Claims Anomaly Detection

Overview

Anomalies in healthcare data come in two flavors that this skill handles together: data errors (a $90,000 office visit, an impossible date, a unit count of 9999) and behavioral outliers (a provider billing a procedure 50× the peer rate). The skill provides a layered detector robust statistics first, unsupervised ML second, optional supervised scoring third with explainable reasons for every flag so an analyst can act on it.

When to Use This Skill

• Catching erroneous records before they are paid or fed to a model.
• Surfacing outlier providers, members, codes, or claim lines for review.
• Building an anomaly score as a feature or a quarantine gate in a pipeline.
• Monitoring utilization/cost time series for sudden shifts.
• Early-stage fraud-waste-abuse (FWA) triage (signal, not adjudication).

Detection Layers

1. Robust univariate z-score, IQR fences, and MAD (median absolute deviation,

resistant to the very outliers we seek) on amounts, units, lengths-of-stay.

1. Rule + peer-group compare each record to its peer cohort (same specialty,

procedure, region) using percentile bands; flag extreme deviation.

1. Unsupervised multivariate Isolation Forest, Local Outlier Factor, and

autoencoder reconstruction error over engineered claim features.

1. Time series STL/seasonal decomposition residuals or Prophet intervals on

per-provider or per-DRG cost/volume trends.

1. Supervised (optional) if labeled audit outcomes exist, train a gradient-boosted

classifier and calibrate; otherwise stay unsupervised.

Example

from sklearn.ensemble import IsolationForest
import numpy as np, pandas as pd

feats = ["billed_amount","units","los_days","amount_per_unit","provider_zscore"]
X = claims[feats].fillna(0).values

iso = IsolationForest(n_estimators=300, contamination=0.01, random_state=42).fit(X)
claims["anomaly_score"] = -iso.score_samples(X)        # higher = more anomalous
claims["is_anomaly"]    = iso.predict(X) == -1

# Robust univariate cross-check (MAD)
def mad_z(s):
    med = s.median(); mad = (s - med).abs().median() or 1e-9
    return 0.6745 * (s - med) / mad
claims["amount_mad_z"] = mad_z(claims["billed_amount"])
flagged = claims[(claims.is_anomaly) | (claims.amount_mad_z.abs() > 5)]

# Per-record explanation: which features pushed the score up
import shap
expl = shap.TreeExplainer(iso)
top_reasons = expl.shap_values(X[flagged.index])   # hand to explainable-ml-healthcare

Evaluation & Tuning

Pure unsupervised detection has no ground truth, so tune contamination to a review- capacity budget (e.g., top 1%), and measure precision@k against analyst dispositions to build a feedback loop. Watch for concept drift and re-fit on a rolling window. Always attach a human-readable reason (which feature, how far from peer) an unexplained flag is an unusable flag.

Outputs

• anomaly_scores.parquet record id, score, binary flag, top contributing features.
• provider_outliers.csv peer-group ranked outliers for FWA triage.
• timeseries_alerts.json points breaching seasonal expectation bands.
• anomaly_report.md method, thresholds, flag rate, example explanations.

Healthcare Context

Designed for 837/claim-line schemas (billed/allowed/paid amounts, units, POS, DRG, revenue codes). Complements healthcare-data-quality-profiling (rule-based validity) by catching the statistical anomalies rules miss. FWA outputs are signals for review, not accusations or adjudication.

References

• scikit-learn outlier detection https://scikit-learn.org/stable/modules/outlier_detection.html
• PyOD (40+ detectors) https://pyod.readthedocs.io
• CMS / NHCAA fraud-waste-abuse guidance.