Wirewalk Instituteof AI
hls-240 Intermediate AvailableFlagship

Building and Deploying AI-Enabled Medical Devices

For RA, QA, ML, biostatistics, risk, security and post-market staff at device manufacturers who must take an AI-enabled SaMD from design controls to clearance, change control and field monitoring under FDA rules.

Level
Intermediate
Study time
16 hours
Lessons
23 in 6 modules
Exam
20 questions

About this course

This is not an overview of how AI devices are regulated. Each module is one or two concrete jobs done by a named role at a device manufacturer, applied to a single running case: NeuroFlag, a fictional software device from the fictional Northbank Imaging AI, Inc. that flags suspected intracranial hemorrhage on non-contrast head CT to re-prioritize a radiologist worklist. For every task you see how the work is done today, where AI helps and where it must not be used, the data and systems the work depends on, the tool categories and historical product-category examples (no current feature or endorsement claim), the controls that apply to that exact task, how to measure success, and a hands-on exercise built on synthetic data. No real patient, complaint or submission data appear anywhere in the course.

The course keeps two kinds of AI apart: AI as the product (the device's own model, where FDA's good machine learning practice principles, the predetermined change control plan (PCCP) guidance and the January 2025 AI-enabled device software functions draft guidance apply) and AI as a tool inside the quality system (an LLM drafting a trace matrix or pre-sorting complaints, where QMSR, ISO 13485 clause 4.1.6 and FDA's computer software assurance guidance apply). Controls covered include the QMSR (21 CFR 820 incorporating ISO 13485:2016, effective February 2, 2026), Section 515C and the PCCP final guidance, 21 CFR 807.81(a)(3)(i) and FDA's software-change 510(k) guidance, ISO 14971:2019 with AAMI TIR34971, IEC 62304, Section 524B and FDA's cybersecurity guidance (February 2026 version, which superseded the June 2025 version), 21 CFR 803, 806 and 820.35, and an optional evidence handoff for US firms considering EU expansion. No current EU transition-date rule is tested in the US exam. Historical examples use October 17, 2026; current US authority checks are dated November 5, 2026, and documents that are still draft or discussion papers are labeled as such. The course is educational, not legal or regulatory advice.

US authority checks and independent factual, instructional and assessment reviews were completed November 5, 2026. Completion requires the lessons, all seven listed Decision Labs and the exam. The integrated evidence packet is practice with a self/facilitator rubric, not a separate certificate gate.

Workload: 16 planned hours is the US core, with supplied-output/no-code practice, packet assembly and AI-feedback time explicitly budgeted. Full optional scripting/tool extensions add nine hours; optional cross-border reading/practice adds 50 minutes. The core includes 383 minutes reading, 300 core practice (including final packet iteration), 90 packet assembly, 45 feedback, 102 labs and 40 exam. These are planning estimates, not measured guarantees.

What you will be able to do

  • Build a predicate comparison table for an AI triage device from openFDA and public 510(k) summaries, verifying every row against the source document and logging each assistant error
  • Audit a dataset manifest for patient-level leakage, scanner confounding and out-of-scope populations, and write the data management section a reviewer would read
  • Write ML hazard/control records with explicit planned/implemented/verified status, and separately justify software classification, FDA documentation and QMS-tool assurance
  • Compute standalone sensitivity and specificity with 95% confidence intervals overall and by subgroup, and report a subgroup that misses its pre-specified threshold
  • Produce a model card and labeling statements in which every performance number traces to the validation report and no limitation is dropped
  • Classify proposed software changes as letter-to-file, within an authorized PCCP or new 510(k), citing the regulation or guidance question that drives each call
  • Draft the three parts of a PCCP with quantitative aggregate and per-subgroup acceptance criteria for a planned retraining
  • Evaluate cyber vulnerabilities, privacy-limited field monitoring and synthetic complaints for investigation, CAPA, Part 806 correction reporting and both MDR branches with their clocks

