low. In an opaque model, identifying the underlying cause may be difficult.
In a structured model, it is possible to investigate the pathway that produced the recommendation. Perhaps a particular symptom was assigned too little evidential weight. Perhaps a risk factor was not sufficiently connected to the relevant condition. Perhaps the prior probability is no longer representative of the population being served.
The point is not that every error can automatically be explained or corrected. Rather, the system provides a framework within which errors can be investigated systematically.
That creates the ability to have a transparent feedback loop:
ongoing monitoring → audit → review → model improvement
In safety-critical healthcare applications, that traceability and oversight supports trust, assurance and accountability.
The most promising approach combines different forms of AI.
An LLM can provide the conversational layer: understanding the patient, asking questions naturally and converting unstructured language into clinically relevant information.
A structured clinical reasoning model can then evaluate that evidence, update probabilities and support a calibrated decision about urgency and next steps.
Additional safeguards can identify missing information, unfamiliar presentations and cases outside the system's validated scope.
This combination separates two capabilities that are often conflated: understanding language and making a clinical decision.
The LLM can be highly flexible in understanding what a patient means without being solely responsible for determining the clinical outcome. The structured reasoning layer can provide explicit probabilistic reasoning and a traceable decision pathway. Together, they can provide both conversational flexibility and greater transparency around the clinical logic being applied.
Uncertainty must also be communicated carefully. A probability alone may confuse or falsely reassure. Patients and clinicians need to understand what is known, what remains uncertain and what that means for the recommended action.
And where appropriate, the system should be able to expose its reasoning pathway to those responsible for clinical oversight ensuring that the underlying decision process can be examined when needed.
In healthcare, the question is rarely just whether an AI system produced the right answer.
We also need to know whether it behaved appropriately when the evidence was incomplete, whether its confidence was justified, and whether its decisions can be understood and scrutinised.
This is where calibrated uncertainty and auditability reinforce one another. A calibrated system tells us how strongly the evidence supports an outcome. An auditable system allows us to understand why that evidence led to that outcome.
Together, they provide something that raw accuracy alone cannot: a basis for trust that can be examined rather than simply assumed.
The future of AI triage will not be determined by which system sounds most confident or achieves the highest benchmark accuracy. It will depend on whether systems remain reliable when information is incomplete, symptoms are unusual and the consequences of error are high- and whether, when something goes wrong, we can understand what happened.
The safest AI will not be the one that always provides a confident answer. It will be the one that knows how strongly the evidence supports that answer, knows what to do when it does not, and allows us to understand exactly how it got there.