Inside the Curavit and Lindus CRO deal
The decentralized trial hype has faded, but the direct-to-patient model has matured, and the teams
The proposed European medical device simplification removes a reporting wrapper while preserving the statutory obligation to generate clinical evidence. The European Commission proposal COM(2025) 1023 final would eliminate the standalone post-market clinical follow-up (PMCF) evaluation report under Annex XIV Section 7. It would also relax several fixed reporting cadences.[1][2] This proposal leaves the core mandate to proactively collect and evaluate clinical data completely intact (Annex XIV Part B Section 5). The PMCF plan requirements in Sections 6, 6.1, and 6.2 also remain in full force.[8] You must still determine what clinical uncertainty remains. You must select a proportionate proactive method and defend that choice to your notified body.
This distinction changes the practical PMCF question. Instead of asking which report to prepare, you must identify the remaining evidence gap. You need to determine which method can close it and how to generate the evidence efficiently without weakening traceability or scientific control.
Evidence convention. Regulatory statements rest on statutory texts and official guidance.[1][4][8] Performance findings originate from peer-reviewed informatics literature[10][17] or Castor operational benchmarks from live production workflows.[36] Financial results rely on modeled first-principles scenarios. Full provenance appears in Appendix G.
The rest of this guide continues below: the method-selection matrix, the notified-body defense dossiers, the AI control domains, and the full financial model. Enter your work email to keep reading and to receive the designed PDF of the complete guide.
The European Commission proposal COM(2025) 1023 final has generated understandable optimism across the medical device sector.[1][3] The proposal aims to remove duplicated documentation, relax fixed update cycles, and shift obligations toward a risk-based cadence. These changes reduce the administrative burden for stable legacy devices on the EU market.
You still bear the statutory responsibility to understand the clinical performance and safety of a device throughout its lifecycle. Annex XIV of Regulation (EU) 2017/745 (MDR) continues to mandate proactive clinical data collection.[8] The proposed simplification alters how you document PMCF findings. It leaves the underlying legal obligation to generate clinical evidence completely intact where residual uncertainty remains.
Under current MDR law, PMCF operates as an ongoing and proactive lifecycle process.[8] You must continuously update your clinical evaluation report (CER), risk management file (RMF), and post-market surveillance (PMS) plans based on real-world clinical experience.
COM(2025) 1023 final currently sits as an ordinary legislative proposal under procedure 2025/0404(COD).[2] The European Parliament Committee on Public Health (SANT) is reviewing the proposal. The Rapporteur tabled Draft Report PE787.987 on 30 June 2026, and it awaits a committee decision.[19] Current MDR requirements remain fully applicable until formal trilogue adoption and subsequent entry into force.
| Proposed administrative relief (COM(2025) 1023) | Clinical obligations that stand intact (MDR) |
|---|---|
| Standalone PMCF report wrapper removed. Findings documented directly in the CER and technical file (Annex XIV Section 7 replaced).[1] | Evidence gaps must still be identified. Annex XIV Part B Section 5 mandates proactive collection and evaluation of clinical data across the lifecycle. Sections 6, 6.1, and 6.2 specify what the PMCF plan must contain.[8] |
| Fixed update cadence relaxed. Annual CER updates for Class III and implants shift to clinically triggered updates (Article 61(11)).[1] | PMCF methods must still be selected. Annex XIV Section 6.2(a) and 6.2(b) require general and specific proactive methods.[8] |
| PSUR delivery schedule simplified. Biennial updates for Class IIb and III after Year 1, Class IIa on request (Article 86).[1] | Method appropriateness must still be justified. Annex XIV Section 6.2(c) mandates a documented defense to notified bodies.[8] |
| Non-applicability clarified. The proposal adds an explicit justification route for mature, low-risk devices.[1] | Non-applicability must still be justified. Under current MDR the post-market surveillance plan must contain a PMCF plan, or a justification as to why PMCF is not applicable, documented in the clinical evaluation (Annex III Section 1.1).[8] |
Regulatory basis (per Sanne Derks review). Annex XIV Part B Section 5 establishes the proactive-collection mandate. It states that “when conducting PMCF, the manufacturer shall proactively collect and evaluate clinical data from the use in or on humans of a device which bears the CE marking.” Sections 6, 6.1, and 6.2 govern what the PMCF plan must specify and how methods are chosen and justified.
