18
synthetic patients, each internally consistent across a decade or more
Five interlocking frameworks — hypertension management, longitudinal care orchestration, lab-pattern description, multi-disease next-best action, and hospitalization reconstruction — built on a shared epistemic discipline and validated against a deeply instrumented synthetic patient cohort.
Most clinical decision support fails in a predictable way: it renders an opinion the moment it has any data at all. Rule engines fire on a single lab value; LLMs confabulate a confident narrative from a sparse chart. The unifying design principle across this portfolio inverts that default. The system must earn the right to make a recommendation, and the process of earning it is explicit, sequential, and inspectable. Every framework in the portfolio implements some version of the same ladder:
Two architectural consequences follow, and both matter to anyone accountable for deploying AI in a regulated clinical enterprise. First, deterministic gates run before any generative step — the LLM is never the arbiter of whether there is enough data; it operates only inside a lane the gates have opened. Second, "I cannot determine this" is a first-class, schema-validated output, not a failure mode. A system that can articulate what it doesn't know is auditable; one that always answers is not.
Each framework below was built, demonstrated as a working HTML application or specification, and exercised against the synthetic cohort described in Section 3. The table gives the conceptual leap each one contributes, a concrete example of who uses it, and — in the final column — where the value lands and, more importantly, whose decision-making has to change for that value to be realized in live workflows. The last point is deliberate: none of these tools creates value by existing; each creates value only when a named role agrees to decide differently because of it.
| Framework | Novel leap / logical understanding | Example use — who, and what they gain | Value opportunity — and whose decisions must change |
|---|---|---|---|
| Hypertension Management Frameworkchecklist + next-best-action demo app · v2.0 | The original proof of the "earned recommendation" pattern. Structured buckets (data adequacy → context/confounders → medication management) with gate conditions: downstream sections stay collapsed until upstream questions are answered. Encounter-type branching (initial vs. prior-data vs. longitudinal follow-up) recognizes that the same BP value warrants different reasoning depending on what is already known. Every item carries visible rationale, disagreement notes, and evidence gaps — transparency as a feature, not documentation. | A nurse or care coordinator works a panel of dialysis patients with uncontrolled BP. Instead of an alert saying "BP high — consider med change," the tool shows which prerequisite is unmet (e.g., no home readings, unassessed adherence, white-coat confound unruled). Who: RN/coordinator, PCP, nephrologistThe gain is a defensible answer to "why is this patient not yet at goal" — often "we haven't earned a med change yet, and here is the missing datum." | Panel-level BP control, fewer reflexive titrations, documentation that survives audit. Highest-yield insertion point: dialysis rounding and CKD clinic pre-visit planning.Whose decisions must changeNephrologists/APPs must accept "not yet earned" as a legitimate output — deferring a med change becomes an ordered action, not inaction. RNs/coordinators must own data-completion (home BPs, adherence assessment) as the new first-line intervention. Medical directors must redefine "care gap" policy from threshold-breach to missing-prerequisite. |
| CKD Care Orchestration Frameworklongitudinal multi-module architecture | Generalizes the HTN pattern from a single visit to a decade-long trajectory. The unit of reasoning is not the encounter but the patient's position on the CKD→ESKD continuum: stage-specific tasks, modality/transplant/palliative conversations tracked as longitudinal threads, cadence logic (labs, visits, counseling), and reconciliation of external feeds (ADT events, discharge summaries, outside labs). Introduced an explicit encounter / event / illness-episode data model so that "what happened" is representable independent of any one document. | A nephrology practice or value-based-care team reviews a panel. The orchestrator shows each patient's trajectory placement and which longitudinal threads have gone silent — a modality conversation never revisited after a hospitalization, a vaccine series interrupted. Who: nephrologist, IDT, population-health analystThe gain is longitudinal accountability: gaps defined against the patient's own timeline, not against a static quality measure. | Earlier modality, access, transplant, and palliative decisions; fewer crash dialysis starts; readmission avoidance. Insertion point: monthly panel review and VBC care-management queues.Whose decisions must changeCare managers must re-prioritize outreach by silent trajectory threads rather than last-contact date. Nephrologists must let trajectory placement set the visit agenda. Population-health leadership must reallocate coordinator time toward thread reactivation, and VBC/payor partners must accept trajectory-anchored metrics alongside conventional quality measures. |
