Skip to main content

AI Features & Transparency

Complete transparency about where and how AI is used in Turtle RCM

Our AI Philosophy​

At Turtle RCM, we believe AI should augment, not replace clinical and billing expertise. Every AI feature:

  • ✅ Requires human review before taking action
  • ✅ Can be overridden by staff at any time
  • ✅ Learns from feedback to improve suggestions
  • ✅ Protects patient privacy through proper de-identification
  • ✅ Is clearly marked so you always know when AI is helping

AI Feature #1: Speech-to-Text Transcription​

What It Does​

Converts clinician dictation into written clinical notes.

Technology Used​

  • Service: Google Cloud Healthcare API
  • Purpose: Convert speech to text in real-time
  • Accuracy: Optimized for medical terminology

How It Works​

  1. Clinician dictates session notes using microphone
  2. Audio is streamed to Google Cloud Speech-to-Text API
  3. Text is returned in real-time and displayed
  4. Clinician reviews, edits, and finalizes the note
🔒 Privacy Protection

  • Audio is processed in real-time - never stored on Google servers
  • Encrypted in transit using TLS 1.3
  • No audio retention - deleted immediately after transcription
  • Google cannot use your data for training (per Healthcare API terms)

Human Oversight​

  • Always required: Clinician must review and approve all transcribed text
  • Edit capability: Full editing available before saving
  • Quality check: System highlights potential transcription errors

AI Feature #2: CPT Code Suggestions​

What It Does​

Suggests appropriate billing codes (CPT) based on session content and duration.

Technology Used​

  • Service: AI-powered code selection
  • Training Data: De-identified synthetic and aggregated healthcare data

How It Works​

  1. System analyzes session duration, complexity, and clinical content
  2. ML model suggests the most appropriate CPT code(s)
  3. Suggestion is displayed with confidence level and reasoning
  4. Biller reviews, approves, or selects different code

Key Features​

  • Considers multiple factors: Duration, diagnosis, complexity
  • Explains reasoning: "Based on 55-minute session with crisis intervention documented"
  • Confidence scores: Shows how certain the model is
  • Easy override: One-click to select alternative code
Model Training & Data Protection

Training Data Sources:
  • Synthetic healthcare datasets
  • De-identified aggregated claims data (with proper Safe Harbor compliance)
Your clinic's identifiable data is NEVER used for training without:
  • ✅ Complete HIPAA Safe Harbor de-identification
  • ✅ Removal of all 18 HIPAA identifiers
  • ✅ Aggregation across multiple clinics
  • ✅ Tokenization of all names and locations

Human Oversight​

  • Always required: Biller must approve code before claim submission
  • Override tracking: System learns from your corrections
  • Audit trail: All code changes are logged

AI Feature #3: Claim Denial Prevention​

Feature in Development

Advanced denial prediction and analytics are currently under development.

Current capabilities:

  • ✅ Basic compliance checks before submission
  • ✅ Manual denial tracking
  • 🔜 Predictive denial warnings (Phase 2)
  • 🔜 Pattern analytics (Phase 2)

See Billing & Claims for current denial management workflows.

AI Feature #4: Appeals Letter Assistance​

Feature in Development

Automated appeal letter generation is currently under development.

Current capabilities:

  • ✅ Manual appeal tracking
  • ✅ Denial reason documentation
  • 🔜 AI-assisted letter drafting (Phase 2)
  • 🔜 Automated policy citations (Phase 2)

See Billing & Claims for current appeal workflows.

Where AI is NOT Used​

To maintain trust and compliance, AI is never used for:

  • ❌ Clinical diagnosis - Only clinicians diagnose
  • ❌ Treatment decisions - Only clinicians create treatment plans
  • ❌ Automatic claim submission - Always requires approval
  • ❌ Access control - Authentication is traditional (JWT-based)
  • ❌ Data breach detection - Uses traditional security monitoring
  • ❌ Financial decisions - No AI in billing rate setting

Data Privacy & Security​

HIPAA Safe Harbor Compliance​

All AI training data goes through our Data Safe Harbor module:

  1. Extraction: Raw data collected from clinic databases
  2. De-identification: Removes all 18 HIPAA identifiers:
    • Names (patients, relatives, providers)
    • Geographic subdivisions smaller than state
    • Dates (except year)
    • Phone/fax numbers
    • Email addresses
    • Social Security numbers
    • Medical record numbers
    • Account numbers
    • Certificate/license numbers
    • Vehicle identifiers
    • Device identifiers
    • URLs
    • IP addresses
    • Biometric identifiers
    • Photos
    • Other unique identifiers
  3. Tokenization: Replaces names/locations with random tokens
  4. Aggregation: Combines data across multiple clinics
  5. Storage: Isolated training database (separate from production)
🛡️ Technical Achievement: Safe Harbor Module

Our Data Safe Harbor implementation ensures:

  • Zero linkability: Training data cannot be traced back to individuals
  • Randomized tokens: No connection between original and de-identified data
  • Isolated storage: Training data completely separate from production
  • Audit logging: All Safe Harbor access logged to GCP
  • No 3rd party access: De-identification happens in-house
Your clinic's identifiable patient data never leaves our secure environment without proper de-identification.

