CRC Data Mining and Predictive Analytics 2 — Questions and Answers
Question 1: An MA plan uses predictive modeling to identify patients with suspected undocumented conditions. Which data source combination is most effective?
- Claims data only
- Claims combined with pharmacy data to identify conditions suggested by medication patterns (Correct answer)
- EHR data only
- Survey data combined with demographics
Correct answer: Claims combined with pharmacy data to identify conditions suggested by medication patterns
Pharmacy data reveals conditions not captured in diagnosis codes, such as a patient on metformin without a diabetes diagnosis.
Medications are prescribed for specific conditions, creating medication-condition logic that can be algorithmically applied to detect gaps.
Question 2: A provider group's mental health HCC capture rate increased 300% year-over-year. What analytical approach should be used?
- Immediately flag as fraud
- Compare against peer trends, evaluate for staffing changes or new screening programs (Correct answer)
- Accept the increase if patient volume increased
- Discount as normal variation
Correct answer: Compare against peer trends, evaluate for staffing changes or new screening programs
A 300% increase requires investigation through peer benchmarking, temporal analysis, and operational review before drawing conclusions.
The analytical approach should determine whether the increase reflects hiring a behavioral health specialist, new screening protocols, or problematic coding.
Question 3: Which technique is most appropriate for estimating future healthcare costs using historical risk adjustment data?
- Simple linear regression using RAF scores
- Gradient boosted trees incorporating HCCs, demographics, utilization, and pharmacy data (Correct answer)
- K-means clustering by cost similarity
- Logistic regression for high-cost vs low-cost prediction
Correct answer: Gradient boosted trees incorporating HCCs, demographics, utilization, and pharmacy data
Gradient boosted trees capture non-linear relationships and interactions that simple models miss.
Healthcare cost prediction involves non-linear relationships. Gradient boosted trees handle mixed data types, capture interactions naturally, and are robust to outliers.
Question 4: Data mining reveals 8% of HCC revenue comes from diagnoses captured only during annual wellness visits. What risk does this represent?
- No risk, as wellness visits are valid encounter types
- Moderate RADV audit risk if documentation lacks clinical depth (Correct answer)
- High fraud risk requiring immediate OIG reporting
- Administrative risk only
Correct answer: Moderate RADV audit risk if documentation lacks clinical depth
While annual wellness visits are valid, conditions captured only once with limited documentation face audit risk.
If providers are carrying forward diagnoses from problem lists without clinical assessment, the documentation may not meet RADV standards.
Question 5: Which feature would be the strongest predictor for a model predicting HCC gaps?
- Member age and gender
- Number of unique providers seen
- HCCs captured in prior year not yet appearing in current year claims (Correct answer)
- Total medical expenditure
Correct answer: HCCs captured in prior year not yet appearing in current year claims
Prior year HCCs not yet recaptured is the strongest predictor because chronic conditions documented last year are likely still present.
Chronic conditions rarely resolve, so prior year captures predict current year conditions with high accuracy. The annual recapture requirement makes this the most direct gap indicator.
Question 6: Data mining reveals providers consistently coding cerebrovascular sequelae without prior stroke documentation. What technique identified this?
- Supervised classification
- Association rule mining
- Anomaly detection identifying deviations from peer norms (Correct answer)
- Time series analysis
Correct answer: Anomaly detection identifying deviations from peer norms
Anomaly detection identifies providers whose coding patterns significantly deviate from expected norms.
Unlike supervised classification which requires labeled training data, anomaly detection can identify new, previously unknown problematic patterns by flagging deviations from baseline behavior.
CRC Data Mining and Predictive Analytics 2
An MA plan uses predictive modeling to identify patients with suspected undocumented conditions.
Which data source combination is most effective?