The Denial Prediction Opportunity

High claim denial rates are one of the largest sources of revenue leakage in ambulatory care. Most denials happen for predictable, mechanical reasons — wrong modifier, missing referral, eligibility lapse, NPI mismatch. The same payer denies the same procedure code patterns over and over again.

This article describes the architecture pattern for a denial prediction system that scores outbound claims before submission and surfaces correction suggestions inside the billing workflow. It's the design I apply when this problem comes up in client work.

The Training Data: 835 Remittance History

The training corpus is the practice's historical 835 remittance advice files — the structured EDI responses payers send back with every claim adjudication. Each 835 record contains:

Claim ID + procedure code(s)
Payer ID
Modifier combinations
Provider NPI
Adjudication status (paid / denied / pending)
Denial reason codes (CARC) when applicable

Two years of 835 history typically yields 50,000–500,000 labeled examples for a mid-sized group practice — enough to train a useful model.

The Model: XGBoost, Not a Neural Network

Denial prediction is a structured-tabular problem with mostly categorical features (codes, IDs, modifiers). Gradient-boosted trees consistently outperform neural networks on this shape of data, train faster, and are easier to interpret — which matters because billing staff need to trust the scores.

import xgboost as xgb
from sklearn.model_selection import train_test_split

features = ['cpt_code', 'payer_id', 'modifier_combo', 'provider_npi',
            'claim_amount_bucket', 'days_since_eligibility_check']
X = df[features]
y = (df['adjudication_status'] == 'denied').astype(int)

X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, stratify=y
)

model = xgb.XGBClassifier(
    n_estimators=300, max_depth=6, learning_rate=0.1, eval_metric='auc'
)
model.fit(X_train, y_train, eval_set=[(X_test, y_test)], verbose=False)

A baseline XGBoost model on 200K examples typically reaches AUC 0.78–0.85 on denial prediction. Not magic — but enough to surface high-risk claims for human review before submission.

Integration: Pre-Submission Scoring API

The model runs as a stateless prediction service. The billing system POSTs an outbound claim payload, the service returns a denial probability and the top contributing features. If denial_probability > 0.6, the billing system flags the claim and shows the top risk factors plus any rule-based correction suggestion. The biller decides whether to submit, edit, or hold.

Continuous Retraining

The model retrains nightly on the prior day's 835 ingestions. New denial patterns (new payer policies, new code edits) get incorporated within 24 hours rather than waiting for a quarterly retraining cycle.

HIPAA Considerations

Claim scoring touches PHI: patient identifiers, procedure codes, diagnosis codes. The prediction service runs in the same HIPAA-compliant cloud environment as the billing system, with KMS-encrypted feature stores and audit logging on all predictions.

For the broader RCM context, see the revenue cycle management software guide and the RCM analytics dashboard architecture. For when ML makes sense vs RPA bots in billing, see the healthcare RPA ROI breakdown.

The Custom Practice Management service builds ML-powered denial prediction integrated with charge capture and claim submission workflows.