As a practical consequence, you must continue PMCF while designing it more intelligently. You can move resources away from repetitive report formatting. You can direct them toward acquiring clinical data that directly answers unresolved CER and risk-management questions.
Methodological note. To illustrate how the regulatory, methodological, technical, and economic principles in this guide apply in practice, we follow a single hypothetical worked example throughout. We examine a Class IIb/III drug-eluting coronary stent facing an Annex XIV Section 6.2(b) long-term follow-up evidence gap. All protocol parameters, data workflows, and cost models are illustrative. They derive from operational benchmarks across Castor’s PMCF and clinical registry portfolio.
A Class IIb/III drug-eluting coronary stent has strong short-term 12-month pre-market trial data, but long-term target lesion revascularization (TLR) and late stent thrombosis in high-risk diabetic subgroups remain uncharacterized at 5 years. Grounded in a Section 6.1(b) PMCF objective (identifying and analyzing previously unknown risks or side-effects in a subpopulation), the notified body issues a formal condition for CE-mark renewal. You must actively collect 5-year post-market clinical data on this diabetic cohort. The selected proactive method to meet that objective sits under Section 6.2(b) (see Section 2). Throughout this guide, we use this example to show how each regulatory requirement translates into method selection, AI-driven execution with human review, and budget optimization.
Objective and method (per Sanne Derks review). The notified body condition maps to a Section 6.1(b) PMCF objective. The retrospective chart-review method chosen to satisfy it is a Section 6.2(b) specific method. Objective and method are distinct clauses and should not be conflated.
You must base your PMCF method selection on a defined clinical uncertainty in the CER or risk-management file that requires resolution. You cannot select a method simply because a registry, survey platform, or extraction tool happens to be available.
A defensible PMCF plan begins with four questions.
MDR Annex XIV Section 6.2 establishes two tiers of proactive clinical data collection, plus a justification duty.[8]
| Residual evidence gap | Best available data source | Candidate PMCF method | Mandatory fitness checks | Minimum defense deliverable |
|---|---|---|---|---|
| Mature device, zero residual design risks, known usability. | High-volume surgical users, published literature. | Literature screening plus HCP eCOA surveys (6.2(a)).[4] | Response-rate validity (at or above 80%), representative user sampling, absence of complaints. | Screening protocol, validated survey tool, statistical summary report. |
| High-quality national registry exists with UDI tracking. | National clinical quality registers (for example NJR, BCIR). | Registry extraction and linkage (6.2(b)).[5][27] | Data completeness (above 90%), device-level UDI capture, longitudinal follow-up. | Registry governance agreement, extraction protocol, data audit. |
| Patient-reported outcomes, recovery quality, pain scores. | Direct patient follow-up cohorts. | Digital ePRO / eCOA follow-up (6.2(a)).[28] | Long-term retention (above 70%), documented Articles 6 and 9 basis, informed consent where required, validated scales. | Validated ePRO audit trail, compliance logs, retention strategy. |
| Implantable device, long-term safety and survivorship gaps. | Hospital electronic health records, surgical notes. | Multi-center retrospective chart review (6.2(b)).[10][13] | Source-document availability, certified-copy integrity, 100% source provenance. | GCP-compliant protocol, eSource audit trail, statistical analysis plan. |
| Novel mechanism of action, unobservable in routine care. | Prospectively enrolled clinical cohort. | Prospective post-market clinical study (6.2(b)).[8] | Ethical-approval feasibility, site recruitment capacity, protocol compliance. | ISO 14155 GCP clinical investigation plan, monitoring dossier.[13] |
ISO 14155 scope (per Sanne Derks review). Registries and any other non-interventional but prospective studies are within the scope of ISO 14155. A study is treated as prospective for the regulatory framework whenever it carries any prospective element, which affects the applicable compliance obligations.
Applying this method-selection logic to the worked coronary stent example.
