Claims and Billing Anomaly Detection for a 30-Office Dental Group's Revenue Cycle
Let's Connect
Overview
What we built
A 30-office dental group was finding out about billing problems only after payers rejected the claims. We built an anomaly detection layer that catches suspect claims before they leave the building.
In plain terms: every dental visit becomes an insurance claim, and a claim with a coding error, a missing attachment or a payer-specific quirk comes back denied. The group's central billing team could hand-check only a small sample of the claims flowing out of its 30 offices, so most problems surfaced weeks later as remittances, the same mistakes repeated at the same offices for months, and write-offs climbed quarter over quarter.
We built a system that scores every outgoing claim against patterns learned from the group's own practice management and clearinghouse history. Claims that look wrong are held before submission in a work queue, with the reason spelled out in plain language, and dashboards show each office's denial and payment trends. First-pass claim acceptance rose from 84% to 93% within 6 months, denial-related write-offs fell 28% year on year, and average days in accounts receivable dropped from 41 to 29.
The Problem
Revenue leaking through denials
Revenue cycle problems in a dental group rarely announce themselves. A claim goes out with a wrong procedure code or a missing attachment, weeks pass, and a remittance arrives paying less than expected or nothing at all. Across 30 offices in the US Mid-Atlantic, that quiet leak had become structural: coding errors, missing attachments and payer-specific denial patterns were draining revenue, and each was discovered only after the money had already been delayed or lost.
The central billing team fought back the only way manual review allows: sampling. A small fraction of claims got human eyes before submission; the rest went out on trust. Recurring problems at individual offices, one provider's coding habit, one front desk's missing documentation step, could run unnoticed for months before the audit cycle caught them, and write-offs rose quarter over quarter as the backlog of undetected patterns grew.
The lag was the real enemy. With remittances arriving well after treatment, the group was always diagnosing last quarter's mistakes, and the audit rhythm meant an office could repeat the same denial-triggering error for months before anyone connected the pattern. Nobody was losing one big amount; the group was losing small amounts everywhere, constantly.
Sampling instead of coverage
Manual audits could touch only a small share of outgoing claims, leaving the majority unreviewed and recurring problems free to repeat unchecked.
Feedback weeks late
Errors surfaced only when remittances came back, so the billing team was always reacting to denials rather than preventing them at source.
Repeating office patterns
Recurring coding and documentation errors at individual offices ran unnoticed for months, with each repetition adding more denials and rework to the pile.
Write-offs compounding quarterly
Revenue that could not be recovered was written off, and the write-off line rose quarter over quarter as undetected patterns accumulated.
What it was costing them
Every denied claim cost the group twice: once in delayed or lost revenue, and again in the rework of correcting and resubmitting it. Write-offs rose quarter over quarter, cash sat in accounts receivable for 41 days on average, and the billing team's capacity went into chasing remittances instead of fixing the upstream habits that generated the denials in the first place.
The Solution
Claim scoring before submission
We built the anomaly detection layer on top of the systems the group already ran, drawing claims and outcomes from its practice management and clearinghouse data rather than asking offices to change how they work. The models learn what normal looks like for each combination of procedure code, payer, provider and office, then score every outgoing claim against those learned patterns.
Claims that score as suspect do not simply disappear into a report. They are held in a pre-submission work queue where each hold carries a plain-language reason a biller can act on, so the queue reads like a checklist rather than a model output. Clean claims pass through untouched, which keeps the workflow fast and concentrates the team's attention on the claims most likely to bounce.
Detection at claim level was only half the design. Dashboards aggregate the same signals upward, surfacing office-level drift in denials, adjustments and days in accounts receivable, so a developing pattern at one office becomes visible within days rather than emerging months later in an audit, while it can still be corrected cheaply.
Key decisions
Score every claim
Anomaly scoring runs on all outgoing claims, not an audit sample, so no office or payer combination sits outside the safety net.
Hold before submission
Suspect claims are intercepted in a pre-submission work queue, converting problems from post-remittance surprises into fixes made before the payer ever sees them.
Explain holds in plain language
Every held claim states why it looks wrong in words a biller can act on, so the queue drives corrections rather than confusion.
Learn per payer and office
Patterns are modelled by procedure code, payer, provider and office, because a claim that is normal for one payer can be a guaranteed denial for another.
Watch drift, not just claims
Office-level dashboards track denials, adjustments and receivables over time, catching slow behavioural drift that no single claim score would reveal on its own.
Measurable Impact
What changed after launch
The submission pipeline got measurably cleaner: first-pass claim acceptance rose from 84% to 93% across all 30 offices within 6 months, and denial-related write-offs fell 28% year on year as anomaly holds caught problems before payers could. Cash followed the cleaner claims, with average days in accounts receivable dropping from 41 to 29 across the group.
The deeper change is in tempo. Recurring office-level coding errors that once ran inside a months-long audit lag are now flagged within days, while the habit is still fresh and fixable. The billing team works a prioritised queue with reasons attached instead of sampling blind, and office conversations about billing quality now start from a shared dashboard rather than a quarterly surprise.
Claim acceptance
First-pass acceptance running at 84%
93% acceptance across all 30 offices
Error detection
Recurring office errors unnoticed for months
Office-level patterns flagged within days
Revenue leakage
Write-offs rising quarter over quarter
Denial-related write-offs down 28% year on year
Cash cycle
Average 41 days in accounts receivable
Down to 29 days across the group
Headline results
First-pass claim acceptance rose from 84% to 93% across all 30 offices within 6 months
Denial-related write-offs reduced by 28% year on year after pre-submission anomaly holds
Recurring office-level coding errors flagged within days versus the previous 2-3 month audit lag
Average days in accounts receivable cut from 41 to 29 across the group
Tech & Tools Used
What powered the build
Every tool below earned its place in this engagement. Here is the part each one played.
Python (FastAPI)
FastAPI services expose claim scoring to the billing workflow, returning holds and plain-language reasons quickly enough to sit inside the submission path.
scikit-learn
Underpins the anomaly detection pipeline, from feature preparation on claim history to the baseline models that define normal per payer and office.
XGBoost
Gradient-boosted models score outgoing claims for denial risk, learning the interactions between procedure code, payer, provider and office that simple rules miss.
PostgreSQL
Stores claims, scores, hold decisions and their eventual outcomes, giving the group a queryable history of every intervention the system has made.
Apache Airflow
Orchestrates the recurring pipelines that ingest practice management and clearinghouse data, retrain the models and refresh the office-level drift metrics.
Metabase
Serves the dashboards where billing leads track denials, adjustments and days in accounts receivable per office and spot drift early.
Redis
Caches scoring features and work-queue state, keeping pre-submission checks fast so the anomaly layer never becomes a bottleneck in daily billing.
Docker
Packages the scoring services, pipelines and dashboard components into consistent containers, so the same build runs identically in development and production.
AWS ECS
Runs the containerised services in production, scaling the scoring API and scheduled pipelines without the group managing its own servers.
Ready to Build your Dental Healthcare Business with Fraud, Risk & Anomaly Detection
Ask Byte
Ask Byte
Typically replies instantly
just Now
Hi! I'm OrganByte's assistant. How can I help you today?
AI-generated content may be incorrect