| Lab Summary Agent — Sequential Axis ModelNephAssist controlled vocabulary · v1.4 | Replaces the naive "trend arrow" with a gate-first descriptive pipeline: Axis 1 counts the data (n=0 says nothing; n=1 earns position only; n≥3 unlocks trend); Axis 2 reads draw cadence, so a hospitalization-clustered Hgb drop is recognized as event-driven rather than physiologic decline; Axis 3 separates population-anchored position from patient-anchored trajectory and per-lab, time-normalized velocity; Axis 4 (interpretation) is structurally quarantined and fires only when the data has earned it. A controlled vocabulary with forbidden phrasings ("moderately elevated," "concerning pattern") makes the output deterministic enough to audit — and a 12-exemplar few-shot set teaches the LLM when to stay silent. | A rounding nephrologist opens a 30-day lab summary. Instead of "Hgb trending down — concerning," they read: "Hgb 9.1, below ESKD range, falling — rate unreliable, hospitalization in window." Who: rounding clinician, anemia manager, downstream NBA modulesThe gain is trustworthy language: every adjective traces to a threshold, event-driven artifacts are pre-flagged, and the same axis outputs become machine-consumable inputs for the multi-disease modules. | Rounding time recovered, safer ESA/iron/MBD titration, one shared lab language across humans and machines. Insertion point: OneView rounding view; second-order value as the upstream data-sufficiency source for every NBA module.Whose decisions must changeRounding nephrologists must trust suppression — accept "all in typical range, no detail lines earned" instead of re-scanning full tables. Anemia/MBD program managers must key protocol actions to the controlled velocity terms. CMIO/informatics must govern the vocabulary as versioned policy, and compliance must sign the description/interpretation firewall as the audit boundary. |
| Complex Multi-Disease Framework (CMDF)17 module-agents + fluid/volume triage · v2 | Scales next-best action across concurrent illnesses without a monolithic engine. Seventeen disease/domain modules each run as an agent emitting one shared, JSON-Schema-validated report contract: data sufficiency, context sufficiency, findings with provenance, NBA, escalations, cross-module signals — and a mandatory cannot_determine field. The coordinator consumes only the contract, never a module's internals; a report missing provenance or claiming "cannot determine" without flagging concern fails validation. The fluid/volume triage pathway (DIAL/CKD) demonstrates the same discipline in an interactive Q&A form. Honesty is enforced by schema, not by prompt engineering. |
A patient with ESKD, heart failure, and diabetes generates competing priorities. The coordinator view shows each module's report side by side: which modules have earned a recommendation, which are blocked on missing data, and where cross-module signals conflict (e.g., volume module vs. BP module on the same weight change). Who: IDT, medical director, informatics architectThe gain for a technical leader: a governance surface — every recommendation is traceable to a validated report, and inter-module disagreement is visible rather than silently averaged away. | Prioritized cross-disease worklists and a standing governance surface where conflicting recommendations become visible and adjudicable. Insertion point: IDT care-plan meetings and medical-director oversight dashboards.Whose decisions must changeThe IDT must let the coordinator's ranked NBAs structure the meeting agenda. Clinical domain leads must each own a module's thresholds and cannot_determine criteria — named ownership per module, like Duke's ABCDS model. Medical directors must adjudicate cross-module conflicts explicitly rather than leaving them implicit, and IT architecture must adopt the report contract as an integration standard for any future module or vendor. |
| BRIDGE Hospitalization Reconstruction6-stage review heuristic + scoreable schema · v0.1 | Reframes chart review from "what does the record say?" to "what actually happened, and do I believe it?" Six stages — Baseline, Run-up, In-hospital evolution, Definitive anchors, Gestalt at discharge, Evaluate & interrogate — reconstruct the hospitalization as a diagnostic trajectory rather than a snapshot. Four outputs ordinary extraction never produces: the trajectory itself (ED impression → discharge diagnosis drift), calibrated certainty on every diagnosis, an independent coherence verdict (the tool may say "this doesn't add up"), and the signature known-unknowns list — the questions a careful clinician would still ask, surfaced because the chart is silent (the "grapefruit-juice problem": the causal fact that appears in no lab, image, or problem list). | A transition-of-care nurse receives a discharge summary for a returning dialysis patient. The BRIDGE reading flags that the discharge diagnosis drifted substantially from the ED impression, rates the final diagnosis "probable, not established," and ranks three unasked questions — including whether a missed-dialysis root cause was ever explored. Who: transition-of-care