A differential should show
- Why each hypothesis is being considered.
- Findings that support and weaken it.
- How likely and how urgent it appears.
- Important missing or contradictory information.
- What would cause the ranking to change.
Diagnify diagnostic support
AI differential diagnosis
Diagnify pairs a clinically fine-tuned open-weight model with separate evidence and guideline retrieval. It organises likely, dangerous and alternative hypotheses for a qualified clinician to inspect—not as an autonomous diagnosis.
From clinical history to reviewable output
A useful differential has a visible path from case to evidence
The model and the evidence library serve different purposes, with human review at both development and clinical-use stages.
Curated histories
Diagnify currently describes its internal corpus as more than 50 million historical consultation records spanning 50 years. Counting, provenance and representation details remain part of the research disclosure.
Clinical fine-tuning
A selected open-weight base model is adapted for clinical reasoning rather than used as a general-purpose model with a prompt wrapper.
Training-time feedback
Human reviewers correct weak, incomplete and unsafe reasoning and feed those corrections into development.
Evidence retrieval
A separate, versioned evidence library retrieves relevant guidelines and sources for the live case.
Clinician review
The clinician checks the hypotheses, urgency, evidence and missing context before deciding what to do.
A differential should not imply
- That the top-ranked item is the diagnosis.
- That low probability means no action is needed.
- That every generated condition is clinically plausible.
- That omitted diagnoses have been ruled out.
- That ranking alone determines testing or treatment.
Two human-control layers: training-time feedback improves model behaviour; clinician oversight at use time remains essential because fine-tuning and retrieval do not eliminate omissions or confident errors.
Four clinical lenses
Separate likelihood from consequence
Ranked hypotheses
Order plausible explanations using the available context, while showing that the order is provisional and sensitive to missing data.
Can’t-miss diagnoses
Keep serious, time-sensitive alternatives visible even when they are not the most likely, and state what makes them actionable.
Red flags
Identify findings that may change urgency or destination of care. Red-flag lists are prompts for assessment, not guarantees that risk has been excluded.
Alternative explanations
Consider mimics, medication effects, psychosocial context and multiple simultaneous processes where clinically relevant.
Likelihood and harm are different dimensions. A lower-probability condition may still require urgent exclusion when the consequences of delay are severe.
Safe sequence
Keep independent clinical reasoning in the loop
Assess urgency first
Stabilise and escalate when required. Do not delay action while entering data or waiting for an AI response.
Form an initial differential
Document the clinician’s problem representation and leading alternatives before reviewing model suggestions where practical.
Compare, do not copy
Use the output to look for omissions and contradictions. Challenge unsupported items and watch for anchoring on the generated order.
Investigate discriminators
Prioritise findings that could meaningfully change probability, urgency or management rather than testing every listed condition.
Reassess over time
Update the ranking when symptoms evolve, results return or the response to management does not fit the working explanation.
Probability updates
Use new findings to revise—not confirm—the initial view
A finding changes probability according to its diagnostic performance and the starting probability in the relevant setting.
Before the finding
Estimate a defensible pre-test probability using prevalence, setting, history, examination and the quality of available data.
Record the uncertainty around that estimate rather than presenting false precision.
After the finding
Apply an appropriate likelihood ratio when one is available and applicable. Reconsider whether findings are independent before combining them.
Use the post-test probability calculator for educational support.
Boundaries
Known limitations of AI-generated differentials
Input sensitivity
Small wording changes, missing findings and inaccurate documentation can materially change the output.
Uneven knowledge
Performance may vary by presentation, population, language, setting and conditions not well represented in development data.
Confident errors
A model may invent relationships, sources or certainty. Plausible language must not bypass verification.
Automation bias
Generated rankings can anchor attention and make clinicians less likely to notice contradictory evidence.
Context loss
Goals, access, examination nuance and longitudinal knowledge may not be captured in the submitted information.
Change over time
Model, evidence and workflow changes can invalidate earlier evaluation and require renewed monitoring.
Clinical safety notice: Diagnify is under evaluation for qualified clinicians. Its output may be incomplete or incorrect and must not be used for emergencies, autonomous diagnosis, or patient self-diagnosis.
Related guidance
Build the complete evidence-first workflow
Two offerings, one foundation
Use the workflow directly or build it into your own clinical AI
Diagnify for clinicians
A clinician-facing workflow for structuring the case, inspecting a differential and reviewing retrieved evidence under human oversight.
Private API for medical AI teams
Request a dedicated hosted model and evidence API for your organisation. Isolation, retention, deletion and data-use controls are documented for each deployment.
Clinical decision support
Inspect hypotheses without surrendering judgement
Explore Diagnify’s clinician-facing workflow or request a dedicated deployment for a medical AI product.