TrustedRisk Tools

TrustedRisk MCP is a SHARP-on-MCP server publishing 145 calibrated clinical decision-support tools across 47 thematic bundles. Every probabilistic tool ships its calibration coefficients, conformal prediction set, fairness audit, and an append-only audit-chain entry per output. The three demos in this video show how the server behaves on one SHARP-bound patient. The demo patient is Marcus Reyes, 63, admitted with urosepsis from a complicated urinary tract infection, with a documented penicillin allergy. His FHIR bundle is bound to the chat session through the SHARP-on-MCP context headers X-FHIR-Server-URL, X-FHIR-Access-Token, and X-Patient-ID, and every tool the agent invokes reads his real chart through that binding. Why this matters: the Epic sepsis early-warning model fired on roughly 40% of unselected ED arrivals because its decision threshold drifted from the calibration cohort it was published on (Wong et al., JAMA Internal Med 2021); the CMS Hospital Readmissions Reduction Program penalises hospitals roughly USD 580M per fiscal year on miscalibrated readmission band selection (CMS HRRP final rules 2023); hospital adoption of FDA-cleared AI clinical tools sits below 30% with audit absence and calibration opacity cited among the top reasons (NEJM AI 2024). TrustedRisk targets the audit-trail surface of that adoption gap. Timestamps: 00:00 Healthcare AI tools publish probabilities without their calibration 00:10 Patient context: Marcus Reyes, urosepsis from complicated UTI 00:25 Demo 1 -- calibrated 30-day readmission probability with conformal set 00:50 Demo 2 -- scribe drafts a consult letter from the chart SOAP DocumentReference 01:20 Demo 3 -- structured abstain when the PGx genotype Observation is missing 01:50 29 chain harnesses, 90 distinct MCP tools exercised, zero crashes 02:15 Regulatory pack at 100% artefact coverage, OpenAPI 3.1, 98-module ML substrate 02:45 Closing slate Architecture summary: 145 compute_* / detect_* / ground_* tools across 47 thematic bundles, each registered with FastMCP via a per-bundle register shim so a marketplace consumer can subscribe to a single bundle without pulling in the rest; every probabilistic tool ships a 5-bin Beta-Binomial posterior with automatic CI95 widening plus a Romano 2020 LAC conformal score that turns the point estimate into a marginal-coverage prediction set; the 98-module trustworthy-ML substrate under src/a2a_agent covers calibration, conformal prediction, fairness, federated learning, and causal inference, and every probabilistic MCP tool calls into it; SHARP-on-MCP middleware enforces the FHIR context headers on every non-framework JSON-RPC method and returns 403 with a spec_reference body when required headers are missing; OAuth client_credentials grant uses HS256 JWT with tenant-allowlist validation, so a token issued for tenant A cannot be used to query tenant B even if the bearer token is otherwise valid. Regulatory and validation surfaces: a 14-section EU AI Act Annex IV regulatory pack reaches 100% artefact coverage, FDA SaMD Class II / III analysis covers the tool surface, ISO 13485 design-controls checklist documents the gap to a Class II audit, Mitchell 2019 Model Card and Gebru 2021 Datasheet cover the calibration cohort, NIST AI RMF 1.0 + OECD AI Principles + HIPAA section 164 + GDPR Article 35 crosswalks are checked in, 4236 unit + integration + golden + adversarial tests pass, ECE 0.0078 on the n=7,880 internal Synthea calibration cohort, 150 end-to-end cases run with zero crashes, and 29 chain harnesses exercise 90 distinct MCP tools at 62% surface coverage with zero crashes. v1.1.0 chart-authoritative hardening: when SHARP context is bound, demographics, vital signs, labs, and medication lists come from the patient's FHIR chart through LOINC-coded Observation lookup and Patient resource extraction, and caller-supplied values for those parameters are discarded so a chat-side LLM cannot inject hallucinated clinical data into a Tier-1 scoring or dosing tool. Built for the "Agents Assemble: The Healthcare AI Endgame" hackathon (Prompt Opinion and Darena Health, May 2026), Alpha track / Superpower / MCP Server. Repository: https://github.com/AlessandroFlati/TrustedRisk #MCP #SHARPonMCP #ClinicalDecisionSupport #FHIR #HealthcareAI #PromptOpinion #AgentsAssemble #CalibratedAbstention #ConformalPrediction #PGx #LACE #EUAIAct #SaMD