Syllabus

6 modules · 23 lessons · about 16 hours

Module 1: RA Specialist and ML Lead: The Pathway File and the Sequestered Test Set

Building the predicate table and the data management plan for NeuroFlag before any model is trained for submission.
4 lessons
  1. The NeuroFlag Case File: Who Owns Which Job at a SaMD Manufacturer Under QMSR, and AI as Product vs. AI as QMS Tool
  2. RA Specialist Task: Building a Verified Predicate Table from openFDA, Public 510(k) Summaries and FDA's AI-Enabled Device List
  3. ML Lead and Clinical Data Manager Task: The Data Management Plan Under the AI-DSF Draft Guidance, GMLP and HIPAA De-identification
  4. Exercise: Finding Patient Leakage, a Training-Only Scanner Vendor and Out-of-Scope Pediatric Studies in a Synthetic 3,000-Study Manifest
  • Decision Lab · 12 minThe Friday Test-Set LockFriday source lock: preserve real owner/action receipts and pending verification through competing priorities.

Module 2: Risk Manager, Software Systems Engineer and CSA Lead: Design Controls with a Model Inside

Adding ML hazards to the risk file, making the model a traceable IEC 62304 software item, and validating the AI tools the team itself uses.
4 lessons
  1. Risk Manager Task: Adding Data Drift, Automation Bias and False-Negative Delay to the ISO 14971 File with AAMI TIR34971
  2. Software Systems Engineer Task: Treating the Model as an IEC 62304 Software Item and ML Frameworks as SOUP
  3. Exercise: AI-Assisted Trace Gap Analysis on a Synthetic Requirements Export, Scoring the Assistant's False Links and Missed Gaps
  4. QA/CSA Lead Task: Validating an LLM Complaint-Summary Feature Under ISO 13485 Clause 4.1.6 and FDA's Computer Software Assurance Guidance
  • Decision Lab · 12 minThe Control That Exists Only in a DraftThe design-review control record: preserve real owner/action receipts and pending verification through competing priorities.

Module 3: Biostatistician and Labeling Lead: Proving NeuroFlag Works and Saying So Honestly

Designing the standalone and reader studies with pre-specified subgroup endpoints, then turning the results into a model card and labeling without over-claims.
4 lessons
  1. Biostatistician Task: Pre-Specifying Standalone Endpoints, Reference Standard and Subgroups Before Unblinding
  2. When a Reader Study Is Needed: MRMC Designs, FDA's CADe Clinical Performance Guidance and Why Triage Devices Differ
  3. Exercise: Wilson Intervals by Site, Scanner, Age, Sex and Hemorrhage Subtype on 1,200 Synthetic Cases, and Reporting the Subgroups That Fail
  4. Labeling Lead Task: The Model Card, Instructions for Use and Public 510(k) Summary Under FDA's Transparency Principles
  • Decision Lab · 12 minA Strong Aggregate and a Weak SAH ResultA claim without its denominator: preserve real owner/action receipts and pending verification through competing priorities.

Module 4: Product-Security Engineer: The Section 524B Cyber-Device Package for an ML Device

Building an SBOM that includes the model and its runtime, threat-modeling AI-specific attacks, and triaging vulnerabilities without filing unsupported VEX statements.
3 lessons
  1. What Section 524B and FDA's February 2026 Cybersecurity Guidance Require from a Cloud-Connected Triage SaMD
  2. Putting Model Weights, the Inference Runtime and the ML Framework into a CycloneDX SBOM
  3. Exercise: Threat-Modeling Data Poisoning, Model Inversion and Evasion, and Triaging Ten Synthetic Vulnerabilities to VEX Decisions
  • Decision Lab · 12 minThe Shared Model Registry AdvisoryWho owns the shared registry gap?: preserve real owner/action receipts and pending verification through competing priorities.

