We're deploying a new agent: Clinical Triage Inbound for the web chat. Can you check the MDR classification?
AGENT · MDR-SAMD AVOIDANCE CHECKER
Healthcare agents stay outside the MDR perimeter without workarounds.
MDR-SaMD Avoidance Checker verifies that every new agent intended for healthcare use does not inadvertently fall within the MDR perimeter (EU Regulation 2017/745, Software as Medical Device). It identifies borderline cases, proposes changes to the functional scope, and escalates to the healthcare compliance officer.
MDR-SaMD Avoidance Checker at work.
Analysis complete. Functions: pre-booking triage, recognition of critical clinical signals against declarative rules written by the clinical team, escalation to the triage physician. The agent does not diagnose, does not prescribe, does not make autonomous clinical decisions. Classification: not SaMD under MDR 2017/745. MDR-avoidance pattern confirmed. Silent pass. Classification recorded in the audit registry.
Good. The deploy can proceed.
Why it exists.
The MDR (EU Regulation 2017/745) classifies software as a 'medical device' if it has a specific use in a medical context (diagnosis, prevention, monitoring, treatment, prediction, prognosis). Software classified as SaMD faces rigorous obligations: CE marking, conformity procedures, post-market surveillance, change management. For healthcare facilities that want to use agents, avoiding inadvertent MDR classification is structural.
How it works each deploy.
MDR-SaMD Avoidance Checker activates on every new healthcare agent deploy. It evaluates the functional description, identifies borderline cases, proposes scope changes, and escalates for cases that would clearly qualify as SaMD. The MDR-SaMD avoidance design pattern is the answer: the agent does not diagnose, does not replace clinical judgement, does not make autonomous clinical decisions.
The decision stays with the compliance officer.
The final validation stays with the healthcare compliance officer. For agents that clearly qualify as SaMD under MDR: automatic block with escalation. For borderline agents: a flag with scope reframing alternatives. For non-MDR agents: silent pass with classification in the audit registry.
Healthcare compliance officer, clinical lead, and DPO.
Healthcare compliance officer
The compliance officer gets automated pre-screening of every new agent. It prevents inadvertent deployment of MDR-classified systems without the required conformity procedures (CE marking, risk management, post-market surveillance).
fnol.receive 09:14:22 ALLOW
triage.classify 09:14:25 ALLOW
idd.check 09:14:31 WARN
liquidation.propose 09:15:02 ALLOW
SELECT * FROM audit_log WHERE claim_id = '2024-0847'
Clinical lead
The clinical lead has structured visibility on the scope of deployed agents. The MDR classification of each agent stays under control and is traced, not left to reactive post-deploy assessments.
Weekly audit 4 systems
HIS · accesses OK
LIS · reports OK
EHR · traceability VERIFY
RIS · privacy OK
1 anomaly · alert to healthcare authority lead
Healthcare DPO
The healthcare DPO holds an inspectable trace of classifications for healthcare compliance audit and for inspection by the competent authorities. The registry contains the classification, the verified pattern, and the compliance officer's sign-off.
outpatient visit · 18:24
«the patient reports exertional dyspnoea for about two months…»
I50.9 Heart failure, unspecified
R06.0 Dyspnoea
EHR populated · physician signature required
Silent pass vs escalation, two scenarios.
Clinical Triage Inbound: MDR-avoidance pattern confirmed.
A healthcare facility is deploying a new Clinical Triage Inbound agent for the web chat. The agent reads the functional description: conducts a preliminary conversation with the patient, recognises critical clinical signals based on declarative rules written by the clinical team, escalates to the triage physician. It does not diagnose, does not prescribe. Silent pass with 'not SaMD' classification in the audit registry. The deploy proceeds.
Predictive risk diagnosis: block with escalation.
Opposite case: a facility is deploying an experimental agent for 'predictive cardiovascular risk diagnosis'. The agent reads patient parameters and produces a risk prediction. MDR-SaMD Avoidance Checker assesses: this agent produces an autonomous clinical prediction and qualifies as SaMD under MDR. Block with escalation to the healthcare compliance officer.
The compliance officer reframes the scope or starts CE marking.
The compliance officer reviews, validates the analysis, and decides whether to start the CE marking procedure for SaMD or to reframe the agent's scope — turning the 'prediction' into a 'risk signal for physician review', moving the clinical decision back to the physician. The trace of the decision stays in the MDR audit registry.
Declarative rules, versioned MDR criteria.
The MDR-SaMD Avoidance Checker rules are declarative. The healthcare compliance officer and clinical team define in a readable format the MDR/SaMD classification criteria, the patterns of agents that can operate without falling into the MDR perimeter, the alert patterns for borderline agents. The rules live in the repository, versioned, and must be updated periodically in line with MDR updates and applicable MDCG guidelines.
- Language
- TypeScript (Node.js)
- LLM model
- customer's choice: Anthropic, OpenAI, Mistral, open source models hosted internally, AWS Bedrock for a private model
- Built-in controls used
- pii-detector, prompt-injection
- Trigger
- new healthcare agent deploy
- MDR/SaMD rules
- declarative, versioned, written by the healthcare compliance officer
- Memory
- persistent per instance, pgvector + PostgreSQL FTS on historical classifications
- Registry
- append-only, queryable with a standard SQL client (MDR audit inspectable)
Frequently asked questions about the agent.
For agents that clearly qualify as SaMD under MDR, yes: automatic block with escalation to the healthcare compliance officer. For borderline agents: a flag with scope reframing alternatives. For non-MDR agents: silent pass with classification in the audit registry.
The facility's healthcare compliance officer, in collaboration with the clinical team. The rules follow the MDR text (EU Regulation 2017/745) and the applicable MDCG (Medical Device Coordination Group) guidelines. The agent executes the configured rules; responsibility for the completeness and currency of the rules stays with the compliance officer.
MDR-SaMD Avoidance Checker focuses on classification under MDR 2017/745. For classification under the AI Act (high-risk systems under Annex III), the reference is the AI Act High-Risk Validator agent. The two agents can work in sequence on the same deploy: MDR check first, then AI Act check.
The typical pattern for MDR-SaMD Avoidance Checker is 8-14 weeks. Discovery and criteria collection with the compliance officer 2-3 weeks, MDR rule configuration 3-5 weeks, testing on agents already in production 2-4 weeks, hand-off to the compliance team 1-2 weeks.
From a 30-minute conversation to the squad in production.
A 30-45 minute conversation to understand how MDR-SaMD Avoidance Checker would configure to the customer's healthcare facility. How many healthcare agents in the deploy pipeline, which compliance officer profile, which level of MDR rule detail.