Guide for doctors
How clinical case matching works
A practical explanation of similarity retrieval, ranking, clinical interpretation, and why a matched case is a lead for reasoning—not a diagnosis.
Start with retrieval, not prediction
Clinical case matching compares the language and concepts in one presentation with narratives already in a case library. The system looks for overlap in meaning—such as symptom pattern, time course, exposures, or the way a workup unfolded—and ranks potentially relevant cases. It is closer to an unusually flexible literature or memory search than to a diagnostic score. The output says “these stories may be worth reading,” not “this patient has the same condition.”
That distinction matters because two patients can use similar words for different processes, and the same condition can be documented in very different language. A useful interface keeps the original narrative visible, identifies why a case may be relevant, and leaves the clinician responsible for interpretation. SameCase treats matching as a way to widen recall and surface questions that deserve attention; it does not replace history, examination, testing, or clinical judgment.
What enters the comparison
A case narrative can carry several kinds of signal at once: the chief concern, onset and sequence, associated features, relevant negatives, prior tests, treatment response, and contextual details such as travel or medicines. Text-based retrieval can connect related terms even when the wording is not identical. That is valuable when one clinician writes “breathlessness” and another writes “dyspnea,” or when the clue is a relationship between events rather than a single keyword.
The quality of a query still depends on the quality of the description. Copying an entire chart may bury the discriminating details, while a one-line symptom label removes the timeline and context. A compact chronological summary usually performs better: what changed, when, what was found, what was already considered, and what remains unexplained. Identifying information should never be added to make a match more specific.
How to read a ranked result
Open a match by asking two separate questions: what is genuinely similar, and what is importantly different? Similarity might come from presentation, laboratory pattern, exposure, or response to a prior intervention. Differences in age, prevalence, immune status, medication, tempo, or examination can make the eventual explanation in the archived case inapplicable. A memorable final diagnosis should not outweigh those differences.
The safest use is hypothesis generation. A case can remind the reader of a missing history question, a dangerous alternative, a useful discriminator, or a reason an apparently reassuring test did not end the workup. Any next step must be justified from the current patient’s evidence and appropriate guidance—not copied because it appeared in a retrieved case.
- Confirm that the timeline, not just the vocabulary, is comparable.
- Identify at least one feature that argues against the archived case’s resolution.
- Separate a test that generated a clue from a test that actually confirmed the explanation.
- Check whether geography, prevalence, referral setting, or selection bias changes the prior probability.
Known limitations
A case library is not a representative epidemiological sample. Interesting, solved, and publishable presentations are more likely to appear than ordinary or unresolved ones. The absence of a match can mean the corpus is small, the terminology differs, or the relevant detail was never recorded. The presence of many matches can reflect common phrasing rather than a common cause. Ranking scores should therefore never be presented as disease probabilities.
Documentation can also contain hindsight. Once a case is solved, its narrative may emphasize clues that were subtle at the time. Readers should distinguish what was known at each stage from what became clear later. Clearly labeled teaching examples are useful for learning a reasoning pattern, but they are illustrative and must not be mistaken for observations about real-world frequency.
A disciplined workflow
Write a de-identified, chronological problem representation; retrieve a manageable set; compare both similarities and differences; and convert the useful parts into explicit questions. Document any hypothesis that affects care through the ordinary clinical record and local decision process. If a match suggests a rare explanation, look for independent support in appropriate references or specialist input before it changes management.
SameCase deliberately separates the public prose layer from structured member tools. Public readers can learn from the narrative, while verified doctors can use the richer matching workflow. In both settings, the same boundary applies: retrieval can support attention and recall, but responsibility remains with the clinician caring for the patient.