laxmanhalaki f3e58dd50a documents added for the code generation context

2026-03-13 19:19:53 +05:30

43 KiB

Raw Blame History

Software Requirements Specification

AI-Powered Medical Billing Automation System

For Out-of-Network Neurosurgical & Orthopedic Practices

Prepared by: Dextra Labs Pte Ltd
Document Version: 2.
Date: February 2026

Document Information

Client Dr. Brian McHugh / McHugh Neurosurgery

Project Name AI-Powered Medical Billing Automation

Implementation Partner Dextra Labs

Development Partner Tech4Biz Solutions Pvt Ltd

Document Status v2.0 — Development-Ready

1. Introduction

1.1 Purpose

This Software Requirements Specification (SRS) document provides a comprehensive, development-ready description of the AI-Powered Medical Billing Automation System. It details functional and non-functional requirements, system architecture, interfaces, constraints, and milestone-aligned deliverables for building an intelligent billing automation platform for out-of- network neurosurgical and orthopedic practices.

This document serves as the primary reference for developers, project managers, technical leads, QA personnel, and stakeholders throughout the development lifecycle. All requirements are mapped to SOW milestones to ensure traceability from specification to delivery.

1.2 Scope

The system is designed to transform the end-to-end claim generation process for specialty medical practices. [UPDATED] The MVP focuses on 5-10 common spine surgery types (covering approximately 98-99% of typical procedures) for a single practice with up to 5 providers.

The system will:

Convert clinical audio dictation to text using Whisper ASR with medical-specialized vocabulary
Extract clinical entities (diagnoses, procedures, anatomical locations) using NLP
Map extracted entities to ICD-10 diagnosis codes and CPT procedure codes
Apply payer-specific billing optimization rules to maximize reimbursement
Perform RAG-powered claim scrubbing against NCCI edits and LCD/NCD coverage determinations
Provide human-in-the-loop review workflow for mandatory claim verification
Integrate with existing EMR systems (Epic, Athena Centricity, CureMD) via FHIR and REST APIs
Support a template-based fast-track workflow for standard procedures (e.g., standard 2- level ACDF)
Maintain HIPAA compliance through cloud-hosted, VPC-isolated AI infrastructure

1.3 System Boundary

1.3.1 In Scope (Our System)

Audio capture via mobile application (smartphone-based, no dedicated hardware)
Speech-to-text conversion with medical terminology using Whisper STT
EMR data retrieval (patient demographics, insurance, clinical documentation) from Epic, Athena, CureMD
AI-powered entity extraction and ICD-10/CPT code mapping with confidence scoring
RAG-powered claim scrubbing against payer-specific rules, NCCI edits, and LCD/NCD
Payer-specific optimization and claim generation (CMS-1500 professional claims)
Template-based fast-track workflow for standard procedures
Human review workflow with mandatory approval before finalization
Export of approved claims for downstream processing
Audit trails, compliance reporting, and analytics dashboard
Provider interface supporting dictate → review → approve → submit workflow

1.3.2 Out of Scope (Existing Practice Workflow)

Claim submission to clearinghouse or insurance payers (EDI 837P/837I transmission)
Payment posting and ERA/835 processing
Patient billing and collections
Denial appeals and follow-up automation
Accounts receivable management
Operative report generation/amendment (future phase consideration)
ModMed integration (mentioned in original brief, formally excluded from MVP)
UB-04 institutional claims (MVP covers CMS-1500 professional claims only)

Note: The practice will continue using their existing systems (Athena Centricity) for claim submission and payment processing. Our system generates the approved claim and hands off to their existing workflow.

[NEW] 1.3.3 Out-of-Network Billing Lifecycle Context

Developers must understand the out-of-network (OON) billing lifecycle to build this system effectively. The typical OON claim lifecycle is: Procedure → Coding & Claim Generation (our system) → Submission to Payer → Initial Denial (common for OON) → Appeal with Documentation → Arbitration (if needed) → Payment. Our system’s output directly impacts denial rates and appeal success by ensuring accurate coding, proper modifier application, and comprehensive documentation from the start.

1.4 Definitions, Acronyms, and Abbreviations

Term Definition

ASR Automatic Speech Recognition — converts spoken words to text
CPT Current Procedural Terminology — medical procedure codes maintained by
AMA

EMR/EHR Electronic Medical/Health Record — digital patient medical history
FHIR Fast Healthcare Interoperability Resources — healthcare data exchange
standard
HIPAA Health Insurance Portability and Accountability Act — US healthcare privacy
law
ICD- 10 International Classification of Diseases, 10th Revision — diagnosis coding
system
LCD/NCD Local/National Coverage Determination — payer coverage policies

LLM Large Language Model — AI model trained on large text datasets

LoRA Low-Rank Adaptation — technique for efficient LLM fine-tuning
MDM Medical Decision Making — complexity level used for E/M coding

MVP Minimum Viable Product — initial product version with core features

NCCI National Correct Coding Initiative — CMS edits to prevent improper code
pairs
NLP Natural Language Processing — AI techniques for understanding text