Primary estimand and acceptance rule (illustrative). In this single-arm retrospective cohort of N=200 diabetic patients evaluated against a historical control benchmark (for example a historical 5-year TLR rate of 4.2% with a pre-specified non-inferiority margin of 2.5%), the primary objective is to demonstrate that the upper bound of the one-sided 95% confidence interval for 5-year TLR remains at or below 6.5%. Cumulative 5-year MACE at or below 7.5% serves as a descriptive secondary endpoint. Final sample size and power must be formally established in the statistical analysis plan. The N=200 figure serves as a standardized operational benchmark for study execution and budget modeling.
Model Section 6.2(c) defense language. “Pursuant to MDR Annex XIV Part B Section 6.2(c), specific proactive clinical data collection is executed via a multi-center retrospective cohort (N=200) linked to national cardiovascular registers under Section 6.2(b). General surveys under Section 6.2(a) are insufficient to detect low-frequency 5-year stent thrombosis. The retrospective design captures longitudinal hard endpoints, while AI-driven source extraction with 100% human review mitigates manual transcription error and preserves full source provenance.”
Selecting a defensible method solves the regulatory design problem, but executing registry extraction and custom PMCF cohorts (site uploads) creates a severe operational bottleneck. When sponsors extract longitudinal outcomes from hospital EHRs, cath-lab records, pathology reports, and paper source worksheets, clinical research coordinators spend hundreds of hours retyping data into eCRFs. The next section introduces Castor Catalyst. This AI-driven extraction and site-upload engine with human review removes this bottleneck while maintaining 100% human-in-the-loop auditability.
For data-intensive PMCF studies, the main operational failure mode occurs during manual chart abstraction. Clinical research coordinators and abstraction teams hit several failure points when reviewing unstructured inpatient and outpatient records.
SDV reframing (per Sanne Derks review). The prior 2012 figure (“15 to 30% of budgets, alters conclusions in under 0.1% of cases”) reflected a 100%-SDV era and has been removed. Risk-based monitoring is now standard, and ISO 14155 requires the monitoring plan to justify the amount of SDV, up to and including no on-site monitoring where risks are adequately mitigated.
Governance: deterministic execution (temperature 0), GAMP 5 Category 4/5, zero model training on client data. Standards: ISO 14155, MDR Annex XIV, FDA RWE guidance.
Catalyst operates through a deterministic, six-stage pipeline designed to satisfy ISO 14155 GCP and to align with the FDA/EMA good AI practice principles (applied by analogy).[13][31]
Governance. Deterministic execution (temperature 0), GAMP 5 Category 4/5, zero model training on client data. Standards include ISO 14155, MDR Annex XIV, and FDA RWE guidance.
Reviewer override rate (99.2% accepted as extracted)[36]
Per-chart abstraction ($9.75 vs $48.75)[36]
Clean document ingestion, 0.2% routed to manual queue[36]
In live production deployments of the Catalyst site-uploads workflow and internal benchmarking against real EMR records, Castor observed a 0.8% adjudication override rate. Medically trained reviewers accepted or confirmed 99.2% of AI-driven proposals, compared to a 6.6% manual chart-abstraction baseline.[10] The system delivered 100% human-in-the-loop completion and a 99.8% clean ingestion rate. We also documented a five- to sixfold abstraction acceleration, dropping time from 39 minutes to 6 minutes per chart. This lowered direct abstraction cost from $48.75 to $9.75 per chart, with zero model training on client data.[36]
Across the Castor platform, device teams run 893 MedTech studies to date (289 live) and 298 post-market studies including PMCF (98 live), the footprint behind the operational telemetry above.[36]
To defend AI-driven PMCF data with human review to notified bodies and health authorities, you must enforce five governance controls.[31][32]
In the worked stent example, the AI-driven system extracts 60,000 target fields from multi-center cath-lab records for human review. Sites upload de-identified operative reports, angiographic discharge summaries, and diabetic clinic follow-up notes via Catalyst site uploads. Catalyst extracts lesion length, reference vessel diameter, stent deployment pressures, and follow-up ischemic events. It maps adverse cardiac events to MedDRA preferred terms. A centralized core-lab research nurse reviews every extracted variable side-by-side against the highlighted operative PDF bounding box. You export an ALCOA+, source-linked dossier demonstrating that 100% of 5-year TLR events are verified against certified hospital source documents, with an adjudicated override rate under 1.0%.[36]
With the Catalyst architecture and human-in-the-loop controls established, how does this translate into study budgets? This section presents a first-principles economic comparison across three operational models for a standardized post-market study.
