Evidence
What supports Synmi today.
A preliminary summary of the evidence behind the product: the study we cite, the architecture we build on, and the limits we state plainly. Advisory only. Each output is reviewed by a qualified clinician before it becomes care.
Cited study
Already demonstrated in Kenyan primary care: a 2025 study across 39,849 patient visits in 15 clinics found 16% fewer diagnostic errors and 13% fewer treatment errors with clinician-facing AI support. Source: Penda Health and OpenAI, 2025.
- Visits
- 39,849 visits
- Clinics
- 15 clinics
- Diagnostic errors
- 16 % fewer diagnostic errors
- Treatment errors
- 13 % fewer treatment errors
Scope note — Kenyan primary care, 2025. Third-party findings: not our deployment, not our measurement. The study describes clinician-facing AI support in that setting; it does not evaluate Synmi.
Safety architecture
Mechanisms with ADR roots, not promises. Advisory only, human-reviewed.
Human-in-the-loop
Every advisory output passes through clinician review gates on every panel before it becomes care. The clinician remains responsible for all clinical decisions.
Data stays with the institution
Patient data remains in the healthcare institution's controlled environment. Synmi is designed as a secure connection to that environment, not a central repository.
Every claim is cited
Evidence and guideline retrieval outputs carry their underlying source citation, so the clinician can verify every recommendation.
No autonomous decisions
Synmi does not diagnose, does not prescribe, and does not finalise records. It informs the professional who does.
Validation posture
Software assurance, as labeled: the automated test suites, the accessibility gate, the frozen synthetic demo case, and deterministic fallbacks that keep every surface working when a model capability is unavailable. This is software assurance — it is not clinical validation, and it is never presented as such.
Limitations
The limits sit on the same page as the claims, by construction.
Common-presentation tuning
Advisory support is tuned around common presentations. Unusual cases stay firmly under clinician judgment; the system flags its own limits rather than guessing past them.
Input budget constraints
Vision and speech input run under explicit token and time budgets, with quality gates and retake prompts. A soft photo or a noisy ward is a stated limit, not a silent failure.
Open gates pending pilot
Plausibility ranges and pending ratifications stay open until facility pilots close them. Nothing here is presented as settled that is still gated.
What comes next
This page is a preliminary summary. A full evidence review follows facility pilots and the sit-down research revamp, and this summary will be revised against what the pilots measure.
Synthetic demonstration data only — every demo runs on invented records. Nothing here writes real data, and every advisory output stays under clinician review.