FDA AI/ML Guidelines

The FDA's framework for AI/ML-based Software as a Medical Device (SaMD) requires continuous monitoring and demographic performance tracking.

Overview

Aspect	Details
Scope	AI/ML-based medical devices
Framework	Good Machine Learning Practice (GMLP)
Enforcement	FDA 510(k), De Novo, PMA pathways
Updates	Predetermined Change Control Plan (PCCP)

Key Requirements

Section IV.A: Performance Monitoring

"Manufacturers should monitor AI/ML device performance in the real-world setting"

ATHENA Solution:

• Real-time trust calibration

• Continuous accuracy tracking

• Performance degradation alerts

const engines = await athena.engines.status();
const metrics = await athena.system.metrics();

console.log('System health:', engines.overallStatus);
console.log('Error rate:', metrics.errorRate);

Section IV.B: Demographic Performance

"Manufacturers should evaluate device performance across demographic subgroups"

ATHENA Solution:

• Subgroup accuracy tracking

• Demographic disparity detection

• FDA-format reporting

const subgroups = await athena.bias.subgroups({ days: 90 });

for (const group of subgroups.data) {
  console.log(`${group.subgroup}: ${group.aiAccuracy}%`);
  if (group.biasDetected) {
    console.log(`  ⚠️ Disparity detected`);
  }
}

Section IV.C: Human Factors

"Manufacturers should understand how users interact with AI/ML outputs"

ATHENA Solution:

• Trust calibration analysis

• Automation bias detection

• Override quality tracking

const calibration = await athena.calibrate.analyze({
  userId: 'clinician_123',
  domain: 'radiology',
  timeWindowDays: 30
});

console.log('Calibration:', calibration.calibration);
console.log('Override quality:', calibration.overrideQuality);

Section V.A: Documentation

"Manufacturers should maintain documentation of AI/ML performance"

ATHENA Solution:

• Audit trail with full context

• FDA-format compliance reports

• Continuous monitoring logs

const report = await athena.export.generate({
  template: 'fda',
  format: 'pdf',
  startDate: '2025-01-01',
  endDate: '2025-03-31'  // Quarterly
});

FDA Report Contents

1. Performance Summary

• Overall accuracy metrics

• Sensitivity/specificity

• Error rate trends

2. Demographic Analysis

• Performance by age group

• Performance by sex

• Performance by ethnicity

• Performance by clinical indication

3. Human Factors Analysis

• User calibration scores

• Automation bias incidents

• Override patterns

• Training effectiveness

4. Adverse Events

• Misdiagnosis incidents

• User intervention required

• Near-misses logged

5. Predetermined Changes

• Model updates applied

• Performance impact

• Validation results

Subgroup Requirements

Subgroup	FDA Priority	ATHENA Tracking
Age (pediatric)	High	age_group:0-17
Age (geriatric)	High	age_group:65+
Sex	High	gender
Race/Ethnicity	High	ethnicity
Disease severity	Medium	Custom attribute
Clinical indication	Medium	Custom attribute

Implementation

Step 1: Track Clinical Decisions

await athena.audit.log({
  userId: 'radiologist_456',
  aiRecommendation: 'Positive for malignancy',
  userDecision: 'Positive for malignancy',
  confidence: 0.94,
  context: 'Mammography screening',
  metadata: {
    modality: 'mammography',
    patientAgeGroup: '55-64',
    patientSex: 'female',
    indication: 'screening'
  }
});

Step 2: Monitor Demographics

// Set up demographic bias monitoring
await athena.webhooks.create({
  url: 'https://hospital.com/fda-alerts',
  events: ['bias.detected'],
  filters: {
    severity: ['high', 'medium'],
    subgroups: ['age_group:0-17', 'age_group:65+']
  }
});

Step 3: Track User Performance

// Monitor clinician calibration
const atRisk = await athena.users.atRisk(10);

for (const user of atRisk.users) {
  console.log(`${user.userId}: ${user.calibration}`);
  if (user.calibration === 'OVERTRUSTING') {
    scheduleRetraining(user.userId);
  }
}

Step 4: Generate Quarterly Reports

// Automated quarterly FDA reports
const schedule = cron.schedule('0 0 1 1,4,7,10 *', async () => {
  const endDate = new Date();
  const startDate = new Date();
  startDate.setMonth(startDate.getMonth() - 3);
  
  const report = await athena.export.generate({
    template: 'fda',
    format: 'pdf',
    startDate: startDate.toISOString().split('T')[0],
    endDate: endDate.toISOString().split('T')[0]
  });
  
  await submitToFdaPortal(report);
});

Continuous Monitoring Dashboard

ATHENA Dashboard provides:

Metric	Refresh Rate	Alert Threshold
Overall accuracy	Real-time	<85%
Subgroup accuracy	Hourly	<80%
Automation bias	Real-time	Any detection
Override rate	Hourly	<10% or >50%

Sample FDA Report

FDA AI/ML PERFORMANCE REPORT
Device: RadAssist AI v2.1
510(k): K234567
Reporting Period: Q1 2025

PERFORMANCE SUMMARY
===================
Total Predictions: 12,456
True Positives: 892
True Negatives: 10,234
False Positives: 456
False Negatives: 78

Sensitivity: 92.0%
Specificity: 95.7%
Overall Accuracy: 94.8%

DEMOGRAPHIC ANALYSIS
====================
Age Group Performance:
  0-17:   N=234,  Accuracy=91.2%
  18-44:  N=3,456, Accuracy=95.1%
  45-64:  N=5,678, Accuracy=95.3%
  65+:    N=3,088, Accuracy=93.1%

Sex Performance:
  Female: N=9,234, Accuracy=94.9%
  Male:   N=3,222, Accuracy=94.6%

Disparity Detected: None exceeding 5% threshold

HUMAN FACTORS
=============
Clinicians Monitored: 45
Well-Calibrated: 41 (91%)
Overtrusting: 3 (7%)
Undertrusting: 1 (2%)

Automation Bias Incidents: 12
  - All low severity
  - Training provided to affected users

ADVERSE EVENTS
==============
Total: 2
  1. Missed diagnosis (age 72, caught on review)
  2. False positive (led to unnecessary biopsy)

COMPLIANCE STATUS: COMPLIANT

---

Next: California SB-53