We model a standard data-intensive PMCF study. It includes N=200 patients across 10 clinical centers, 5-year follow-up, and 300 variables per patient (60,000 total target fields).
Model A. Traditional on-site ($5,510 / patient)
Model B. Centralized manual ($3,945 / patient)
Model C. Castor Catalyst AI + HITL ($2,755 / patient)
| Budget line item | A: traditional | B: centralized manual | C: Catalyst AI + HITL | Primary driver of efficiency |
|---|---|---|---|---|
| Protocol and regulatory design | $45,000 | $45,000 | $45,000 | Standardized regulatory dossier and SAP. |
| Site identification and contracting | $100,000 | $60,000 | $60,000 | Centralized contracting, no EDC training at sites. |
| EDC build and system validation | $55,000 | $55,000 | $45,000 | Pre-validated extraction schemas, CDISC CDASH. |
| Site management and CRA monitoring | $180,000 | $120,000 | $96,000 | Remote SDV replaces routine CRA travel. |
| Direct chart abstraction labor | $120,000 | $100,000 | $70,000 | 6.67 hrs/patient manual reduced to a fixed $350/patient AI + HITL fee (a $50,000, 41.7% direct-labor saving). |
| Data management and query resolution | $145,000 | $85,000 | $45,000 | AI-driven pre-validation with human review eliminates transcription syntax errors. |
| Medical writing, CER and technical file | $65,000 | $50,000 | $40,000 | Automated structured evidence export into the CER. |
| Project management and governance | $92,000 | $74,000 | $50,000 | Accelerated timeline reduces monthly CRO fees. |
| Site grant fees ($1,500/patient) | $300,000 | $200,000 | $100,000 | Reduced site burden lowers per-patient compensation. |
| Total fully loaded study budget | $1,102,000 | $789,000 | $551,000 | 50.0% net saving vs traditional, 30.2% vs centralized manual. |
| Per-patient fully loaded cost | $5,510 | $3,945 | $2,755 | Capital efficiency from combining centralization with AI-driven extraction and human review. |
For the traditional and centralized manual build, 300 variables per patient across multi-year unstructured records take an average of 6.67 hours of coordinator labor per patient (about 1.33 minutes per variable). At a fully loaded $90/hour, manual abstraction costs $600 per patient, or $120,000 across 200 patients. For the Catalyst build, the AI-driven system ingests source files and extracts 300 variables in seconds for human review. A centralized research nurse then reviews the highlighted bounding boxes in about 45 minutes per patient. Including platform software and managed HITL review, the fully loaded cost is $350 per patient, or $70,000 across 200 patients. This creates a $50,000 (41.7%) direct-labor reduction.
At a 6.6% error rate,[10] manual abstraction produces about 3,960 discrepant data points. Internal validation rules catch about 2,000 syntax errors. The remaining 900 complex clinical queries require formal cross-functional resolution. Resolving 900 queries at $180 each costs $162,000 in gross operational friction. With a 0.8% override rate and AI-driven pre-commit validation with human review, complex queries drop to about 200. This reduces gross friction to $36,000, representing a 77.8% friction saving.
Sensitivity modeling across cohort sizes shows consistent behavior. N=100 saves 49.3% ($345k vs $680k), and N=500 saves 51.8% ($1.18M vs $2.45M). Studies capturing more than 150 variables per patient from unstructured hospital records achieve the highest return.
For the worked stent study (N=200 diabetic patients, 10 centers, 60,000 target fields), applying the AI-driven Catalyst system with human review reduces the fully loaded 5-year PMCF budget from $1,102,000 to $551,000. This is a $551,000 (50.0%) reduction. Direct abstraction labor falls from $120,000 to $70,000. Gross query friction drops from $162,000 to $36,000. This lets you maintain CE-mark compliance within commercial margin targets.
Clinical evidence generated for MDR PMCF should not stay trapped in a single regulatory silo. When structured correctly, a governed real-world dataset can serve multiple commercial and scientific objectives.
