← 返回期刊

NEUROLOGY · 2026-09-12 · en

Clinical context: why a symptom is not a diagnosis

What must an AI assistant preserve when it connects complaints, examination findings and a patient’s history? We examine the difference between recognising a pattern and establishing a clinical conclusion.

Read in English · Читать по-русски
AI 生成的编辑插图 · 人物与场景均为虚构

AI 生成的编辑插图 · 人物与场景均为虚构

A complaint becomes clinically meaningful through context. The same description of forgetfulness can refer to different difficulties: sustaining attention, learning new information, retrieving a name or organising everyday activity. A useful clinical record must preserve these distinctions rather than replacing them with a single broad label. This is the starting point for our work on connected clinical information.

The research question is not simply whether a model can name a disease. It is whether an assistant can reconstruct the sequence of observations, distinguish their sources and identify missing information without inventing a link. A symptom reported by a patient, a finding documented during examination and an interpretation in a previous letter are different kinds of evidence. Combining them should not erase that difference.

A suitable evaluation begins with a case that unfolds over time. Initial complaints are followed by history, examination, additional results and a repeat consultation. At each stage, the assistant should state what the available information supports and what remains uncertain. A later finding may change the interpretation; the earlier record should remain traceable rather than being rewritten as though the answer had been known from the start.

Pattern recognition and causal explanation need separate assessment. Co-occurring complaints can suggest a question for further examination, but their association does not establish a shared cause. Similarly, a result outside a reference range does not automatically explain a neurological symptom. A clinically useful assistant should preserve alternative explanations and direct attention to the information needed to distinguish them.

We therefore consider several evaluation dimensions: fidelity to the source, separation of observation from interpretation, preservation of time and negation, and the number of unsupported claims. A fluent summary that changes “no loss of consciousness” into an episode of loss of consciousness fails an essential requirement even if the rest of the narrative reads well.

The practical objective is a better starting point for a specialist: a coherent case overview, explicit uncertainties and questions that can be checked during a consultation. For the patient, the corresponding benefit is an explanation that connects the discussion to their own experience. These are evaluation goals for the platform, not a claim that automated diagnosis has been clinically validated.