How We Score Prescribing Risk

OpenPrescriber uses a unified 10-component risk scoring model that combines two complementary approaches: specialty-adjusted peer comparison (how far does a provider deviate from others in their specialty?) and population-level analysis (where do they fall nationally?). This dual approach avoids the critical flaw of universal thresholds and produces more accurate results than either method alone.

Why Specialty-Adjusted Scoring?

Most fraud detection tools use universal thresholds — for example, flagging anyone with an opioid prescribing rate above 20%. This approach generates massive false positives. A pain management specialist at 50% opioid prescribing is entirely normal for their field. A family doctor at 50% is 13 standard deviations above their specialty average — a genuinely extreme outlier.

Our model computes baseline statistics (mean, standard deviation, percentile distributions) for 110 medical specialties. Each provider is then scored against their own specialty peers. This means a pain management doctor is compared to other pain management doctors, an oncologist to other oncologists, and a family physician to other family physicians.

The technical implementation uses z-scores — the number of standard deviations a provider's metric falls above or below their specialty mean. A z-score of 2 means the provider is roughly in the top 2-3% of their specialty. A z-score of 5 means they're far beyond what any normal clinical variation can explain.

The 10 Scoring Components

Each component contributes points to a composite score (max 100). Points are additive, with a volume multiplier applied at the end.

1. Specialty-Adjusted Opioid Rate (0–25 points)

The most important component. We compute each provider's opioid prescribing z-score against their specialty. A family doctor with 3.78% mean opioid rate (σ=10.2) gets flagged very differently than a pain management specialist with a 35.3% mean (σ=22.1).

• >5σ above peers: 25 points (extreme — virtually impossible by chance)
• >3σ above peers: 18 points (very high — top ~0.1% of specialty)
• >2σ above peers: 10 points (high — top ~2% of specialty)

2. Population Percentile Opioid Rate (0–15 points)

Independent of specialty, we check where the provider falls in the national distribution of all 433,324 providers with opioid claims. This catches cases where a specialty itself has unusual norms.

• 99th percentile (rate >70.6%): 15 points
• 95th percentile (>50.3%): 10 points
• 90th percentile (>37.2%): 5 points

3. Cost Outlier (0–10 points)

Requires a dual condition: the provider must be high both in population percentile AND relative to their specialty peers (z-score). This prevents flagging oncologists whose high costs are normal for their field.

4. Brand-Name Preference (0–8 points)

Flags providers prescribing far more brand-name drugs than their specialty peers. Only triggered when z-score exceeds 2–3 AND the absolute brand percentage exceeds 30–50%. This avoids flagging specialties where brand-name drugs are clinically necessary.

5. Long-Acting Opioid Rate (0–8 points)

Long-acting opioids (OxyContin, MS Contin, fentanyl patches) carry higher diversion and abuse potential. We flag providers whose long-acting opioid share exceeds 3 standard deviations above their specialty mean.

6. Elderly Antipsychotic Prescribing (0–10 points)

CMS specifically tracks antipsychotic prescribing to patients 65+ as a quality concern — these drugs carry FDA Black Box Warnings for increased mortality in elderly dementia patients.

7. Opioid + Benzodiazepine Co-Prescribing (0–8 points)

By analyzing the full 11.9 million-row provider-drug dataset, we identify providers who prescribe both opioids and benzodiazepines. The FDA issued a Black Box Warning about this combination due to life-threatening respiratory depression. We identified 6,149 co-prescribers.

8. OIG Exclusion Match (0–20 points)

We cross-reference every provider NPI against the Office of Inspector General's List of Excluded Individuals/Entities (LEIE) — individuals convicted of healthcare fraud, patient abuse, or related offenses. We found 372 excluded providers still actively prescribing in Medicare Part D.

9. Low Drug Diversity (0–5 points)

Providers who prescribe a very narrow range of drugs (≤5–10 unique medications) while also prescribing opioids may indicate a “pill mill” operation rather than a genuine medical practice. Legitimate practices typically prescribe across a broad range of drug categories.

10. High Fills Per Patient (0–5 points)

Providers whose patients average more than 15–20 prescription fills per year may indicate over-prescribing or patients being used as conduits for drug diversion. This metric is specialty-adjusted to account for fields like psychiatry where multiple maintenance medications are common.

Volume Adjustments

Two volume-related adjustments improve accuracy:

• Minimum threshold: Providers with fewer than 100 claims are excluded entirely — too little data for meaningful statistical analysis.
• Low-volume cap: Providers with 100–200 claims have their percentile-based flags capped at 60% — acknowledging higher statistical noise.
• High-volume multiplier: Providers with >5,000 claims receive a 15% score boost, because the same risky patterns at scale affect vastly more patients. Providers with >2,000 claims get 5%.

Risk Levels

What This Looks Like in Practice

Consider our highest-scoring provider (77/100): an internal medicine doctor who:

• Prescribes opioids at 62% of claims — 2,687% above the internal medicine average of 2.2%
• Co-prescribes opioids and benzodiazepines (FDA Black Box Warning combination)
• Appears on the OIG exclusion list (convicted of healthcare offense)
• Prescribes only 6 unique drugs (extremely low diversity)
• Has elevated long-acting opioid prescribing vs peers

Every single component is transparent and explainable. You can see exactly why they scored 77 — 25 points for extreme opioid vs peers, 10 for 95th percentile nationally, 20 for LEIE exclusion, 8 for opioid+benzo combo, 8 for LA opioid, 3 for elevated cost, and 3 for low diversity.

Explore flagged providers yourself on the Risk Explorer, or see the full flagged providers list.

Key Limitations

• No clinical context: We cannot see diagnoses, treatment plans, or medical necessity. High opioid rates may be entirely appropriate for hospice, palliative care, or addiction treatment.
• Specialty misclassification: If CMS lists a provider under the wrong specialty, the peer comparison may be inaccurate.
• Single-year snapshot: 2023 data only — providers may have changed practices since then.
• Right-skewed distributions: Some specialties (like NPs, with mean 3.8% and σ=10.2%) have highly skewed distributions where the standard deviation is larger than the mean. Z-scores work less well for heavily skewed data.
• Correlation ≠ fraud: Statistical outliers are not inherently problematic. Our scores identify unusual patterns, not bad actors.

Machine Learning Extension

Beyond the rule-based model, OpenPrescriber now includes a machine learning fraud detection system trained on 281 confirmed fraud cases from the OIG LEIE exclusion list. The ML model — a Bagged Decision Trees ensemble — analyzes the same 20 features and identifies non-obvious pattern combinations that hand-tuned rules miss.

The ML model flagged 4,100+ providers at ≥80% confidence. Cross-validated performance: 83% precision, 66.6% recall. See the ML Fraud Detection page for full results and the complete list of flagged providers.