Morning report in your pocket.

A differential you can think with — plus an AI your patients use before the visit, so they arrive already on the map, not on Google.

A note from the founder

The Rub

In a sentence

Generative AI is changing how patients arrive in your office. TerrainDx tries to put you and the patient on the same map — with the unknown explicitly included — so you can spend the visit on what you're best at.

Skip to the live demo →
Paul Foster, MD, in his exam room — stethoscope and credentials visible

We've known for decades that educating patients and sharing decisions improves outcomes. We've also known it eats time. That tension is old.

What's new is generative AI. Patients now arrive misinformed and anxious, with details in hand that make us look uninformed — and that often force unwise use of resources and treatments.

I built TerrainDx as a doctor living the same experience, wondering if we could regain control of the narrative. Could we give patients an understanding of their diagnostic possibilities in a way that actually helps us take care of them?

For many patients, I think the answer is yes.

Sharing the 70 diagnoses that could explain their symptoms gives the patient a sense of the overwhelming probability that those symptoms are benign — and that we have the time to work through options sensibly. More importantly, in our testing the AI surfaces details of the history that help point the patient in the right direction and help us reach a diagnosis. It does surface zebras, but it explains the risk factors and time patterns that lead there, past the many more common alternatives. For the few scary diagnoses, or ones that progress rapidly, it points out the signs of progression that signal the need for quicker evaluation. This may be especially useful for the patient who previously would have come to us demanding specific testing well before their pretest probability made that testing meaningful.

One of the design journeys I most valued was the question that produced "the wider view" lake — the unknown node on every map.

Even with 70 diagnoses on the list, what's the chance the real one isn't?

Clinically, I know I typically underestimated it, and I'd never have put it on the table with a patient. I found some guidance in the ecology literature. The math used to estimate species in a forest after a brief survey by ten grad students turns out to share many properties with estimating how many diagnoses ten AI agents missed in a patient history. Most of the unknown comes from parts of the story that are wrong or left out. Some comes from misunderstood science. Some isn't yet known to science.

Having the unknown on the table — 10%, 30%, 50% — takes the pressure off. It also emphasizes what we're actually good at: looking at a well-organized array of diagnoses and intuiting which need immediate attention, which need exploration, and which to keep an eye on.

I've found myself changing the way I present myself to my patients. I'm no longer racing to find the answer right away — though sometimes I still do. TerrainDx helps me show off what I'm best at, which is helping the individual in front of me find a way forward. I think most of you will find the same.

A few honest limits. The tool will sometimes waste time or open Pandora's box. It is not a genius diagnostic engine. It has far less subjectivity and far fewer hallucinations than most AI tools, but it can be incorrect or misled. It is evidence-based, and every decision is traceable — every diagnosis points to external references for further perspective.

The full analysis takes a while, so it may be best to have your patient enter the case ahead of time. We're working with partners on quicker ways to capture data while you're doing other things.

The four tools are designed so you can use TerrainDx in just the amount that's helpful:

  • TerrainDx Chat — for quick answers, dose checks, evidence verification
  • Key Questions — quick and compulsive use as you close your notes, to think briefly about how you might have gotten it wrong
  • The Condensed Case — for sharing with colleagues
  • The Terrain Map — for brainstorming a case where you feel you might be missing something

Take some time with the demo, or register and enter a case of your own. If you have knowledge-hungry patients, or complex patients where something feels missing, have them go home and explore. I think you'll find it well worth the time.

Signature of Paul N. Foster, MD

Paul N. Foster, MD

Internist · Former Internal Medicine Residency Program Director
Assistant Clinical Professor, Hackensack Meridian School of Medicine
Founder, TerrainDx

An example case

Look at how TerrainDx renders a real case.

A complete diagnostic landscape from the production pipeline — the same view a beta clinician sees. Open it, alt-click a few diagnoses to inspect the Bayesian math, and read the Key Questions.

A real production case — full landscape with action lanes, Key Questions, and provenance chains visible on every probability and likelihood ratio.

No login required.

See an example case →
⏱ About 15 minutes for a new case. Faster if it matches our template library. Quality over speed — eight specialist agents with Bayesian calibration. Best workflow: patients pre-enter through patient.terraindx.co, the map builds while they wait, you open a completed landscape when you're ready.