team, hospitalist QA, readmission-reduction programThe gain is a review that catches what the chart omits, scored objectively against ground-truth answer keys before any real chart is touched. | Readmission reduction, faster and sharper post-hospitalization reconciliation, hospitalist QA with teeth. Insertion point: the first 48 hours after discharge, when the transition call still shapes outcomes.Whose decisions must changeTransition nurses must restructure post-discharge calls around the ranked known-unknowns instead of a generic script. Nephrologists must treat a "coherence: doubtful" verdict as a trigger to contact the hospital team. QI leadership must adopt diagnostic-drift and unknowns-resolution as measured outcomes, and HIM/coding must consult certainty ratings before finalizing diagnosis queries. |
| Synthetic Cohort & Ground-Truth Infrastructure18 patients · canonical repository · 6 scored hospital packets | The insight that evaluation infrastructure is itself a framework. Eighteen synthetic patients with 10+ years of coherent history each — families, care teams, SDOH pressures, medication journeys, dated ADT events, hypertension episodes, comprehensive health evaluations — maintained in a canonical repository (master JSON + timeline + lab time-series + NBA snapshots) from which every app's data is a regenerated mirror. Six full hospitalization packets were drafted by different frontier LLMs, red-teamed and white-teamed, and shipped with machine-readable answer keys that encode not just facts but diagnostic drift as an extraction target. | Any new tool — an extraction agent, an NBA module, a summarizer — is benchmarked against the cohort before it sees production data. Because every case has a known answer, evaluation is objective, repeatable, and free of PHI. Who: informatics/AI engineering team, validation & compliance, vendor evaluationThe gain for a CIO: an in-house eval harness that turns "does this AI work?" from a vendor claim into a measured score — and a reusable asset for demos, training, and regression testing. | Objective, PHI-free pre-deployment evaluation; vendor accountability; a regression suite that catches silent degradation on model updates. Insertion point: AI intake/procurement and the validation step of every deployment.Whose decisions must changeCIO/procurement must make a cohort benchmark score a gate condition — no score, no pilot. Validation & compliance must accept answer-key scoring as formal acceptance evidence. Vendors must agree to run against the cohort as a condition of consideration, and clinical educators can shift onboarding and training onto the synthetic cases rather than live charts. |
None of the frameworks above could be honestly demonstrated on fabricated single-visit vignettes, because their central claims are longitudinal and contextual: that cadence distinguishes signal from artifact, that trajectory placement changes the right action, that silence in a chart is detectable. Testing those claims requires patients with enough depth that the frameworks' gates have something real to gate on.
The cohort spans etiologies, trajectories, and social contexts deliberately: diabetic and lupus nephropathy, cardiorenal decline, a missed-dialysis hyperkalemia crisis rooted in access barriers, immunosuppression-related sepsis. Each patient carries a family, a care team, and SDOH circumstances that materially alter what the "right" recommendation is — which is precisely what forces the frameworks to demonstrate context-awareness rather than merely assert it. Each hospitalization was constructed so the ED's initial impression drifts to a different discharge diagnosis, making diagnostic-trajectory tracking a scoreable capability rather than a marketing phrase.
Read as a portfolio, the work makes an architectural argument. Trustworthy clinical AI is a pipeline of earned privileges, not a model choice. The pieces that make it trustworthy are unglamorous and deterministic: sufficiency gates, cadence analysis, controlled vocabularies with forbidden phrasings, schema-validated agent contracts, quarantined interpretation layers, and ground-truth evaluation harnesses. The LLM sits inside that structure and is at its best doing what LLMs do well — fluent description within a constrained vocabulary, and pattern-matching against curated exemplars — while the structure decides when it is allowed to speak and how specific it may be.
The portfolio also demonstrates a second, quieter claim: that a small clinically-led team can produce the reasoning specifications — the checklists, axes, stage models, and vocabularies — that engineering teams and vendors otherwise have to guess at. The frameworks are the transferable asset. Any of them could be re-implemented on a different stack tomorrow, because the intellectual property is the gating logic and the controlled language, not the code.
Natural next steps, in rough order of leverage: clinical sign-off on the lab-velocity thresholds; expansion of the BRIDGE evaluation from six packets to the full cohort's event library; and a pilot wiring the Lab Summary Agent's axis outputs into the CMDF modules as their upstream data-sufficiency source — closing the loop between description and orchestration.