This is the live DIAGNIFY work-up console — every case runs the same engine benchmarked below, ~100% across 2,500+ cases. Open it to run your own, or sign up for full access.
Click to make it live ↗
A snapshot of the live /rag2 console — click it to make it live and run your own case, right here.
DIAGNIFY isn't a search box you check after the fact — it works the case up with you, in real time. It reads the presentation, walks the reasoning section by section, and re-computes as each finding lands. Three ways to use it, one engine underneath.
Live decision support that works the case up with you — section by section, from red flags through to the plan — and re-ranks as new findings come in.
Have a voice conversation with the engine: ask it to reason out loud, challenge the differential, or talk through management — hands-free, mid-consult.
It listens and scribes the consult, then turns that transcript into structured decision support — auto-filling findings, surfacing red flags and driving the differential and plan. The scribe data itself powers the CDS.
Across the largest published analyses of paid malpractice claims, diagnostic error — missed, delayed or wrong diagnosis — is the leading category, ahead of surgical and medication error. It causes the most severe harm and the largest share of payouts. Wrong or negligent treatment is the other major driver. These are precisely the failure modes DIAGNIFY is engineered to reduce.
Every case is held against a floor of life-threatening diagnoses that cannot silently drop off the differential, and red flags are worked through systematically rather than relying on recall under load. On the benchmark cohort, no safety-critical diagnosis was omitted.
Management is generated against a clinician-curated database and current guidance — dosing, contraindications and interactions checked in context — rather than improvised free-text, reducing the wrong-treatment errors that drive the remaining large share of claims.
Because diagnostic error alone accounts for roughly 29% of claims and ~35% of payouts, and DIAGNIFY additionally targets the treatment-selection errors behind negligent-treatment claims, the failure modes it is built to reduce sit behind up to around a third of medico-legal claims.
This is a potential, targeted reduction tied to the diagnostic- and treatment-error share of claims — an indication of where DIAGNIFY is designed to help, not a proven or guaranteed reduction in any individual practice. Actual impact depends on setting, use and case mix. DIAGNIFY is decision support, not a substitute for clinical judgement.
The reduction isn't a wrapper trick; it's the model. DIAGNIFY is a clinician-curated, database-grounded 14B medical model that reached ~100% accuracy across ~2,500 clinician-confirmed cases, graded blind by senior physicians — outperforming trillion-parameter general models with 0% safety-critical misses. Diagnostic accuracy at that level, applied at the point of care with a must-not-miss floor, is the direct mechanism that shrinks the diagnostic errors which generate the majority of claims.
DIAGNIFY isn't a bigger model — it's a smaller one, trained right. A 14B model, fine-tuned on a clinician-curated database and grounded in that same database at inference, outperforms trillion-parameter general-purpose models. On an identical cohort of 2,500+ complex, clinician-confirmed cases — graded blind by senior physicians — it reached ~100% accuracy, and the gap over every comparator was statistically significant (p<0.001).
The result above is a property of the system design — a domain-specific model, multi-agent reasoning, and retrieval from a curated knowledge base, on one path.
A domain-specific model trained on more than a million real, de-identified case histories spanning roughly half a century of practice, curated by clinicians — so its reasoning reflects observed case patterns rather than general web text.
Each likelihood ratio, red flag and contraindication is entered and reviewed by clinicians, so every retrieved item carries a traceable source rather than unattributed web text.
A standing pipeline aligns the knowledge base with current guidelines and literature, so outputs are not bound to a fixed training cut-off.
A single reasoning path for any presentation. An orchestrator dispatches specialist agents — each grounded in the curated knowledge base — and verifies every claim before output. The four stages:
Age-band priors, context flags and archetype detection frame the case; life-threatening possibilities are raised first.
Posteriors update with each finding using likelihood ratios drawn from the database rather than model memory.
Tests and imaging are selected for discriminating power, limiting unnecessary investigations.
Each dose, differential and recommendation is checked against the curated database and guidelines before it is returned.
Each reasoning step resolves against a purpose-built base of cited signs, symptoms, likelihood ratios, red flags and contraindications, authored by clinicians and kept current by the update pipeline.
A domain-specific model for the language and logic of medicine. In a blinded, head-to-head study, it was run against five leading general-purpose models on an identical cohort of 2,500+ complex, de-identified cases, each output graded against the confirmed final diagnosis by an independent panel of senior physicians. Primary and secondary outcomes follow.
| Table 1 · Primary outcome — diagnostic accuracy, 2,500+ cases, identical cohort | Correct | 95% CI | vs DIAGNIFY |
|---|---|---|---|
| DIAGNIFY Medical LLM structured expert mode | ~100%#1 | 99.8–100% | — |
| GPT-5.5 OpenAI | 83% | 81.5–84.5% | p<0.001 |
| Claude Opus 4.8 Anthropic | 83% | 81.5–84.5% | p<0.001 |
| Gemini 3.1 Pro Google DeepMind | 75% | 73.3–76.7% | p<0.001 |
| Grok 4.20 xAI | 58% | 56.1–59.9% | p<0.001 |
| DeepSeek V4 DeepSeek | 42% | 40.1–43.9% | p<0.001 |
| GPT-4o historical reference · 2025 study, not re-tested | 49% | — | — |
| Unaided physicians historical reference · 2025 study, not re-tested | 20% | — | — |
Pre-specified subgroups: ultra-rare, multi-system, safety-critical misses, and investigation efficiency. The margin over comparators was larger in these subgroups than in the primary outcome.