The literature on diagnostic error

Diagnostic error contributes to ~10% of US patient deaths and 800,000 serious harms each year. (Singh et al., BMJ Qual Saf 2014; NAS 2015)

Graber's taxonomy describes three overlapping failure modes: no-fault errors (atypical presentations, rare diseases), system errors (communication failures, inadequate follow-up), and cognitive errors — which account for roughly 75% of cases, often in combination with system factors. (Graber et al., Arch Intern Med 2005)

Dual process theory explains why: System 1 thinking (fast, pattern-based) handles most clinical encounters efficiently, but it's vulnerable to premature closure and anchoring — especially when the presentation partially matches a familiar pattern. The challenge isn't engaging System 2 on every case; it's knowing when to. (Croskerry, Ann Emerg Med 2009)

LLMs introduce new versions of these same problems. Studies show they anchor on early information in a clinical vignette, generate plausible but unsourced reasoning, and can reinforce existing biases rather than challenge them. (Kanjee et al., JAMA 2024)

A more recent stress test makes the gap concrete. A Mass General team (JAMA Network Open, April 2026) tested 21 leading LLMs against complete vs. disturbed cases. Accuracy on complete clinical data: roughly 85%. Accuracy when patient information was incomplete or sequentially revealed: as low as 20%. Patients live in exactly that gap — stories arrive partial and out of order. A useful tool needs to add diagnostic range while keeping the reasoning transparent and challengeable.

"When we disturbed the stories — removed data, put in lies, pointed at another diagnosis — TerrainDx found the same diagnoses. It was not thrown off." — Paul Foster, MD, founder

Four tools, each with a distinct purpose

TerrainDx Chat — informed by the Bayesian math

A medical conversation grounded in the full probability distribution of the case. As you think through the differential, the chat draws on calibrated likelihood ratios and your reasoning style — not generic LLM output. Quick answers, dose checks, evidence verification. Think of it as talking through a case with a colleague who has already computed the differential and can show you the math when you want it.

Key Questions — where is the evidence weakest?

The differential pit stop — a quick stop as you close your notes, to catch how you might have gotten it wrong.

Identifies the history elements most likely to shift the differential, the findings with the highest diagnostic leverage, and the pivots that could change your leading diagnosis. On straightforward cases this is where it's most useful — the case that looks simple is exactly when anchoring is hardest to catch.

The Condensed Case

A structured clinical summary — chief complaint, key positives, key negatives, timeline — distilled from the patient's narrative. Built for sharing with colleagues and for quick orientation when you pick up a case. When the patient has pre-entered their story through patient.terraindx.co, the condensed case is ready before they walk in.

The Terrain Map

See the zebras — where they sit, in context — so you can rule them out and move on.

60 to 120 diagnoses organized into territories and assigned to action lanes: what needs empiric treatment, what key questions to answer, what to monitor, what's been cleared. Not a ranked list that drops off after 5 items — the full differential in its widest form. For brainstorming a case where you feel you might be missing something. Each diagnosis links to the Bayesian math behind it. The map responds to new data — labs, findings, patient diary entries — and the entire landscape recalibrates.

Alt-click any data point

Pulmonary Embolism — 4.2%
Community Pneumonia — 12.8%
Sarcoidosis — 1.7%

Provenance

Prior: 0.3% (population prevalence, age-adjusted).
LR+ bilateral hilar lymphadenopathy: 8.2.
LR+ dry cough + fatigue: 2.1.
Posterior: 1.7% after calibration.

PubMed: Systematic review →
ATS Guidelines 2024 →

Every number is traceable

Alt-click any diagnosis, probability, likelihood ratio, or lane assignment to see the evidence chain — the Bayesian math, the reasoning at each step, and links to published literature.

If a number doesn't look right, trace it and challenge it. That's the point. The tool supports your independent judgment by making its reasoning visible, not by asking you to trust it.

Patients build their map before the visit

Through patient.terraindx.co, patients enter their case before the appointment. The map builds while they wait. By the time you open the case, you have a completed landscape and a structured condensed case — and the patient has been exploring their map instead of doomscrolling WebMD.

See what they're seeing

The patient's map view is available to you — their neighborhoods, their questions, their diary entries. Understand their perspective before the conversation starts.