To reuse real-world PMCF data legally and scientifically, sponsors establish six governance controls at protocol inception. These include prospective estimand prespecification, full source traceability, data completeness and quality thresholds (above 90% completeness, under 1.0% error rate[36]), transparent bias quantification, a defensible GDPR compliance architecture,[9][35] and regulatory consistency with published SSCP files.[8]
Governance note. Illustrative allocation only. Controller and processor roles, the Articles 6 and 9 conditions, and applicable national-law requirements must be confirmed for each processing operation based on the actual purposes and means of processing. The primary PMCF conduct rests on the MDR legal obligation (Art 6(1)(c)) with Art 9(2)(i). Where processing is for secondary scientific research beyond that legal duty, the Article 6(1)(e)/(f) and 9(2)(j) research basis applies (row 3). Reference [9] (EDPB Opinion 3/2019) is a clinical-trials-regulation opinion applied here by analogy.
| Processing operation | Controller | Processor | Lawful basis (Art 6 / 9) | Core safeguards |
|---|---|---|---|---|
| 1. PMCF study conduct and abstraction | Device sponsor | Technology vendor | Art 6(1)(c) (MDR duty), Art 9(2)(i) (public health / safety) | Art 4(5) pseudonymization at site, unique study codes, strict RBAC. |
| 2. Vigilance and safety reporting | Device sponsor | Clinical site | Art 6(1)(c) (MDR duty), Art 9(2)(i) (high safety standards) | Secure E2B XML reporting, pseudonymized adverse-event logs. |
| 3. Secondary HTA and reimbursement analysis | Device sponsor | Academic partner | Art 6(1)(e)/(f) (public / legitimate), Art 9(2)(j) (scientific research) | Fully anonymized dataset, cell suppression for low-frequency categories. |
| 4. Technical support and system hosting | Device sponsor | Technology vendor (sub-processors) | Processing under the controller’s documented instructions and underlying lawful basis. Art 28 DPA controls apply. | Zero PHI/PII visibility, encrypted logs, Chapter V SCCs and TIAs.[35] |
| 5. AI model training and fine-tuning | Prohibited | Prohibited | Prohibited (strict zero retention) | Client clinical data is segregated and never used for foundation-model training. |
For the worked stent example, the 5-year retrospective dataset resolves the CER safety question for the notified body. Because the protocol prospectively prespecifies the long-term outcomes to be extracted and preserves complete source provenance, you can reuse the governed evidence asset for a purpose-specific HTA analysis (for example with France’s HAS or the UK’s NICE[33]). You can evaluate whether it supports a narrowly defined external claim under Article 7 MDR. This potentially avoids a separate seven-figure evidence-generation study.
Transitioning to a modern, AI-driven PMCF operating model with human review does not require an immediate overhaul of the entire portfolio. You should follow a phased 90-day roadmap.
COM(2025) 1023 final would signal a welcome maturation of the European medical device framework. By removing redundant report wrappers and shifting toward risk-based cadences, the proposal creates space for you to focus on substantive clinical science. The statutory duty to demonstrate long-term safety and performance under MDR Annex XIV Part B remains. If you rely on manual chart transcription and fragmented site monitoring, you will keep facing cost inflation and audit vulnerability. By pairing rigorous Annex XIV method selection with centralized operations and validated AI-driven extraction with 100% human-in-the-loop review, you can turn PMCF from a costly compliance burden into a permanent, high-trust clinical-evidence asset.
No. COM(2025) 1023 would delete the standalone PMCF evaluation report under Annex XIV Section 7 and relax some reporting cadences, but the mandate to proactively collect and evaluate clinical data (Annex XIV Part B Section 5) and the PMCF plan requirements in Sections 6, 6.1, and 6.2 remain fully in force. Until the proposal is adopted and enters into force, current MDR requirements apply.
Annex XIV Part B Section 5. It requires the manufacturer to proactively collect and evaluate clinical data from the use of a CE-marked device across its lifetime. Sections 6, 6.1, and 6.2 govern what the PMCF plan must specify and how methods are chosen and justified.
Start from the residual evidence gap in the clinical evaluation, select a proportionate method under Annex XIV Section 6.2 (general methods under 6.2(a) or specific methods under 6.2(b)), and provide the documented justification required by Section 6.2(c) explaining why the chosen method meets the Section 6.1 objectives.