OON Out-of-Network — healthcare providers not contracted with insurance plans

PHI Protected Health Information — patient health data protected by HIPAA
RAG Retrieval-Augmented Generation — AI technique combining semantic search
with LLM
STT Speech-to-Text — audio to text conversion technology

WER Word Error Rate — standard metric for measuring STT accuracy

1.5 References

SO070FY26: Real-Time Dictation and Automated Billing AI Platform (December 2025)
SOW: AI Dictation Application Development, Dextra Labs v
Product Vision Document — Dr. Brian McHugh Discovery Session Notes (1st & 2nd calls)
Project Kickoff Presentation (February 2026)
AI-Enabled Real-Time Dictation and Automated Billing Project Brief
HIPAA Security Rule (45 CFR Part 164)
HL7 FHIR R4 Specification
Epic FHIR API Documentation

2. Overall Description

2.1 Product Perspective

The AI-Powered Medical Billing Automation System addresses a significant gap in the healthcare revenue cycle management market. While in-network billing has seen substantial automation advances, out-of-network specialty billing remains largely manual due to complex payer-specific coding requirements, high-value procedures requiring precise documentation, institutional knowledge locked within experienced human coders, and stringent HIPAA requirements limiting cloud-based AI adoption.

This system serves as a standalone application integrating with existing EMR infrastructure while providing its own AI-powered processing layer.

2.2 Why Historical Billing Data is Critical

The system requires access to historical billing data from the practice (filtered for relevance). This data serves several critical purposes:

Pattern Recognition: Learning which CPT codes are most successful for specific procedures with specific payers
Denial Avoidance: Understanding which code combinations and documentation patterns lead to claim denials
Payer-Specific Optimization: Different insurers have different preferences (e.g., BCBS prefers certain code sequences while Cigna accepts others)
Institutional Knowledge Capture: Expert coders have accumulated decades of knowledge about what works

Important: Historical data will be filtered to exclude outdated policy periods. For example, if United Healthcare restructured policies 3 years ago, denial patterns from before that date are excluded as no longer relevant. Data filtering criteria must be defined per-payer based on policy change dates.

2.3 Product Functions

2.3.1 Audio Capture and Transcription

Smartphone-based audio recording for clinical dictation (no dedicated hardware in MVP)
AI-powered speech-to-text conversion with medical vocabulary fine-tuning
Noise reduction optimized for hospital environments (post-procedure hallway, dictation room, OR)
Single speaker per dictation session (clinic claims may have PA voice, surgical claims have surgeon voice)

2.3.2 Clinical Entity Extraction

Automated extraction of diagnoses, procedures, anatomical locations, and laterality
Temporal relationship mapping for procedure sequencing
Confidence scoring for each extracted entity

2.3.3 Medical Code Mapping

Automatic mapping of diagnoses to ICD-10 codes and procedures to CPT codes
Modifier suggestions based on context (-59, -51, -LT/RT, etc.)
Confidence scoring with alternative code suggestions for edge cases

[NEW] 2.3.4 Template-Based Fast-Track Workflow

For standard, high-volume procedures (e.g., standard 2-level ACDF), the system provides a fast-track workflow where the surgeon selects a procedure template and associates it with a patient. The system auto-generates the claim with pre-mapped CPT/ICD codes, modifiers, and justification. The surgeon can add dictation notes for any deviations. This significantly reduces time-to-claim for routine procedures.

2.3.5 RAG-Powered Claim Scrubbing

Automated claim review using Retrieval-Augmented Generation against payer-specific rules, NCCI edits, and Local/National Coverage Determinations (LCD/NCD). The RAG corpus includes policy documents per individual plan, coding manuals, billing guidelines, and the practice’s laminated cheat sheets.

2.3.6 Payer-Specific Optimization

Business rules engine for payer-specific coding strategies
Historical denial pattern analysis for proactive optimization
Coverage for top 30 local insurance plans in MVP (starting with top 10 for initial validation)

2.3.7 Claim Generation and Review

Automated generation of submission-ready CMS- 15 00 professional claims
Mandatory human review interface with AI-generated content highlighting
Correction workflow with feedback captured for model improvement
Complete audit trails for compliance and quality tracking

2.4 User Classes and Characteristics

User Class Description Technical Proficiency

Surgeon/Physician Primary users who record clinical dictation via
mobile app. Up to 5 providers in MVP.