Structured history

The condensed case is organized: chief complaint, key positives, key negatives, timeline. Built from the patient's own words, structured for clinical use.

Go deeper

The story behind this

25 years of internal medicine and diagnostic-error work. Why a physician decided to build it.

How to use the tools

Walkthrough of all four tools with case examples. Coming soon.

Clinician experience

Physicians using TerrainDx in practice. Coming soon.

About TerrainDx

The reasoning layer for diagnostic medicine, and how the patient and clinician tools work together.

Common questions

Things clinicians want to know.

Is this FDA-cleared? Does it meet the CDS exception?
TerrainDx is structured to operate under the Clinical Decision Support exception in Section 520(o)(1)(E) of the FD&C Act. The output is informational and traceable; every probability is backed by an inspectable Bayesian chain anchored to published evidence, the recommended action lane is a suggestion not a directive, and the clinician retains full independent judgment. We are preparing a Q-Sub voluntary pre-submission with the FDA.
What's the evidence behind the likelihood ratios?
Every LR is anchored to a published source — clinical studies, validated diagnostic criteria, or peer-reviewed databases. Alt-click any probability in the live tool and you see the chain: prior, every LR contribution, posterior, and the citation each LR was anchored to. There is no opaque LLM intuition driving the math.
How long does a case actually take?
10–20 minutes from case entry to a complete landscape. Eight specialty agents work in parallel, then the system calibrates each diagnosis against the published evidence. Best workflow: have the patient pre-enter through patient.terraindx.co — the map builds while they're in the waiting room. You walk in to a completed landscape.
Will it integrate with my EHR?
Not yet. v1 is a standalone web tool. The Condensed Case feature produces a structured summary you can paste into your note — we've designed it to be MyChart- and Epic-friendly. Direct EHR integration (FHIR-based) is on the roadmap; we're scoping partner conversations.
What about liability if I follow a recommendation?
TerrainDx is decision support, not a directive. Every output is a starting point for clinical reasoning — the action-lane suggestions surface what's worth considering, but the clinician's judgment governs. Because every probability and every Key Question is traceable to its evidence, the tool produces a more defensible reasoning trail than a typical encounter, not less.
My patient already shopped me with WebMD — how does this help?
This is the case TerrainDx is built for. When a patient arrives convinced they have something rare, the full landscape gives you a structured way to validate their concern (it is on the map), put it in proportion against everything else (most maps have 60–120 possibilities), and redirect to the high-leverage next questions. You're working from the same picture, not arguing across one.
Where does the patient input come from? Is it reliable enough to trust?
Patients enter their own story; the system also runs case perturbation — deliberately varying findings, timelines, and adjacent details to identify which diagnoses are robust to data error and which are fragile. Fragile diagnoses get a "needs confirmation" treatment automatically. You see, on the map, exactly which conclusions depend on the data being right.
What does it cost?
Free to view a patient's shared map — no account needed. Clinician subscriptions for actively running cases yourself (entering vignettes, using all four tools) are launching post-beta; pricing details available on request via the Beta Access form below.
More questions?
We answer every note: our contact form. Deeper materials — technical white paper, defensibility brief, validation data — available on request via the about page.

Hear when TerrainDx launches for clinicians.

We're refining the four tools and the Bayesian provenance layer with a small group of clinicians before broader release. Send a note and we'll keep you posted.

Request information on the release of TerrainDx paul@terraindx.co · we read and reply to every note

About the founder

TerrainDx is built by Paul Nathan Foster, MD, FACP — a practicing internist at Englewood Hospital and Hackensack University Medical Center, former Internal Medicine residency program director, and Assistant Clinical Professor at Hackensack Meridian School of Medicine. Co-author of "Defining excellence: next steps for practicing clinicians seeking to prevent diagnostic error" (J Community Hosp Intern Med Perspect, 2016) — a peer-reviewed call for redesigning the diagnostic process. TerrainDx is the redesign.

Meet Dr. Foster →

Clinical decision support: TerrainDx is designed to support, not replace, independent clinical judgment. It meets the criteria for clinical decision support under Section 520(o)(1)(E) of the Federal Food, Drug, and Cosmetic Act. All diagnostic information, probabilities, and action lane assignments should be evaluated by a licensed clinician in the context of the individual patient. TerrainDx does not diagnose, treat, cure, or prevent any disease or condition.