Google Cloud AI Service Protections​

When using Google Cloud AI services:

  • Healthcare API: Uses HIPAA-compliant endpoints
  • Data Processing Addendum: Google signs BAA (Business Associate Agreement)
  • No data retention: Google cannot store or use your data
  • Encrypted transit: TLS 1.3 encryption
  • Encrypted at rest: AES-256 encryption
  • Regional isolation: Data stays in US regions

AI Transparency Commitments​

We commit to:

  1. Clear labeling: Every AI feature is marked with an AI badge
  2. Explanation: You always see why AI made a suggestion
  3. Control: You can override any AI recommendation
  4. Opt-out: AI features can be disabled (except transcription)
  5. Feedback: System learns from your corrections
  6. Documentation: This page explains every AI use case
  7. Updates: We notify you when AI features change

Improving AI Performance [WIP]​

Your Feedback Matters​

When you override an AI suggestion, the system learns:

  • Code corrections: Improves future CPT suggestions
  • Denial patterns: Better denial prediction
  • Appeal outcomes: Refines appeal letter templates
📊 Continuous Improvement

How feedback works:
  1. You override an AI suggestion (e.g., change CPT code)
  2. System logs the override (anonymized)
  3. Monthly model retraining incorporates feedback
  4. Next version of model is more accurate
Your expertise makes the AI better for everyone.

Model Updates​

  • Frequency: Models retrained monthly with new data
  • Testing: Validation on holdout dataset before deployment
  • Rollback: Previous model version kept for 30 days
  • Notification: Admins notified of model updates
  • Performance tracking: Accuracy metrics monitored

AI Limitations​

What AI Cannot Do​

Current Limitations:

  • Cannot understand complex clinical nuance
  • May miss rare or unusual cases
  • Depends on quality of input data
  • Can be fooled by incomplete documentation
  • Does not replace clinical judgment

Known Issues:

  • Speech-to-text accuracy drops with background noise
  • Code suggestions may be conservative (under-coding risk)
  • Denial prediction is probabilistic (not certain)
  • Appeal letters may need significant editing

When to Trust AI Less​

Be extra cautious when:

  • 🚨 Case is unusual or complex
  • 🚨 Patient has rare diagnosis
  • 🚨 Documentation is incomplete
  • 🚨 You disagree with AI suggestion
  • 🚨 High financial or clinical stakes

Always trust your professional judgment over AI.

Questions & Concerns​

Frequently Asked Questions​

Q: Can AI make mistakes?
A: Yes. AI is a tool that can make errors, which is why human review is always required.

Q: Will AI replace billers or clinicians?
A: No. AI is designed to assist, not replace. It handles repetitive tasks so you can focus on complex cases.

Q: What if I disagree with an AI suggestion?
A: Always override it. Your expertise is paramount. The system will learn from your corrections.

Q: Is my clinic's data used to train Google's models?
A: No. Per our Business Associate Agreement, Google cannot use your data for their own training.

Q: Can I see how AI made a decision?
A: Yes. Every AI suggestion includes an explanation of key factors.

Q: Can I turn off AI features?
A: Most AI features can be disabled per user or per clinic. Contact your administrator.

Regulatory Compliance​

FDA Guidance​

Turtle RCM's AI features are classified as Clinical Decision Support (CDS) tools that:

  • Display/analyze medical information
  • Provide recommendations (not decisions)
  • Allow human review and override
  • Are therefore not subject to FDA premarket review per FDA guidance

ONC Certification​

Our FHIR implementation follows ONC (Office of the National Coordinator) standards for health IT certification, including:

  • Transparent AI/ML information
  • Source attribute listing
  • Predictive model intervention
  • User notification of AI involvement

Contact & Feedback​

Have questions about AI features? Contact:

  • Your clinic administrator
  • Turtle RCM support team
  • Refer to the FAQ for common questions
We value your feedback!

If you notice AI behaving unexpectedly or have suggestions for improvement, please report it to your administrator. Your insights help us build better AI tools.

Last updated: November 2025 | AI Features Version 1.0