Low — Requires simple,
intuitive interface
Medical Biller Reviews AI-generated claims, makes
corrections, approves submissions

Moderate — Familiar with
billing systems

Billing Supervisor Oversees billing operations, manages rules
and policies

Moderate — Dashboard and
reporting focus

System
Administrator

Manages system configuration, users, and
integrations

High — Technical
administration

2.5 Operating Environment

Web Application: Modern browsers (Chrome, Firefox, Safari, Edge) on Windows, macOS
Mobile Application: iOS 14+ and Android 11+ smartphones for audio capture
Server Infrastructure: Cloud-hosted (AWS/GCP/Azure) with HIPAA-compliant VPC isolation
AI Infrastructure: GPU-enabled cloud instances or Mac Mini M-series for LLM inference
Database: PostgreSQL for structured data; vector database for RAG/semantic search
Network: Always-online connectivity required (no offline mode in MVP)

2.6 Design and Implementation Constraints

HIPAA Compliance: All PHI must remain within HIPAA-compliant, VPC-isolated infrastructure
Self-Hosted AI: No external API calls for patient data processing (LLM inference is self- hosted)
Human-in-the-Loop: Mandatory human review before any claim finalization
Single Practice: MVP limited to Dr. McHugh’s practice only (up to 5 providers)
Surgical Focus: MVP focused on 5-10 common spine surgery types (CMS- 1500 professional claims only)
English Only: Speech-to-text optimized for English medical dictation, single speaker per session
Open-Source LLM: Llama/Mistral/Mixtral/Qwen with LoRA fine-tuning

2.7 Assumptions and Dependencies

2.7.1 Assumptions

Internet connectivity is always available at all practice locations
Practice will provide access to historical billing and denial data (filtered for relevance)
Payer policy documents, coding manuals, and cheat sheets will be provided for RAG corpus
Human billers will be available for review and correction during MVP
Epic FHIR API access will be granted through proper channels (timeline dependency)
No major scope changes during the execution of agreed phases
Client will provide list of 5-10 common spine surgeries with typical CPT/ICD code combinations
Client will provide prioritized list of top 30 insurance plans for payer rules configuration

2.7.2 Dependencies

Epic Systems: FHIR R4 API availability and access credentials (critical path — can take 3 - 6 months)
Athena Centricity: Legacy system data export capabilities and API availability
CureMD: REST API documentation and access
Open-source LLM: Availability of Llama 2/3 or Mistral/Mixtral models for self-hosting
Whisper ASR: Model availability for self-hosting with medical vocabulary fine-tuning
Client SMEs: Availability of billing and clinical subject matter experts for validation

3. Functional Requirements

All functional requirements are mapped to SOW milestones for traceability. Priority levels: Must (MVP-critical), Should (important but can be deferred), Could (nice-to-have).

3.1 Audio Capture Module (Mobile App) — Milestone 2

Req ID Requirement Description Milestone Priority
FR-AC-
001

System shall allow users to record audio dictation using
smartphone microphone

M2 Must

FR-AC-
002

System shall support audio recording in standard formats
(AAC, MP3, WAV)

M2 Must

FR-AC-
003

System shall display recording duration and audio level
indicator

M2 Must

FR-AC-
004

System shall allow pause and resume during recording
session

M2 Must^

FR-AC-
005

System shall require user to associate recording with patient
identifier (MRN or encounter ID from EMR lookup)

M2 Must^

FR-AC-
006

System shall encrypt audio files at rest on device before
upload (AES-256)

M2 Must^

FR-AC-
007

System shall automatically upload recordings when network is
available

M2 Must^

FR-AC-
008

System shall support multiple recording sessions per patient
encounter

M2 Must^

FR-AC-
009

System shall support template-based fast-track: surgeon
selects standard procedure template and patient, bypassing
full dictation

M2 Must^

3.2 Speech-to-Text Module — Milestones 2- 3

Req ID Requirement Description Milestone Priority

FR-ST- 001 System shall convert audio dictation to text with ≥97% word
accuracy (WER) on a mutually agreed medical dictation test
corpus

M3 Must^

FR-ST- 002 System shall use medical-specific vocabulary (ICD-10 terms,
CPT terms, drug names, anatomical terms) for recognition

M3 Must

FR-ST- 003 System shall recognize ICD-10 and CPT codes when spoken
directly

M3 Should

FR-ST- 004 System shall handle medical drug names and dosages
accurately

M3 Must

FR-ST- 005 System shall apply AI-powered noise reduction for hospital
environments (hallway, dictation room, OR)

M3 Must

FR-ST- 006 System shall provide transcript with timestamps for review M2 Must

FR-ST- 007 System shall allow manual correction of transcription errors M2 Must^
FR-ST- 008 System shall mark low-confidence words/phrases for human
review

M3 Must^

3.3 Clinical Entity Extraction Module — Milestone 3

Req ID Requirement Description Milestone Priority

FR-EE- 001 System shall extract diagnoses from clinical documentation M3 Must^
FR-EE- 002 System shall identify procedures and treatments performed M3 Must^
FR-EE- 003 System shall recognize anatomical locations and laterality
(left/right)

M3 Must^

FR-EE- 004 System shall extract temporal relationships between
procedures

M3 Should^

FR-EE- 005 System shall identify patient demographics from EMR
integration

M2 Must^

FR-EE- 006 System shall extract insurance/payer information for the
encounter

M2 Must^

FR-EE- 007 System shall provide confidence scores for each extracted
entity

M3 Must^

3.4 Code Mapping Module — Milestone 3

Req ID Requirement Description Milestone Priority
FR-CM-
001

System shall map extracted diagnoses to ICD-10 codes M3 Must

FR-CM-
002

System shall map procedures to appropriate CPT codes M3 Must

FR-CM-
003

System shall suggest appropriate CPT modifiers based on
context (-59, -51, -LT/RT, etc.)