Yes, with human review. AI-driven source extraction is defensible when every extracted value is reviewed by a qualified person before it enters the record (100% human-in-the-loop), with full source traceability, deterministic version-locked execution, and ISO 14155 source-data integrity.
In a first-principles model of a 200-patient, 10-center, 5-year study, a traditional on-site build runs about $1,102,000, a centralized manual build about $789,000, and Castor Catalyst with human review about $551,000, a 50% reduction versus traditional. The savings come from the data work, direct abstraction labor and query friction, not the study design.
Specific amendments proposed under COM(2025) 1023 final and European Parliament SANT Committee Draft Report PE787.987 (30 June 2026).[1][19]
| MDR provision | Current MDR (2017/745) | Commission proposal COM(2025) 1023 | SANT draft report (PE787.987) | Operational impact |
|---|---|---|---|---|
| Annex XIV Part B Section 5 (proactive-collection mandate) | Proactive collection mandatory. Non-applicability must be justified and documented in the clinical evaluation (Annex III Section 1.1).[8] | Adds an explicit non-applicability justification route for mature, low-risk devices. The collection mandate is unchanged.[1] | Requires documented clinical proof of zero residual risk per Annex III Section 1.1.[19] | Notified bodies require rigorous clinical proof before granting non-applicability. |
| Annex XIV Part B Section 6 (PMCF plan and duties) | PMCF plan, objectives, and method justification mandatory.[8] | Unchanged. Planning and method selection remain mandatory.[1] | Emphasizes risk-proportionate application of specific versus general methods.[19] | Core plan obligations remain fully intact. |
| Annex XIV Part B Section 7 (PMCF report) | Standalone annual PMCF report mandatory for Class IIa, IIb, III.[8] | Deletes Section 7. Findings documented in the CER.[1] | Endorses replacement, mandates clear cross-referencing in technical files.[19] | Eliminates standalone report formatting. |
| Article 61(11) (CER cadence) | Annual CER updates for Class III and implants.[8] | Shifts to clinically triggered updates.[1] | Clarifies implants reviewed at least biennially.[19] | Reduces calendar reporting, requires continuous PMS signal detection. |
| Article 86 (PSUR) | Annual PSUR for Class IIb/III, biennial for IIa.[8] | Biennial for IIb/III after Year 1, IIa on request.[1] | Clarifies electronic submission via EUDAMED.[19] | Aligns PSUR delivery with CER updates. |
Reusable rotary surgical reamer for total hip arthroplasty (Class IIa, Rule 6). Mature technology on the EU market for more than 12 years with more than 250,000 lifetime uses. Residual gap. Long-term cutting efficiency, dimensional wear, and sterilization resilience after more than 50 autoclave cycles. Selected method. Systematic literature screening plus a validated surgeon eCOA survey at high-volume centers (6.2(a)).[4] Acceptance. N=60 surgeons across 4 member states, at or above 85% completion, at or above 90% satisfaction, zero intraoperative fractures. Model 6.2(c) defense. “Pursuant to MDR Annex XIV Part B Section 6.2(c), PMCF is executed via high-volume surgeon eCOA survey and systematic literature surveillance under Section 6.2(a). Specific clinical investigations under Section 6.2(b) are unjustified as the reamer is a well-established Class IIa technology with well-characterized clinical risks.”
Standalone SaMD deep-learning algorithm (Class IIb, Rule 11) analyzing emergency head CT scans for large-vessel-occlusion ischemic stroke and hemorrhage. Residual gap. Real-world sensitivity and specificity across heterogeneous CT hardware and elderly patients with leukoaraiosis. Selected method. Multi-center retrospective imaging database extraction and diagnostic-accuracy evaluation (6.2(b)).[5][10] Acceptance. N=500 consecutive emergency stroke CTs across 5 centers with neuroradiologist consensus ground truth, lower bound of the 95% CI sensitivity above 90.0% (target at or above 93.0%), specificity above 88.0% (target at or above 91.0%). Model 6.2(c) defense. “Pursuant to MDR Annex XIV Part B Section 6.2(c), clinical follow-up is conducted via multi-center retrospective diagnostic-accuracy evaluation (N=500) under Section 6.2(b). General surveys under Section 6.2(a) cannot validate diagnostic accuracy.”