Blinded grading, paired testing on an identical cohort, and cases post-dating the models.
A photograph of a 12-lead ECG contains no direct access to the underlying voltages; the diagnostic signal is encoded in the printed tracing. General-purpose vision-language models can describe such an image but interpret it unreliably — 3–30% accuracy in our testing, with unmistakable findings missed. The system instead reconstructs the underlying signal from the image and classifies it with a specialist model.
Four stages convert a photograph or scan of a printout into a structured cardiology read. Each stage performs one function; classification is performed by a trained model rather than a language model.
The digitiser that won the PhysioNet 2024 Challenge — an nnU-Net model — reconstructs the 12-lead waveform from a photograph or scan of the printout, recovering the underlying electrical signal.
A convolutional neural network trained on PTB-XL — 21,000+ cardiologist-labelled ECGs — classifies the reconstructed signal across 23 diagnostic classes: rhythm, bundle-branch blocks, chamber enlargement and ischaemia territories among them.
A vision model produces a structured, human-readable report — rate, rhythm, axis, intervals, morphology — constrained to the classifier's findings, so it describes rather than introduces competing diagnoses.
Findings are cross-referenced against a 5,102-topic ECG teaching corpus (LITFL, ECGpedia, Dr Smith's ECG Blog and others), so the output is anchored to established literature rather than model memory.
After fine-tuning on 15,000 real-world-style ECG images, the specialist model reaches top-1 70% / top-3 89% against a cardiologist gold standard — with better handling of everyday phone photos. The system is positioned as a second read, not a replacement for the clinician's interpretation.
Constraints developed with cardiologists are enforced in the pipeline to prevent specific high-consequence errors, independent of the classifier's confidence.
It's flagged for formal Sgarbossa assessment — never asserted as a reason to activate the cath lab.
Discordant ST changes are correctly treated as expected in LBBB, not mistaken for acute ischaemia.
Interval and measurement estimates are labelled as estimates; any value failing a physiologic sanity check is suppressed rather than shown wrong.
Training uses reconstructed rather than idealised signals: ECGs are rendered to images, digitised back, and the model is trained on that reconstructed signal, matching the production input distribution. The latest model was fine-tuned on 15,000 ECGs rendered with realistic phone-photo conditions — rotation, noise and contrast variation — so it handles messy real uploads, not just clean scans. Each pipeline change is evaluated against a held-out gold set before release, and nothing ships unless it beats the previous version; one recent optimisation was rejected on failing this gate rather than shipped.
Most electrocardiograms exist only as paper or images and are not machine-readable. Reconstructing the signal from an image makes these tracings accessible for interpretation and research.
Large volumes of ECGs exist only on paper in institutional records and the published literature. Reconstructing them to signal makes historical cohorts machine-readable — enabling longitudinal studies, rare-arrhythmia datasets, and retrospective analyses not previously feasible.
The pipeline requires only a photograph of a printout — no proprietary export or on-site cardiologist — which is relevant to telecardiology and screening in settings where the digital signal is otherwise unavailable.
The model runs serverless at approximately $0.02 per ECG and scales to zero when idle, keeping the marginal cost of interpretation low at population scale.
Structured ECG findings are passed to the reasoning engine evaluated above (~100% across 2,500+ cases), where they update the differential and work-up under the same grounding constraints. A photographed printout is reconstructed, interpreted, and incorporated into the whole-patient assessment.
Dr Diagnify is in testing for non-urgent routine consultations and chronic-disease reviews — a telehealth-style workflow intended for settings with limited clinician access, including rural and remote communities. It applies the same reasoning architecture described above.
The voice pipeline used in DIAGNIFY's patient triage takes a structured history — red flags first, then the discriminating questions relevant to the presentation.
Limited to checks that are safe and reliable at home — temperature, pulse, a rash, neck stiffness, the FAST stroke check — each explained, recorded, and incorporated into the case.
The camera is used only at the examination step — a single frame, analysed on the spot, then incorporated into the reasoning as any other finding.
The consultation runs on the reasoning architecture described above — full differential, work-up logic, and a structured assessment — not a reduced variant.
The system is intended to extend clinician reach where access is limited, under human oversight — not to operate unsupervised.
Inference runs on GPUs deployed in Sydney. The models that reason over a case never execute offshore.
All data is stored in an Australian (Sydney) data centre. Nothing leaves the country — patient information stays onshore, end to end.
Data is encrypted end-to-end — in transit and at rest — so records are protected at every step of the consultation.
Decision support only, currently in testing. DIAGNIFY and Dr Diagnify are designed to support and, in time, extend the reach of qualified health professionals — not to replace a real consultation without one. Every output must be reviewed and verified by a registered practitioner before any clinical decision. These results describe performance on a focused benchmark and a clinician-confirmed validation set; they are not a regulatory approval, and formal external validation and regulatory clearance are required before any unsupervised clinical use. In an emergency, call 000.