Medical Fraud Data Investigation

1. Data Processing

We were dealing with a million plus rows and 84 columns, so I decided to eliminate some incomplete data and convert some columns to numeric types.

# Convert key columns, turning errors into 'NaN' (Not a Number)
numeric_cols = ['tot_clms', 'tot_drug_cst', 'tot_benes', ...]
for col in numeric_cols:
    df[col] = pd.to_numeric(df[col], errors='coerce')

# Remove rows with missing essential data
df.dropna(subset=numeric_cols, inplace=True)

After processing the data we got around 414,000 providers with complete columns ready for evaluation.

2. Feature Engineering

Then I decided to engineer a couple more datapoints upon which I would use the isolation forest algorithm and try to find outliers based on these newly created metrics.

# Creating primary investigative metrics
df['claims_per_beneficiary'] = df['tot_clms'] / df['tot_benes']
df['cost_per_claim'] = df['tot_drug_cst'] / df['tot_clms']
df['opioid_claim_rate'] = df['opioid_tot_clms'] / df['tot_clms']
df['brnd_tot_clms'] = df['brnd_tot_clms'].fillna(0) 
df['brand_claim_rate'] = df['brnd_tot_clms'] / df['tot_clms']

3. Anomaly Detection Results

After running the Isolation forest algorithm on these metrics we flagged over 12k anomalies, who were clearly different from their peers:

Since the average patient risk score was higher in the anomaly group we are assured that these anomalies are not just fraud but most surely are just prescribers who work in fields which prescribe expensive drugs. This was proven correct in my next checks.

4. False Positives

Looking at the top anomalies at the cost per claim, I checked the prescribers with the highest costs per claims and the lowest risk scores, which is where I thought I found my smoking gun.

Ngoc Trieu and Ada Noh were optometrists with patient risk scores around 1.00 but had insanely high costs per claim over $100k and over $30k per claim respectively. However when I cross referenced their prescriptions in this database:

https://data.cms.gov/provider-summary-by-type-of-service/medicare-part-d-prescribers/medicare-part-d-prescribers-by-provider-and-drug/data

I found that they were prescribing Oxervate, a legitimately expensive and rare drug. There goes my smoking gun.