For ICU intensivists, hospitalists, and clinical-AMC physicians making individualized treatment decisions. The risk surface arrives with a Rosenbaum-Γ bound printed beside every recommendation — "this estimate is fragile to unobserved confounding at Γ = X" — so the clinician makes the call with the uncertainty visible. Targeted FDA Class II SaMD via 510(k) with a Predetermined Change Control Plan. Available post-2027 after clinical-deployment validation.
A direct clinician-facing decision-support tool that influences individual patient treatment is a Class II Software-as-a-Medical-Device under the FDA framework. It requires a 510(k) submission with a substantial-equivalence claim, a Predetermined Change Control Plan (PCCP) so the model can be updated post-clearance, and clinical-deployment validation evidence that the Class II workflow actually moves outcomes vs. usual care.
That sequence — Q-Submission Pre-Sub → PCCP draft → 510(k) submission → clinical validation → clearance — takes 18-24 months at minimum, even with a Breakthrough Device designation (which itself only accelerates ~12% of designated devices to clearance per a 2024 JAMA Internal Medicine analysis).
Building this on top of an unproven methodology platform is the wrong sequencing. The right sequencing is: ship pharmacovigilance and real-world-evidence in 2026 (under the 21 CFR § 3060 non-device CDS exemption), accumulate the clinical-deployment evidence base, then formally enter the 510(k) pathway with a Pre-Sub in 2027.
The underlying causal-inference engine, calibration infrastructure, and audit ledger that will power the physician-CDS product are already in production today — they're what powers the 2026 PV and RWE products. The 510(k)-specific layer (clinical workflow UX, EHR integration, alert pathway, intended-use labeling, PCCP modification protocol) is the 2027+ build.
Every treatment-effect estimate already arrives with Rosenbaum-Γ sensitivity bounds today. The technology is in production on the RWE and PV products; the physician-CDS application surfaces those bounds beside per-patient recommendations.
Beta-Bayes and isotonic calibration with per-workload auto-selection; ECE under 0.005 on MIMIC-IV W1 mortality. Calibration is non-negotiable for a SaMD — the probability has to mean what it says it does.
The system knows when it doesn't know. Out-of-distribution patients trigger abstention rather than over-confident extrapolation. This is the single biggest regulatory objection to ML in clinical care — addressed at the architecture level, not via warning labels.
Every recommendation, every patient-record access, every override committed to a SHA-256 chained ledger. Already in production today — the SaMD product inherits it.
No LLM in the prediction path. The model's input feature space is frozen at training; modifications happen via the PCCP's bounded modification protocol, not via stochastic drift. This is exactly what the December 2024 PCCP final guidance rewards.
Every prediction is reproducible from its inputs + the model artifact hash + the platform git commit. FDA reviewers can independently verify any output without access to live data.
Stage 1 (Q3 2027): Q-Submission Pre-Sub to FDA CDRH for substantial-equivalence claim discussion. Predicate-device identification (existing FDA-cleared ICU CDS tools). Intended-use statement covering treatment-selection in a specified clinical context (initial scope: ICU anticoagulant choice, given the W12 + W13 benchmark depth).
Stage 2 (Q4 2027): Predetermined Change Control Plan drafted to the December 2024 final guidance: Description of Modifications, Modification Protocol, Impact Assessment. The frozen-feature deterministic architecture bounds the modification space cleanly.
Stage 3 (2028): 510(k) submission. Class II software function, no clinical investigation required if predicate-device equivalence holds. Breakthrough Device application contingent on credible clinical-benefit evidence from pharmacovigilance/RWE deployment.
Stage 4 (2028–2029): Clinical deployment + post-market surveillance. EU AI Act Article 6(1) (device-embedded AI) conformity with Notified Body if deploying in EU; standalone high-risk AI track if separate from the device CE mark.
Watch the product walkthrough at rosenbound.ai — three moments that define the platform: the Cognitive Validation Report refusing incoherent data, the live Γ-bound sensitivity visualization, and the reproducibility certificate generated on every study. The full platform stays gated for Founding Partners.
Watch the preview →
pip install rosenbound
— Official Python SDK for programmatic access: cohort upload, sensitivity-bounded study runs, and reproducibility certificate retrieval. Apache 2.0; Pydantic v2 typed; py.typed for IDE autocomplete + mypy. Platform access gated by Bearer token + RBAC + tenant scoping — the SDK is open, the audit substrate is not.
ICU intensivist groups at academic medical centers who'd like input on the intended-use statement, the initial clinical scope, and the PCCP modification protocol are welcome to engage now as Q-Sub partnership candidates. The 510(k) submission strength is directly proportional to the clinical-deployment validation depth — AMC partnerships in 2027 set the trajectory.