M3 Must

FR-CM-
004

System shall provide confidence scores for each code
mapping

M3 Must^

FR-CM-
005

System shall suggest alternative codes for low-confidence
mappings (<80%)

M3 Must^

FR-CM-
006

System shall validate code combinations against NCCI/CCI
edits

M3 Must^

FR-CM-
007

System shall support neurosurgery-specific code sets (5- 10
common spine procedures)

M3 Must^

FR-CM-
008

System shall support orthopedic surgery-specific code sets M3 Should^

FR-CM-
009

System shall apply confidence thresholds: >90% auto-
suggest, 70-90% flag for review, <70% require manual coding

M3 Must^

FR-CM-
010

System shall determine and assign Medical Decision Making
(MDM) level based on clinical documentation complexity

M3 Must^

FR-CM-
011

System shall generate medical necessity justification text to
support assigned codes

M3 Must^

3.5 RAG-Powered Claim Scrubbing Module — Milestone 3

Req ID Requirement Description Milestone Priority

FR-RS-
001

System shall scrub generated claims against payer-specific
rules using RAG

M3 Must^

FR-RS-
002

System shall validate claims against NCCI edits for improper
code pair detection

M3 Must^

FR-RS-
003

System shall check claims against Local Coverage
Determinations (LCD)

M3 Must

FR-RS-
004

System shall check claims against National Coverage
Determinations (NCD)

M3 Must

FR-RS-
005

System shall flag claims failing scrubbing rules with specific
failure reasons

M3 Must

FR-RS-
006

System shall suggest corrective actions for failed scrubbing
checks

M3 Should

FR-RS-
007

RAG corpus shall include: policy docs per plan, coding
manuals, cheat sheets, billing guidelines

M3 Must

3.6 Payer Rules Engine — Milestones 3- 4

Req ID Requirement Description Milestone Priority
FR-PR-
001

System shall maintain payer-specific coding rules for top 30
local plans

M3 Must

FR-PR-
002

System shall apply different coding strategies based on payer M3 Must^

FR-PR-
003

System shall incorporate historical denial pattern analysis M3 Must^

FR-PR-
004

System shall flag claims at high risk of denial based on
historical patterns

M4 Must^

FR-PR-
005

System shall optimize code selection for maximum
reimbursement

M3 Must^

FR-PR-
006

System shall allow manual update of payer rules by billing
staff

M4 Must^

FR-PR-
007

System shall track payer rule changes with version history M4 Should^

3.7 Claim Generation Module — Milestone 4

Req ID Requirement Description Milestone Priority
FR-CG-
001

System shall generate complete billing claims from processed
data

M4 Must

FR-CG-
002

System shall populate all required CMS-1500 claim fields
automatically

M4 Must

FR-CG-
003

System shall consolidate multi-session audio into single
patient claim

M4 Must

FR-CG-
004

System shall validate claim completeness before presenting
for review

M4 Must

FR-CG-
005

System shall generate claims in CMS-1500 professional
format

M4 Must

FR-CG-
006

System shall support multi-session audio consolidation
including recordings spanning multiple days for a single
surgical encounter

M4 Must

FR-CG-
007

System shall implement claim state machine with defined
states: Draft, Pending STT, Pending AI Review, Ready for
Human Review, Approved, Rejected, Exported

M4 Must

3.8 Human Review Workflow — Milestone 4

Req ID Requirement Description Milestone Priority

FR-HR-
001

System shall present claims for mandatory human review
before finalization

M4 Must^

FR-HR-
002

System shall highlight AI-generated content for reviewer
attention

M4 Must^

FR-HR-
003

System shall display original audio/transcript alongside
generated claim (side-by-side view)

M4 Must^

FR-HR-
004

System shall allow reviewers to modify any claim field M4 Must^

FR-HR-
005

System shall track all reviewer corrections for model
retraining feedback loop

M4 Must^

FR-HR-
006

System shall require explicit approval before claim finalization M4 Must

FR-HR-
007

System shall support claim rejection with reason capture and
re-routing to biller for manual correction

M4 Must

FR-HR-
008

System shall maintain complete audit trail of all claim
modifications

M4 Must

FR-HR-
009

System shall display EMR data (demographics, insurance)
alongside claim for cross-reference

M4 Must

3.9 EMR Integration — Milestones 2- 4

Req ID Requirement Description Milestone Priority
FR-EMR-
001

System shall integrate with Epic via FHIR R4 API for hospital
records (operative reports, clinical notes)

M4 Must

FR-EMR-
002

System shall integrate with Athena Centricity for billing data,
denied claims, patient records

