Core

Diagnostic Reasoning

Differentials for complex, undifferentiated cases

Diagnostic Reasoning is the engine for the case that does not yet have a diagnosis: the patient whose findings cross organ systems, or whose working diagnosis no longer fits. Describe the case and it reasons from the most telling findings to a ranked differential, says what fits and what does not for each candidate, flags what must be excluded first, and sets out the tests that would change the picture.

Who it is for

  • Physicians and registrars on the medical take with a patient who does not fit a single diagnosis
  • Ward doctors whose patient is not improving as expected, or whose working diagnosis no longer explains the findings
  • Residents preparing a structured presentation of a complex case for a senior colleague or a multidisciplinary meeting

What it does

  • Reasons from the most telling findings

    It picks out the findings that few conditions explain and builds the differential around them, so that a common but non-specific symptom does not lead the reasoning.

  • A ranked differential with the reasoning shown

    Each candidate diagnosis is set against the case: the features that support it, the features that argue against it, and what would be expected if it were the answer but is missing.

  • Looks for one explanation

    It asks whether one disease process could account for findings in several organ systems, and considers the medicines, recent exposures and family history as possible causes before settling on several separate diagnoses.

  • What cannot wait comes first

    Diagnoses that are dangerous if missed are named at the top, with the findings that point to them and the steps needed today, so that urgent work is not lost inside a long differential.

  • Investigations in a sensible order

    Suggested tests come in order, starting with those that are quick, widely available and most likely to settle the question, and each is tied to the diagnosis it would confirm or exclude.

  • Verified references

    The report ends with a numbered list of references read from PubMed and ClinicalTrials.gov when the question is asked, with guidelines and systematic reviews searched first and retracted papers excluded.

How it works

  1. Choose Diagnostic Reasoning and describe the case: the presenting problem and its time course, examination findings and observations, results with their units, medicines and background. Reports and images can be attached.
  2. The engine works out which findings narrow the differential most, and puts any time-critical problem first.
  3. It searches PubMed and ClinicalTrials.gov for the case, with guidelines, systematic reviews, meta-analyses and randomised trials first, and reads the US product information for any named medicine from DailyMed.
  4. It writes a structured report that cites the retrieved records by number. Any PubMed link it writes itself is checked against PubMed.
  5. The report typically arrives in under a minute. The clinician weighs it against the patient in front of them and decides what to do.

What to include

  • Age, sex and the presenting problem with its timeline
  • Examination findings and observations
  • Investigations so far, with values and units
  • Medications, allergies and relevant background

Example cases

  • 58M on empagliflozin, vomiting for 2 days, pH 7.12, bicarbonate 9, glucose 9.8 mmol/L, ketones 5.4. Diagnosis and immediate management?
  • 34F with 6 weeks of intermittent fever, night sweats and pleuritic chest pain after travel to India, new pericardial rub. Differential and work-up?
  • 67M with progressive exertional dyspnoea, macroglossia, nephrotic-range proteinuria and a low-voltage ECG. What unifies this?

What the report contains

  1. Immediate Safety Actions, when a red flag is present
  2. Conclusion and Recommendation
  3. Clinical Dashboard
  4. Differential Diagnosis, ranked, with the findings for and against each
  5. Recommended Work-up, with what each result would change
  6. Management Plan
  7. Reasoning and Pitfalls, including what would argue against the leading diagnosis
  8. Information Needed
  9. Key Evidence and References

See a full example report

Limits

  • It reasons only from what it is given. It cannot examine the patient, take a further history or see results that were not entered, and a missing detail can change the differential.
  • A ranked differential supports clinical judgement; it is not a diagnosis. The treating clinician remains responsible for every decision and should seek senior or specialist advice as usual.
  • Drug information attached to an answer is US product information. Licensed doses and indications differ between countries, and local guidelines and formularies take precedence.

Frequently asked questions

When should I use Diagnostic Reasoning rather than a specialist engine?

Use it when the diagnosis is not yet clear, when the findings cross organ systems, or when the working diagnosis no longer explains the picture. When a problem clearly belongs to one specialty, such as a suspected stroke or a medical problem in pregnancy, the specialist engine follows that specialty's criteria and scores more closely.

What should I include in the case description?

Give age and sex, the presenting problem with its time course, examination findings and observations, results with their values and units, medicines, allergies and relevant background. Normal results are worth including when they help to exclude a diagnosis. Remove names, record numbers and other identifiers before entering the case.

Does it give a single diagnosis?

No. It gives a ranked differential with the reasoning for each candidate, the diagnoses that must be excluded first and the tests that would separate them. It also says what would make the leading diagnosis less likely, so that the ranking can be revisited as results come back.

Are the references in the report real?

The reference list is read from PubMed and ClinicalTrials.gov, and from DailyMed for named medicines, at the time the question is asked. The model does not write the list, so the links cannot be invented. Each entry links to its record, and the report cites the list by number.