Patient context, made useful
Bring the relevant history, outcomes, and constraints into one reviewable view.
In development with practicing physicians
Make treatment tradeoffs easier to see and discuss.
We're exploring clinician-controlled decision support that compares eligible treatments using patient context, clinical evidence, expected outcomes, and uncertainty.
No sales pitch. No patient information. No product commitment.
Posterior outcome comparison
STAR*D treatment set · Demonstration output
Example decision context
Adult with nonpsychotic major depression after no remission with citalopram
Eligible option
020%40%
P(best)Bupropion SR
NDRI
21% · 90% CrI 14–30%
Sertraline
SSRI
18% · 90% CrI 11–27%
Venlafaxine XR
SNRI
25% · 90% CrI 16–35%
Intervals overlap. The model expresses uncertainty; it does not identify a clinically reliable winner.
Simulated patient profile and posterior draws. Treatment set and published benchmarks are from STAR*D. Not for clinical use.
Designed for hospital deployment
Evidence and uncertainty visible
Clinician remains in control
Built for patient conversations
The clinical problem
Treatment choice is rarely a simple lookup.
Guidelines identify reasonable options, but they cannot resolve every patient-specific tradeoff. Clinicians still have to connect prior response, comorbidities, safety, treatment burden, and what matters to the patient.
What stan4bart returns
A distribution, not a verdict.
The model produces posterior draws for expected and predicted outcomes. The product layer can summarize those draws as estimates, credible intervals, and treatment contrasts while preserving the uncertainty underneath.
- Expected outcome
- Posterior draws of E[Y | patient, option]
- Predicted outcome
- Posterior predictive draws, including outcome noise
- Treatment effect
- A derived distribution from paired option draws
Pairwise remission effects
Posterior mean and 90% credible interval, percentage points
−20No difference+20
Venlafaxine XR vs sertraline
+7 pp · −4 to +17
Bupropion SR vs sertraline
+3 pp · −6 to +11
Venlafaxine XR vs bupropion SR
+4 pp · −5 to +13
Trial benchmark
Published remission rates were 21.3%, 17.6%, and 24.8%.
Interpretation
Every interval crosses zero, so the direction of each contrast remains uncertain.
Evidence you can inspect
See what supports each estimate, how well it applies, and where important gaps remain.
A better patient conversation
Translate benefit, harm, burden, and uncertainty into choices patients can discuss.
Designed around clinical review
One view of the options. Not one answer from an algorithm.
The intended output is a reviewable comparison that supports independent clinical judgment—not an autonomous prescription.
- 01
Bring forward context
Read only the patient information needed for the decision.
- 02
Confirm eligible options
Apply explicit indication, contraindication, and safety rules.
- 03
Compare possible outcomes
Estimate benefit and harm for each option, including uncertainty.
- 04
Review and discuss
Inspect the basis, then choose what to share with the patient.
Proposed deployment architecture
Designed to keep patient data inside the hospital.
The target architecture runs the application and model within the hospital-controlled environment. Patient data, predictions, and clinical logs would remain onsite.
On-premises or hospital-private-cloud deployment
Subject to institutional review and validation.
Narrowly scoped, read-only EHR access
Subject to institutional review and validation.
Signed and versioned model updates
Subject to institutional review and validation.
No patient data sent to Causal Care servers
Subject to institutional review and validation.
Founding team
Building Causal Care alongside clinicians.
We're forming a physician design group to challenge the assumptions, workflows, and evidence behind this concept.
For practicing physicians
Help us design this around how treatment decisions actually happen.
We're speaking with physicians who regularly choose among multiple reasonable therapies. Share what would be useful, distracting, or unsafe in a 20-minute research conversation.
Request a conversationNo patient-identifying information requested.