M2 Must

FR-EMR-
003

System shall integrate with CureMD via REST API for patient
demographics and clinical docs

M2 Must

FR-EMR-
004

System shall retrieve patient demographics from EMR M2 Must

FR-EMR-
005

System shall retrieve insurance/payer information from EMR M2 Must

FR-EMR-
006

System shall retrieve clinical documentation and notes from
EMR

M2 Must

FR-EMR-
007

System shall handle EMR connection failures gracefully with
retry logic

M4 Must

3.10 Reporting and Analytics — Milestone 5

Req ID Requirement Description Milestone Priority
FR-RA-
001

System shall provide dashboard with claim processing status
(pending, approved, rejected, exported)

M5 Must^

FR-RA-
002

System shall display AI accuracy metrics (STT WER, code
mapping accuracy)

M5 Must^

FR-RA-
003

System shall track and report claim approval/rejection rates M5 Must^

FR-RA-
004

System shall provide human correction rate analytics (per
coder, per payer)

M5 Must^

FR-RA-
005

System shall generate audit reports for compliance review M5 Must^

FR-RA-
006

System shall support search/filter claims by patient, date
range, status, payer

M5 Must^

FR-RA-
007

System shall track same-day charge capture rate (target:
80%)

M5 Must^

[NEW] 3.11 Notification System — Milestone 4

Req ID Requirement Description Milestone Priority
FR-NT- 001 System shall notify billers when a new claim is ready for
review

M4 Must

FR-NT- 002 System shall notify providers when a claim is approved or
requires attention

M4 Should

FR-NT- 003 System shall support in-app push notifications on mobile M4 Should

FR-NT- 004 System shall support email notifications for critical events M4 Should^

[NEW] 3.12 Data Migration & Model Training Pipeline — Milestones 1, 3,

6

Req ID Requirement Description Milestone Priority

FR-DM-
001

System shall provide ETL pipeline to extract, clean, and load
historical billing data from Athena Centricity

M1 Must^

FR-DM-
002

System shall filter historical data per-payer based on policy
change dates to exclude outdated patterns

M1 Must^

FR-DM-
003

System shall ingest and index policy documents, coding
manuals, and cheat sheets into the RAG vector store

M1 Must^

FR-DM-
004

System shall capture human reviewer corrections in a
structured format suitable for model retraining

M4 Must^

FR-DM-
005

System shall support periodic model retraining/fine-tuning
using accumulated correction data (manual trigger in MVP)

M6 Should^

4. Non-Functional Requirements

4.1 Performance Requirements

Req ID Requirement Target Milestone
NFR-P-
001

Speech-to-text processing time per minute of
audio

< 90 seconds M

NFR-P-
002

Claim generation time from completed transcript < 90 seconds M

NFR-P-
003

Web dashboard page load time < 3 seconds M

NFR-P-
004

Mobile app recording start latency < 2 seconds M

NFR-P-
005

Provider submission workflow completion
(dictate to approve)

< 1 minute M

NFR-P-
006

Concurrent users supported 20+ simultaneous M

4.2 Security Requirements

Req ID Requirement Milestone

NFR-S-
001

All PHI shall be encrypted at rest using AES-256 encryption M

NFR-S-
002

All data in transit shall be encrypted using TLS 1.3 M

NFR-S-
003

User authentication shall use OAuth 2.0 with multi-factor authentication M

NFR-S-
004

Role-based access control shall restrict data access by user role M

NFR-S-
005

All user actions shall be logged in tamper-proof audit trail M

NFR-S-
006

Session timeout shall occur after 15 minutes of inactivity M

NFR-S-
007

AI inference shall occur entirely within VPC-isolated, HIPAA-compliant
infrastructure

M

NFR-S-
008

No PHI shall be transmitted to external cloud AI services M

4.3 Reliability and Availability

Req ID Requirement Milestone

NFR-R-
001

System availability target: 99.5% during business hours (6 AM - 10 PM
ET)

M

NFR-R-
002

Mean time to recovery from failure: < 4 hours M

NFR-R-
003

Database backup frequency: Daily with 30-day retention M

NFR-R-
004

Audio file retention: 7 years (aligned with HIPAA audit trail requirements) M

NFR-R-
005

Graceful degradation when EMR integration is unavailable M

4.4 Compliance Requirements

Req ID Requirement Milestone
NFR-C-
001

System shall comply with HIPAA Privacy Rule requirements M

NFR-C-
002

System shall comply with HIPAA Security Rule requirements M

NFR-C-
003

Business Associate Agreement (BAA) required with all cloud providers M

NFR-C-
004

Complete audit trails shall be maintained for minimum 7 years M

NFR-C-
005

System shall support compliance reporting for audits M

4.5 Usability Requirements

Req ID Requirement Milestone

NFR-U-
001

Mobile app shall require < 3 taps to start recording M

NFR-U-
002

New billers shall be productive within 4 hours of training M

NFR-U-
003

Web interface shall be accessible on screen sizes >= 1280px width M

NFR-U-
004

System shall provide clear error messages with suggested corrective
actions

M

5. System Architecture

5.1 Architecture Overview

The system follows a modular, layered architecture designed for scalability and future expansion to agentic AI capabilities. The architecture separates concerns across four distinct layers.