| Device risk class | Residual gap | Primary data source | Recommended method | PMCF method clause | Est. budget | Duration |
|---|---|---|---|---|---|---|
| Class I / IIa (mature) | Usability, routine feedback | Clinicians, published literature | Literature screening + HCP survey[4] | Annex XIV 6.2(a) | $15k to $40k | 2 to 4 months |
| Class IIa / IIb (orthopaedic) | Long-term survival, UDI tracking | National registries (NJR, EPRD)[27] | Registry extraction and linkage[5] | Annex XIV 6.2(b) | $80k to $220k | 6 to 12 months |
| Class IIa / IIb (digital / SaMD) | Patient recovery, quality of life | Enrolled patient app users | Digital ePRO / eCOA cohort[28] | Annex XIV 6.2(a) | $50k to $140k | 4 to 8 months |
| Class IIb / III (implantable) | Survivorship, rare complications, subgroups | Hospital EHRs, surgical records[10] | Multi-center retrospective chart review[13] | Annex XIV 6.2(b) | $200k to $550k | 6 to 14 months |
| Class III / novel tech | Uncharacterized biological response | De novo enrolled clinical sites | Prospective PMCF clinical study[8] | Annex XIV 6.2(b) | $850k to $2.5M+ | 18 to 36 months |
| Control domain | Verification requirement | Mandatory deliverable | Regulatory benchmark |
|---|---|---|---|
| 1. Validation and QMS | GAMP 5 validation, ISO 13485 / ISO 9001 quality system. | IQ/OQ/PQ dossier, system traceability matrix, release reports. | GAMP 5 Cat 4/5, 21 CFR 11, MDR Annex IX. |
| 2. Traceability | Deterministic execution (T=0), visual source bounding-box audit trail. | Model architecture spec, bounding-box verification demo. | ISO 14155 source-data integrity, FDA RWE guidance.[13][34] |
| 3. Human-in-the-loop | 100% human verification workflow, medical qualification logs. | Reviewer UI workflow docs, adjudication role SOPs. | FDA/EMA good AI practice, MDCG 2020-7.[4][31] |
Operational basis. N=200 patients across 10 centers, 5-year follow-up, 300 variables/patient (60,000 total target fields across operative notes, cath logs, and outpatient visits). Traditional decentralized model. Site activation $10k/site ($100k), CRA monitoring and travel $18k/site ($180k), labor 6.67 hrs/patient at $90/hr ($120k), 900 complex queries at $180 ($162k gross friction), site grants $1,500/patient ($300k), total $1,102,000 ($5,510/patient). Centralized manual model. Centralized onboarding ($60k), remote monitoring ($120k), manual abstraction $100k, DM and queries $85k, site grants $1,000/patient ($200k), total $789,000 ($3,945/patient, saves 28.4%). Castor Catalyst AI + HITL model. Pre-validated schemas ($45k), remote SDV ($96k), Catalyst software plus managed HITL review fixed $350/patient ($70k, 41.7% direct-labor saving), post-HITL query friction $36k (77.8% saving), site grants $500/patient ($100k), total $551,000 ($2,755/patient, 50.0% net saving). Sensitivity. Across N=100 to N=500, total savings range 49.3% to 51.8% versus traditional on-site execution.
Governance note. Illustrative allocation only. Controller and processor roles, the Articles 6 and 9 conditions, and applicable national-law requirements must be confirmed for each processing operation. This matrix expands Section 5.3 with detailed privacy safeguards per operation.
