The science
Six innovations behind the map.
Most diagnostic AI gives you a ranked list. TerrainDx gives you a navigable map of clinical possibility — calibrated against what's known and what's missing, with the reasoning visible at every step.
Six innovations make that possible. What follows is a tour — enough to see why the system behaves the way it does. The full math, calibration data, and worked examples are in the technical white paper, available on request.
The map: visual chunking, not a ranked list.
Working memory holds 5–7 named chunks reliably; ranked lists past that point look readable but aren't actually processed4. A complex case has 40–120 plausible diagnoses — far beyond what either patient or clinician can hold from a list.
TerrainDx organizes the diagnostic landscape into territories of ~7 related diagnoses, each named for what its members share clinically. The eye reads the map as a place; the mind navigates by chunks.
Bayesian likelihood ratios for every probability.
Every probability on the map is a real Bayesian posterior — not LLM intuition. Each finding carries a calibrated likelihood ratio anchored to published evidence; the posterior is normalized across the full landscape and tier-anchored against population prevalence.
Alt-click any probability in the live tool and the chain unfolds: prior, every LR contribution, posterior, and the source each LR was anchored to. No black box, because there is no box.
Factor-X: estimating the unknown using ecology.
Every clinical case has a hidden dimension: the diagnoses that weren't generated. Most diagnostic AI ignores it. TerrainDx estimates it directly, using methods from ecology that estimate species richness from incomplete sampling5.
The intuition: imagine ten graduate students surveying a forest. The number of species they collectively find isn't the answer — the structure of how often each species was seen tells you how much they probably missed.
Case A — saturated
10 students return with 10 individuals across 3 species, well-distributed.
Estimated unknown: small.Case B — sparse
10 students return with 30 individuals across 30 species — each seen exactly once.
Estimated unknown: roughly an order of magnitude larger.The same principle applies clinically. TerrainDx runs the case through a panel of specialty-trained agents, then estimates from the structure of their output how many plausible diagnoses likely sit outside what was generated. The output, normalized as Factor-X, appears to the user as “the wider view” — an explicit acknowledgment of what's still unsampled.
The psychology of the unknown — different for each user.
Putting the unknown explicitly on the map matters — but it matters differently for the clinician and the patient.
For the clinician
An explicit estimate of what the system likely missed prevents premature closure. When Factor-X is high, it's a signal to broaden the differential or seek additional history. When it's low, it's reassurance that the landscape is approximately complete — the prompt to engage System 2 only when needed.
For the patient
Counterintuitively, an explicit unknown is reassuring. It places the “what if it's something nobody's thought of?” fear into a bounded probability rather than an unbounded dread. Honesty about what we don't know is more comforting than a confident-sounding list that hides the gap.
Capturing case urgency: Severity max, Steptime, and Clinical Threat Density.
Standard triage collapses risk into one dimension. Real clinical risk is two: how bad it could get and how fast. TerrainDx captures both and combines them into a metric that reflects actual decision pressure.
| Variable | Definition | What it captures |
|---|---|---|
| Severity max | Worst plausible functional outcome | How bad this could be if it goes the wrong way (death, permanent harm, hospitalization, vs. self-limiting). |
| Steptime | Days to next clinically meaningful step | How quickly a decision must be made — minutes for a STEMI, weeks for chronic fatigue workup. |
| CTD | Clinical Threat Density | The composite that reflects actual decision pressure. |
The map renders CTD visually: territory color shifts as the metric rises; elevated diagnoses flag for attention. A 10% chance of an emergency at minutes-to-step is a different decision than the same probability at months-to-step — flat triage scales miss the gradient; CTD captures it. The clinician sees decision pressure as a place on the map, not as another numerical alarm.
Patent pending — Innovation 05Diagnostic Leverage Index and case perturbation.
Two innovations close the loop on bedside utility.
Diagnostic Leverage Index (DLI)
For every potential next step — a question, a finding, a test — TerrainDx computes how much that step would shift the differential. High-DLI questions surface as “Key Questions”; low-DLI ones are de-prioritized regardless of how routine. As findings come in, the ranking updates in real time.
Case perturbation
Patient stories are imperfect. Conventional decision-support takes the case as given and breaks silently when the input is off. TerrainDx perturbs the input deliberately and re-runs the analysis to see how much the differential shifts. Diagnoses that survive are robust; diagnoses that vanish under small perturbations are fragile and get a “needs confirmation” treatment.
Experienced clinicians do this implicitly (“if she's wrong about the timing, this changes”). TerrainDx makes it a visible quantity attached to every diagnosis on the map.
Patent pending — Innovation 06A more detailed technical white paper covering each of these innovations, with citations and worked examples, is available on request.
Bibliography
- Newman-Toker DE, Peterson SM, Badihian S, et al. Diagnostic Errors in the Emergency Department: A Systematic Review. Agency for Healthcare Research and Quality (AHRQ) Comparative Effectiveness Review, 2022; updated estimates in Newman-Toker et al., BMJ Quality & Safety 2023. ↑ back
- Gallup. Honesty & Ethics in Professions, annual survey 2020–2024 — physicians’ “very high / high” rating fell from 77% (2020) to 53% (2024). ↑ back
- Shrank WH, Rogstad TL, Parekh N. Waste in the US Health Care System: Estimated Costs and Potential for Savings. JAMA 2019; 322(15): 1501–1509 — ~$210B+ annually attributed to overtreatment, failure of care coordination, and failure of care delivery. ↑ back
- Miller GA. The magical number seven, plus or minus two: some limits on our capacity for processing information. Psychological Review 1956; 63(2): 81–97. ↑ back
- Chao A. Nonparametric estimation of the number of classes in a population. Scandinavian Journal of Statistics 1984; 11: 265–270 (the original Chao1 estimator); bias-corrected variants subsequently developed by Chao and colleagues for sparse-sampling regimes. ↑ back
- Foster PN, Klein JR. Defining excellence: next steps for practicing clinicians seeking to prevent diagnostic error. Journal of Community Hospital Internal Medicine Perspectives 2016; 6(4) — the peer-reviewed call to redesign the diagnostic process that TerrainDx implements.