5.1.1 Experience Layer

Web Dashboard (React.js + TypeScript): Primary interface for billers to review claims and view analytics
Mobile App (React Native / Expo): Smartphone-based audio capture, template selection, and claim status
REST API Gateway: Secure API with OAuth 2.0 authentication, rate limiting, and routing

5.1.2 Processing Layer

Workflow Engine: Orchestrates the claim generation pipeline (audio → transcript → entities → codes → claim)
Business Rules Engine: Implements payer-specific billing rules and optimization logic
Claim Scrubbing Engine: RAG-powered validation against NCCI, LCD/NCD, and payer rules
NLP Processor: Coordinates speech-to-text and clinical entity extraction

5.1.3 AI Layer

Open Source LLM: Self-hosted Llama/Mistral/Mixtral/Qwen with LoRA fine-tuning
Speech-to-Text Engine: Whisper with medical vocabulary fine-tuning
Vector Database: For semantic search and RAG (pgvector or Weaviate/Pinecone)

5.1.4 Data Layer

Knowledge Base: Vector database for semantic search across policy docs, coding manuals, cheat sheets
Historical Data: Billing and denial history (filtered per-payer by policy change dates)
Operational Database: PostgreSQL for claims, users, transactions, and audit logs

5.2 Technology Stack

Layer Technology Rationale
Frontend React.js + TypeScript Modern, maintainable UI

Mobile React Native / Expo Cross-platform iOS/Android
API Server FastAPI (Python) High performance, async, ML
ecosystem

LLM Llama/Mistral/Mixtral + LoRA Open source, fine-tunable, self-
hostable

Speech-to-Text Whisper (OpenAI) High accuracy, self-hostable, medical
fine-tuning

Database PostgreSQL + pgvector Structured + vector search in single
DB
Cache Redis Session management, API caching

Hosting AWS/GCP/Azure (HIPAA BAA) Cloud-hosted, VPC isolation, scalable

5.3 Deployment Architecture

The MVP deployment is designed for cloud-hosted operation within a HIPAA-compliant VPC:

Application Server: Cloud VM (8+ cores, 32GB RAM) for API, workflow engine, and web app
AI Inference: GPU-enabled cloud instance or Mac Mini M-series for LLM and Whisper
Database: PostgreSQL with pgvector extension on managed service (RDS/Cloud SQL)
Storage: Encrypted object storage for audio files and documents
Estimated monthly infrastructure cost: $1,100-1,600/month

6. EMR Integration Specifications

6.1 Epic Integration (Hospital)

Attribute Specification
Integration Method FHIR R4 API

Authentication OAuth 2.0 with SMART on FHIR
Data Retrieved Patient demographics, encounters, operative reports, clinical notes,
insurance

Environment Hospital inpatient and surgical records
Risk Epic FHIR approval can take 3-6 months. Timeline contingency required.

6.2 Athena Centricity Integration (Outpatient Legacy)

Attribute Specification

Integration Method Custom API Connector (legacy Centricity version — confirm actual version
and API limitations)

Data Retrieved Historical billing data, denied claims with reason codes, patient records
Note Existing connector available — reduces integration time. Emily has flagged
potential roadblocks with older Centricity version vs modern Athena API.

6.3 CureMD Integration (Outpatient)

Attribute Specification
Integration Method REST API

Data Retrieved Patient demographics, appointments, clinical documentation
Note Existing connector available — reduces integration time

[NEW] 6.4 ModMed — Formally Excluded from MVP

ModMed was listed in the original project brief but has been dropped from all subsequent documents. It is formally excluded from MVP scope. If needed in a future phase, a separate integration SOW will be required.

7. MVP Scope Definition

7.1 In Scope (MVP)

Integration with Epic (hospital) and Athena Centricity/CureMD (outpatient)
Speech-to-text with medical dictionary (smartphone-based capture via Expo/React Native)
AI-powered claim generation from clinical documentation (5-10 common spine surgery types)
RAG-powered claim scrubbing against NCCI edits and LCD/NCD coverage determinations
Payer-specific CPT code optimization for top 30 local plans (starting with top 10)
Template-based fast-track workflow for standard procedures
Human review interface with mandatory approval workflow
Complete audit trails for compliance
Manual policy updates by billing staff
Single-practice deployment (Dr. McHugh’s practice, up to 5 providers)
CMS-1500 professional claims (surgical billing focus)
Provider interface supporting dictate → review → approve → submit workflow
Analytics dashboard with claim status, AI accuracy, and correction rates

