What a Diagnosis Carries

Clinical AI, evolving medical knowledge, and the decisions that follow

By Luis Cisneros, CEO of Serelora

Imagine a patient presents with bilateral lower extremity edema, dyspnea, orthopnea, and reduced exercise tolerance. You press a finger firmly against the skin over the shin. When you lift it, an indentation remains. Two or three seconds pass, and it is still there. The patient describes becoming breathless during activities they previously tolerated and needing to sit upright to breathe comfortably.

You are using an AI scribe during the encounter. As you take the history and examine the patient, you describe the findings aloud. You ask when the symptoms began, review the medications, and discuss the possible explanations. The system captures the conversation and organizes it into a clinical note.

At the end of the encounter, you open the draft. The history appears accurate. The examination includes the edema you described. Under the assessment, the system has written CCF, congestive cardiac failure. The interpretation may be reasonable. It fits a familiar clinical presentation, and the rest of the note reads coherently. You review the draft, make the changes you consider necessary, and accept the assessment.

Then you ask the system to review the patient’s treatment. Its decision support function identifies heart failure in the note, retrieves a treatment pathway, and compares the medication list with the therapies associated with that diagnosis. It flags apparent omissions and proposes changes. If the practice also uses an agent to carry out approved work, those recommendations may lead to orders prepared for review, follow up tasks, or requests for authorization. What began as an entry in the assessment is now informing a sequence of decisions.

The system has reached a point where it needs more information than CCF provides. It needs to establish whether the patient has heart failure with reduced ejection fraction, known as HFrEF, or heart failure with preserved ejection fraction, known as HFpEF. It needs to account for significant valvular disease, right ventricular dysfunction, and whether the current symptoms represent an acute deterioration. CCF may serve as shorthand among clinicians who share the relevant context. A system using it to select a treatment pathway has to establish that context before relying on it.

Even then, a documented cardiac condition does not necessarily explain every finding. Renal disease, hepatic disease, venous insufficiency, or medication effects may contribute to the edema. Heart failure may be established while its contribution to the present symptoms remains uncertain. The initial assessment can therefore be reasonable and still leave important questions unresolved for the decision the system is trying to make.

The need for this specificity becomes clearer when we consider how the treatment recommendations came to exist. In 2014, a large trial comparing sacubitril/valsartan with enalapril demonstrated reductions in cardiovascular death and hospitalization for heart failure in patients with reduced ejection fraction. The finding changed the treatment of a defined population. Applying it required knowing whether the patient being treated belonged to the clinical population for whom that benefit had been demonstrated.

Five years later, dapagliflozin was shown to reduce the combined risk of worsening heart failure or cardiovascular death in patients with reduced ejection fraction, including those without diabetes. A medicine developed to treat diabetes had acquired an established role in heart failure beyond its original indication. Its clinical significance had changed as the evidence developed. An AI system reviewing treatment would need to understand that development to interpret why the medicine might be appropriate for someone who did not have diabetes.

These findings joined established evidence for other therapies and helped shape the four pillars of treatment for chronic HFrEF. The first is inhibition of the renin angiotensin system, preferably with an angiotensin receptor neprilysin inhibitor when appropriate. The other three are a beta blocker with established benefit, a mineralocorticoid receptor antagonist, and an SGLT2 inhibitor. Applying this framework still requires attention to clinical stability, blood pressure, renal function, potassium, tolerability, and the patient’s circumstances. Its presence in a guideline does not settle whether each treatment is appropriate for a particular patient at a particular moment.

Behind that framework are decades of evolving understanding of cardiac physiology, left ventricular function and remodeling, and the effects of treatment on disease progression and mortality. Landmark clinical trials and increasingly refined phenotyping have made treatment more specific because the conditions grouped under heart failure differ in ways that affect care. That specificity allows us to apply accumulated knowledge without treating fundamentally different syndromes as though they are interchangeable. As the knowledge passes through guidelines, specialist services, clinical training, and ordinary practice, the conditions attached to each recommendation remain part of its meaning.

By the time a recommendation reaches a clinical AI system, it may look like a straightforward item on a treatment pathway. The system still needs to understand the population studied, the benefit demonstrated, and the limits of the finding. A reduction in a combined outcome involving hospitalization and cardiovascular death does not automatically establish an independent mortality benefit. Nor can evidence from one population be carried into another simply because both have heart failure documented in their records. Retrieving the right reference is only useful when the system can determine what it supports.

This brings us back to the note. The scribe may have reduced a specific phenotype discussed during the encounter to CCF. It may have presented a possible diagnosis as established. It may also have recorded the encounter faithfully, while the decision support function assumed that the documented diagnosis supplied details that had never been established. These are different failures, arising at different points. Once the familiar term is accepted as a sufficient account of the patient’s condition, however, each can become difficult to recognize.

Suppose the system looks for supporting evidence and finds an echocardiogram reporting an ejection fraction of 30 percent. That result appears to answer an important question. But the examination was performed several years earlier. A later report may show improvement or deterioration, treatment may have changed, and the current symptoms may differ from those present at the time. The earlier finding remains clinically relevant, even if ventricular function subsequently improved. Understanding the patient requires retaining both the historical finding and the subsequent course. An accurately extracted number can still be used inaccurately when its date and relationship to later events are lost.