| Processing operation | Controller | Processor | Lawful basis (Art 6 / 9) | Privacy safeguards and controls |
|---|---|---|---|---|
| 1. Primary study conduct and abstraction | Device sponsor | Technology vendor | Art 6(1)(c) (MDR duty), Art 9(2)(i) (public health / safety) | Pseudonymization at the clinical site, unique patient study codes, role-based access control. |
| 2. Vigilance and safety reporting | Device sponsor | Clinical site | Art 6(1)(c) (MDR duty), Art 9(2)(i) (high safety standards) | Secure E2B XML reporting channels, pseudonymized adverse-event logs. |
| 3. Secondary HTA and reimbursement analysis | Device sponsor | Academic partner | Art 6(1)(e)/(f) (public / legitimate), Art 9(2)(j) (scientific research) | Fully anonymized dataset generation, cell suppression for low-frequency categories. |
| 4. Technical support and system hosting | Device sponsor | Technology vendor (sub-processors) | Processing under the controller’s documented instructions and underlying lawful basis. Art 28 DPA controls apply. | Zero access to clinical records, encrypted logs, Chapter V SCCs and TIAs.[35] |
| 5. AI model training and fine-tuning | Prohibited | Prohibited | Prohibited (strict zero retention) | Client clinical data is strictly segregated and never used for foundation-model training. |
Evidence tiers. Tier 1 (statutory texts and regulatory authorities). Regulation (EU) 2017/745 (MDR),[8] COM(2025) 1023 final,[1] SANT Committee Report PE787.987,[19] MDCG guidance (2020-5, 2020-7, 2020-8, 2021-24, 2022-21, 2024-10), ISO 14155,[13] FDA guidances,[25][34] and joint FDA/EMA good AI practice.[31] Tier 2 (Castor production benchmarks[36]). Operational telemetry from live Catalyst site-uploads deployments. Definitions. 100% HITL review = every extracted field reviewed and approved by a qualified clinical professional before commit. 0.8% override rate = adjudicated human edits divided by total verified fields. 99.8% clean ingestion = source documents parsed without exception-queue escalation. Time and cost efficiency = 6 minutes vs 39 minutes per chart ($9.75 vs $48.75). Tier 3 (modeled financial scenarios). First-principles bottom-up cost model for an illustrative 200-patient, 10-center, 5-year study.
Reference note. Reference [14] (Tudur Smith et al., PLoS ONE 2012) was removed with the stale 100%-SDV statistic per Sanne Derks’ review. The numbering is otherwise held stable to avoid citation drift, renumber on a final pass if preferred.
Build changelog (not for publication, strip before publish). v3, 2026-09-15. v2 was worked up FRESH from Derk draft_v6 (full long-form incl. Appendices A to G), superseding condensed v1, with Sanne Derks’ four review comments applied (Section 5 mandate vs 6/6.1/6.2 plan; stent worked example grounded in a 6.1(b) objective with method held at 6.2(b); ISO 14155 scope note at 2.2; SDV reframed with the 2012 Tudur Smith stat and ref [14] removed). v3 applies the medical-device-reviewer pre-screen. ISO edition year DROPPED per Kevin (bare “ISO 14155”, the MDR-harmonized edition is EN ISO 14155:2020/A11:2024, not the 4th ed), F1 Annex III cite corrected 1.1(b) to 1.1, F4 dropped the ISO Section 7.3 subclause (source-data integrity / ALCOA+), F5 FDA/EMA good AI practice cast as principles applied by analogy, not requirements, F6 proposal-not-law tense fixed in the callout and conclusion, F8 “audit-ready” to ALCOA+, F10 IMDRF device-event coding added, F11 Appendix F matrix restored, F13 21 CFR Part 11 paired with EU Annex 11, Appendix C column relabelled “PMCF method clause”. F2 resolved in-house. The non-applicability justification is anchored to current Annex III Section 1.1 rather than asserted as added to Annex XIV Section 5. GDPR resolved in-house. Primary PMCF conduct on Art 6(1)(c) + 9(2)(i), secondary research on 6(1)(e)/(f) + 9(2)(j), EDPB 3/2019 noted as applied by analogy. Castor stats reconciled to the cleared claims register. Footprint set to the current cleared device figures (893 MedTech / 289 live, 298 post-market incl PMCF / 98 live). Core Catalyst stats cleared (0.8%/6.6% per Pollux 2026-09-14, $48.75 to $9.75, 6 vs 39 min, 99.8% ingestion ref [36]). Retired <0.7% removed. Gates waived by Kevin (Derk-authored, Sanne review only), qc_scan done, Gemini voice pass run, human Sanne signs off on the FINAL DESIGN FILE (Kevin, 2026-09-15).
The decentralized trial hype has faded, but the direct-to-patient model has matured, and the teams
Castor Catalyst, Castor’s AI-driven clinical data platform, reads data directly from source records and proposes
Curavit has acquired Lindus Health’s CRO assets, bringing US and European trial delivery under one
Discover all the features offered by Castor EDC
Discover Now