7.2 Out of Scope (Future Phases)

Self-learning/auto-updating rules engine
Agentic AI capabilities
Dedicated hardware dictation devices
Multi-accent speech recognition optimization
Offline/mobile-only processing
Multi-practice/multi-tenant deployment
Full automation without human review
Outpatient clinic visit (E/M) billing automation
Direct clearinghouse/EDI submission integration (837P/837I)
Denial appeals automation
Payment posting and ERA/835 processing
Patient billing and collections
UB-04 institutional claims
ModMed integration
Operative report generation/amendment

8. Project Timeline — SOW Milestone Aligned

8.1 Milestone Schedule

Phase Deliverable Target Date Duration SOW Milestone
Kick
Off

Scope & Team
Walkthrough

04 Feb 2026 — —

Phase
1

Foundation & Architecture 11 Mar 2026 5 weeks M1: Infrastructure

Phase
2

Core Platform
Development

22 Apr 2026 6 weeks M2: Core Platform

Phase
3

AI Engine Development 24 Jun 2026 9 weeks M3: AI Engine

Phase
4

Integration & Workflow 12 Aug 2026 7 weeks M4: Integration

Phase
5

Testing & Go-Live 23 Sep 2026 6 weeks M5: Go-Live

Phase
6

Support & Maintenance Post go-live Ongoing M6: Support

8.2 Phase Details with Deliverables

Milestone 1 — Foundation & Architecture (Weeks 1-5)

HIPAA-compliant cloud infrastructure setup (VPC, encryption, IAM)
Database schema design and deployment (PostgreSQL + pgvector)
Authentication and RBAC framework (OAuth 2.0 + MFA)
API gateway and security layer
CI/CD pipeline configuration
Historical data extraction from Athena Centricity (ETL pipeline)
Data cleaning and filtering (per-payer policy change date filtering)
Policy document collection and RAG corpus preparation

Acceptance: Infrastructure operational, data pipeline running, security audit passed.

Milestone 2 — Core Platform Development (Weeks 6-11)

Web dashboard foundation (React.js + TypeScript)
Mobile app with audio capture (Expo/React Native)
Speech-to-text module integration (Whisper base)
Athena Centricity connector implementation
CureMD API connector implementation
Patient lookup and encounter linking via EMR
Template-based fast-track UI for standard procedures

Acceptance: Audio capture working end-to-end, EMR data retrieval functional, basic transcript generation.

Milestone 3 — AI Engine Development (Weeks 8-17)

LLM fine-tuning on billing data using LoRA (orthopedic/neurosurgery domain)
Medical vocabulary fine-tuning for Whisper STT
Clinical entity extraction pipeline (NLP)
ICD-10 and CPT code mapping engine with confidence scoring
RAG pipeline for claim scrubbing (NCCI, LCD/NCD, payer rules)
Business rules engine with payer-specific optimization
Modifier logic implementation

Acceptance: AI generates claims for test cases with ≥90% code mapping accuracy, STT ≥97% WER on test corpus.

Milestone 4 — Integration & Workflow (Weeks 15-21)

End-to-end pipeline integration (audio → transcript → entities → codes → claim)
Human review interface with side-by-side transcript/claim/EMR view
Claim correction workflow with model feedback loop
Epic FHIR integration (contingent on API access approval)
Audit trail and compliance logging
Notification system for claim events
Claim export mechanism for downstream processing

Acceptance: Full workflow functional, human review working, audit trails complete.

Milestone 5 — Testing & Go-Live (Weeks 19-24)

User acceptance testing with billing staff (real claim scenarios)
Bug fixes and performance optimization
Analytics dashboard and reporting
Documentation and training materials
Go-live preparation and cutover support
Pilot with 1-3 surgeons, then rapid expansion

Acceptance: UAT sign-off, 80% same-day charge capture target met, system stable in production.

Milestone 6 — Support & Maintenance (Post Go-Live)

Ongoing bug fixes and performance monitoring
Model retraining based on correction feedback
Payer rule updates as policies change
System monitoring and infrastructure management

9. Success Metrics

Metric Target Measured At
Same-day charge capture rate 80% M5 Go-Live

Claim denial reduction 10 - 25% reduction 3 months post go-live
A/R cycle improvement 5 - 10 days faster 3 months post go-live

Provider submission time (dictate to approve) < 1 minute M5 UAT

STT accuracy (Word Error Rate) ≥97% on test corpus M3 acceptance
Code mapping accuracy ≥90% on test cases M3 acceptance

10. Risk Assessment

Risk Impact Detail Mitigation
AI Hallucination HIGH^ Incorrect codes generated and
submitted

Mandatory human review for all
claims; confidence thresholds
with escalation

Epic API Delay HIGH^ FHIR approval can take 3- 6
months; blocks hospital
integration

Begin Epic approval process
immediately; build with mock
data; Epic integration is M4
deliverable allowing parallel work

STT Accuracy HIGH^ 99% target unrealistic; industry
standard is 95-98%

Revised to ≥97% WER; medical
vocabulary fine-tuning; noise
reduction; iterative improvement
via feedback loop

Policy Staleness MED^ Outdated rules cause denials Manual updates by billing staff in
MVP; per-payer relevance
filtering; future auto-learning