Module 5: RA Lead and ML Lead: Changing the Model Without Losing Clearance

Deciding which changes need a new 510(k) and writing a PCCP that pre-authorizes the retraining the business actually plans.
4 lessons
  1. RA Specialist Task: Running Six NeuroFlag Change Requests Through 21 CFR 807.81(a)(3)(i) and FDA's Software-Change 510(k) Guidance
  2. RA Lead and ML Lead Task: Writing the Description of Modifications, Modification Protocol and Impact Assessment of a PCCP Under Section 515C
  3. Exercise: Turning a Vague Retraining Memo into Quantitative Aggregate and Per-Subgroup Acceptance Criteria
  4. US Manufacturer Handoff for an EU Expansion: Evidence Boundaries and Literature-Screening Recall
  • Decision Lab · 12 minThe Candidate That Improves the AverageThe failed candidate stays failed: preserve real owner/action receipts and pending verification through competing priorities.

Module 6: Post-Market Engineer and Complaint Handler: Living with NeuroFlag in the Field

Detecting drift after a scanner firmware change, deciding on CAPA and field action, triaging missed-bleed complaints for reportability, and scoping a generative-AI feature.
4 lessons
  1. Post-Market Engineer Task: Monitoring Flag Rates, Input Metadata and Report-Concordance Sensitivity After a Scanner Firmware Change
  2. Exercise: Control Charts on 18 Months of Synthetic Six-Site Data, and the CAPA, Part 806 and Within-PCCP Decision Memo
  3. Complaint Handler Task: Triaging 'The AI Missed a Bleed' Complaints for MDR Reportability Under 21 CFR 803 and 820.35
  4. Capstone: The NeuroFlag Evidence Packet and a Proposed LLM Findings Feature
  • Decision Lab · 12 minThe Quiet Outage and the Missing Study IDThe field handoff under interruption: preserve real owner/action receipts and pending verification through competing priorities.
  • Capstone simulation · 30 minThe Northbank Review BoardThe board receives the actual six-domain register: preserve real owner/action receipts and pending verification through competing priorities.

Exam and certificate

The exam unlocks once you have completed every lesson and played each Decision Lab to an ending. It draws 20 questions from a larger bank, spread across the modules, with the answer options shuffled for each attempt. You have 40 minutes and need 80% to pass. After a failed attempt you can try again after 24 hours, up to 3 attempts in total. Feedback is given per module, never as the answer key.

Passing earns a certificate of completion and assessment from the Wirewalk Institute of AI with a unique ID that anyone can check at institute.wirewalk.com/verify. It is not an accredited qualification or a vendor certification. How certification works.

The Wirewalk Institute of AI is independent. This course teaches publicly documented products and is not affiliated with or endorsed by Flywheel, V7, Ketryx, Jama Software, Veeva Systems, Medcrypt, Finite State, Basil Systems, Siemens (Polarion), or IBM (DOORS). Product and company names are trademarks of their respective owners.

More in AI in Health and Life Sciences

All 21 courses in this track
hls-001 Free intro Available

AI in Health and Life Sciences: A Practical Introduction

A free introduction built on eight healthcare jobs: auditing an ambient-scribe note, answering a portal message from an AI draft, acting on a sepsis alert, reading a triage-flagged CT worklist, reviewing a prior-authorization request, checking AI-suggested codes, and triaging drug and device safety reports. Each lesson shows where AI helps, where it must not decide, and the rule that governs that exact task.

3 h · 8 lessons Free intro
hls-110 Basic Available

AI for Nursing and Care Team Professionals

Use a synthetic US care-team case to check AI drafts against source records, preserve nursing judgment and coordinate accountable follow-up. Build a practical evidence packet without patient data or a model account.

10 h · 18 lessons See price at checkout
hls-210 Intermediate Available

Building and Deploying AI in Healthcare Delivery

Evaluate clinical AI through a fictional US health-system case: procurement, source audits, silent validation, monitoring, privacy and integration. Produce a bounded evidence portfolio using synthetic materials and branching decisions.

16 h · 24 lessons See price at checkout