The medication review raises a similar problem. Before classifying an apparent omission as a care gap, the system needs to determine whether the medicine is truly absent or simply missing from an incomplete list. If the patient is not taking it, the reason matters. Perhaps it was prescribed previously and held after symptomatic hypotension. Perhaps there was hyperkalemia or a deterioration in renal function. Another clinician may have recommended a temporary interruption. The patient may have been unable to afford it or may have declined it after a discussion.

Each explanation changes the work that follows. A previous adverse effect requires reassessment in light of the patient’s current circumstances. A temporary interruption calls for reviewing the conditions under which treatment was supposed to resume. A financial barrier may require help obtaining a medicine that has already been judged appropriate. A decision to decline treatment requires understanding the patient’s preferences and the discussion that informed them. Treating all of these situations as an instruction to initiate a missing medication would discard the information needed to address the actual problem.

That information may be distributed across the chart. The prescription appears in one place, the laboratory result in another, and the decision to hold treatment in a telephone encounter. Connecting those records requires care about what their relationship establishes. A symptom occurring after a medication change may suggest an adverse effect, but timing alone does not prove causation. The clinician may have documented the relationship as suspected, established, or sufficiently concerning to justify a precautionary decision. The system needs to preserve that distinction. Otherwise, it can turn a tentative concern into a permanent contraindication or overlook a documented reason for caution.

The meaning of a clinical finding depends on these relationships. Its presence, timing, possible cause, and relevance to the decision all matter. A medication list tells us something different when it is read alongside the events that produced it. A diagnosis tells us something different when we know whether it was considered, confirmed, questioned, or copied from an earlier note. Clinical AI has to retain those differences as information moves between documentation, interpretation, and action.

This is the practical purpose of traceability. If the system concludes that the patient has HFrEF, the clinician should be able to inspect the evidence supporting that conclusion. A statement about ventricular function should lead to the relevant report and its date. A statement about intolerance should lead to the documented symptoms, measurements, and clinical assessment. When the system recommends a treatment, it should make visible both the patient information and the clinical evidence on which the recommendation depends. A guideline can explain why a therapy is appropriate for a particular phenotype. It cannot establish that the patient has that phenotype.

Questions become necessary wherever that connection remains incomplete. If the phenotype is unclear, the system should seek the relevant evidence or identify the need for clinical clarification. If a medication was stopped, it should investigate why. If the records disagree, it should make the disagreement visible. The question should be chosen because its answer could change the interpretation or the next action. Asking whether a medication was previously attempted serves a different purpose from asking whether the patient could obtain it, even when both questions begin with the same apparent omission.

The system also needs to recognize which uncertainties prevent a particular step and which can be investigated while appropriate care continues. The patient’s current condition may require attention before every historical detail has been reconciled. A treatment review should not obscure a change in symptoms that alters urgency. Clinical significance depends partly on the decision being considered, and a useful system must keep that decision in view as it gathers more information.

Training and evaluation should follow this same logic. A system should respond differently when a treatment has never been considered, when it was stopped after an adverse event, and when the patient could not obtain it. It should distinguish a current finding from a historical one, a diagnosis considered from a diagnosis excluded, and a patient’s account from a clinician’s interpretation. An evaluation that checks only whether the expected diagnostic term appeared in the output can miss the errors that emerge when that term begins directing work. The question is whether the system used the available information appropriately and recognized what it could not yet establish.

Those expectations become more consequential when an agent can act on a recommendation. The workflow needs to distinguish gathering information, preparing work for review, obtaining clinical authorization, and carrying out the authorized task. The agent’s permissions should reflect those differences. If a proposed action depends on an unresolved assumption, that dependency needs to remain visible to the person reviewing it. Approval is meaningful when the clinician can understand what is being proposed and why.

The record must then distinguish what was recommended from what actually occurred. A prescription can be transmitted without the patient obtaining the medicine. A test can be performed without its result being reviewed. A visit can be scheduled without taking place. Each event may complete one part of the work while leaving another unresolved. Responsibility for what remains should be clear, and a task should be marked complete only when there is evidence appropriate to that task.

Auditability allows the team to reconstruct this sequence. It should be possible to establish what information was available, which system version was used, what the AI generated or changed, who reviewed or authorized it, and what action followed. This matters because later steps can be internally consistent with an interpretation that was inadequate from the beginning. A recommendation may follow correctly from an assumed phenotype while the assumption itself remains unsupported. Recording only the final output would conceal where the reasoning failed.

Return to the moment CCF appeared in the draft note. The abbreviation seemed unremarkable because it belonged to the vocabulary of the encounter. Its consequences emerged as the clinician and the system began relying on it to determine what should happen next. Behind those decisions were decades of evolving understanding, along with conditions that governed how that knowledge could be applied. The patient’s record still had to establish whether those conditions were met.

Clinical AI needs to preserve the reasoning that connects the encounter to the action. The patient’s words, the examination findings, the working diagnosis, the supporting investigations, and the proposed treatment each contribute something different. Their relationships have to remain visible as the work proceeds. Otherwise, a system can document a plausible diagnosis, retrieve a respected guideline, and complete an entire sequence of tasks without ever resolving whether those tasks were appropriate for the patient who presented in the first place.