Data Quality MED^ Historical data has gaps or
errors

Data validation phase in M1; filter
obsolete records per-payer; SME
review of training data
Adoption Friction MED^ Users resist workflow changes Smartphone-only capture (no
hardware); template fast-track for
routine procedures; < 1 min
submission target
Security Breach HIGH^ PHI exposure Self-hosted LLM within VPC;
AES-256 encryption; HIPAA
controls; no external AI API calls

Integration Delays MED^ EMR API access issues
(especially Athena legacy)

Existing connectors for
Athena/CureMD; confirm
Centricity version; early API
testing in M1

Scope Creep MED^ Requirements grow beyond
agreed MVP

Strict scope freeze after M1;
change requests require written
approval with cost/timeline
impact

11. Open Questions Requiring Resolution

The following items must be resolved before or during Phase 1. Items marked CRITICAL block development if unresolved.

11.1 Development-Critical Questions

# Question Priority Owner

Q1 What are the 5-10 common spine surgery types for
MVP with their typical CPT + ICD-10 code
combinations?

CRITICAL (^) Dr. McHugh / Billing team Q2 What is the Athena Centricity version? What APIs are available? (Emily flagged potential legacy roadblocks) CRITICAL (^) Emily / Tech team Q3 Is Epic FHIR API access already granted or is approval still pending? What is the expected timeline? CRITICAL (^) Emily Q4 What is the exact claim handoff mechanism to Athena? (API push, file import, manual re-entry?) CRITICAL (^) Emily Q5 What is the timeline for receiving historical billing data and payer policy documents? CRITICAL (^) Emily Q6 What is the prioritized list of top 30 insurance plans? (Start with top 10 for initial validation) HIGH (^) Emily / Billing team Q7 What patient identifier links audio to EMR? (MRN, encounter ID, manual entry, or EMR lookup?) HIGH (^) Dr. McHugh / Emily Q8 Mobile platforms: iOS only, Android only, or both required for MVP? HIGH (^) Dr. McHugh Q9 What specific data fields should be pulled from Epic vs Athena vs CureMD for a given claim? HIGH (^) Emily / Billing team Q10 Expected daily volume: how many audio recordings per day, average length per recording? HIGH (^) Dr. McHugh Q11 Total concurrent users (providers + billers + admin)? HIGH (^) Emily Q12 What format should approved claims be exported in? (PDF CMS-1500, data file, API push to Athena) HIGH (^) Emily Q13 Cloud provider preference: AWS vs GCP vs Azure? (GCP recommended in cost analysis) MEDIUM (^) Emily Q14 Has infrastructure budget (~$1,100-1,600/month) been confirmed? MEDIUM (^) Emily Q15 What notification methods are required? (Email, SMS, in-app push, or combination) MEDIUM (^) Emily

Note: CRITICAL questions must be resolved before Phase 1 can be completed. HIGH questions must be resolved before Phase 2 begins.

12. Stakeholders and Team Structure

12.1 Client Stakeholders

Role Name Contact
Executive Sponsor Dr. Brian McHugh DrMcHugh@mchughneurosurgery.com

Project Manager Emily Clifford emily@farmtotablehealth.com / +1 (716) 983- 2572

12.2 Implementation Partner (Dextra Labs)

Role Name Email
Engagement
Partner

Vijay Agarwal vijay@dextralabs.com

Project Manager Gaurang Ghadigaonkar gaurang.ghadigaonkar@dextralabs.com
Solution Architect Yasha Khandelwal yasha@dextralabs.com

Technical Lead TBD —

12.3 Development Partner (Tech4Biz Solutions Pvt Ltd)

Tech4Biz Solutions serves as the development partner, providing technical execution and existing EMR connector expertise (Athena Centricity, CureMD).

13. Document Approval

This Software Requirements Specification requires approval from the following stakeholders before development proceeds:

Role Name Signature Date

Executive Sponsor Dr. Brian McHugh

Client PM Emily Clifford (^) Engagement Partner Vijay Agarwal (Dextra) (^) Solution Architect Yasha Khandelwal (Dextra) (^)

Document Version History:

Version Date Author Changes

1.0 03 Feb 2026 Yasha Initial draft based on discovery documents
1.1 03 Feb 2026 Yasha Added open questions (Sec 10), clarified system
boundary (Sec 1.3)
2.0 16 Feb 2026 Yasha Major revision: SOW milestone alignment, template fast-
track workflow, RAG claim scrubbing, OON billing
context, scope narrowing (5-10 spine procedures, 5
providers, CMS-1500 only), STT accuracy revised to 97%
WER, ModMed/UB-04 formally excluded, confidence
thresholds, notification system, success metrics, MDM
level generation, medical necessity justification, claim
state machine, data migration ETL pipeline, model
retraining pipeline, all requirements mapped to
milestones. Removed all references to specific AI model
names — SRS is now technology-agnostic on AI model
selection.

— End of Document —

43 KiB Raw Blame History Unescape Escape