Infrastructure Health
Clinical Coding Pipeline
Document Processing & Code Assignment
Prior Authorization Pipeline
PA Generation, Criteria Matching & Clearinghouse Filing
ML Outcome Feedback Loop
ERA/835 Capture → Label Accumulation → Adaptive Retraining → Auto-Deploy
Model Training Status
Phase
—
0 labeled outcomes
/ 500 required
Last trained
Never
Schedule
Daily check → monthly cadence
PA Submission Statistics
Total submissions—
Approved—
Denied—
Overturned on appeal—
Awaiting ERA decision—
With high-risk clinical flags—
Auto-Escalation Status
● L1 — Re-queried (0–24h past deadline)
—
▲ L2 — Escalation Fax (24–72h past deadline)
—
✕ L3 — Peer-to-Peer (72h+ past deadline)
—
Exception queue depth—
Peer-to-peer queue depth—
Self-Learning
Verified cases in learning index
—
PA Appeal & Outreach
PA appeals drafted (AI)—
PA appeals pending provider signature—
Patient outreach sent (email + SMS)—
EHR Write-Back
Coding results written to EHR (ICD+CPT)—
Write-back failed / pending—
PA outcomes pushed to EHR encounter—
Appeal letters attached to EHR encounter—
Claim Denials & Appeals
Total claim denials received—
Claim appeal letters drafted (AI)—
Overturned on claim appeal—
+ Enter claim denial manually
Open form ›
ML Confidence — Payer/Procedure Combos
● ML High (≥500 samples)
—
◐ ML Low (1–499 samples)
—
○ LCD Fallback (no data yet)
—
Recent Pipeline Log
No log entries.
Policy Change Alerts
CMS LCD/NCD + payer policies — weekly feed
Open alerts
—
Critical (pending submissions affected)
—
Warning (policy changed, no active PA)
—
Claim Denials & Appeals
Total received—
Appeal letters drafted (AI)—
Overturned on appeal—
Upheld / closed—
Upload Doctor's Note
PDF processed on EC2 — results saved automatically
No file selected
Processing Queue
Queue empty — upload a note to begin
Self-Learning — Verified Coding Cases
Human-approved cases feed future coding as few-shot examples
Verified cases in Qdrant index—
How it works: when you approve a coding result, it is embedded and stored in the
verified_cases Qdrant collection. On the next note, the system retrieves the top-3 most similar past cases and injects them as few-shot examples into the ICD, CPT, and HCPCS coding prompts — so the model learns your practice's specific coding patterns over time.