Data Project. Overview. Home Health Overview Fraud Indicators Decision Trees. Zone Program Integrity Contractor Zone 4 Decision Tree Modeling

Transcription

1 Zone Program Integrity Contractor Zone 4 Decision Tree Modeling Holly Pu, M.S. Chief Statistician October 14, 2009 Data Project Home Health Overview Fraud Indicators Decision Trees Overview 1

2 Home Health Rules and Regulations 42 CFR (Code of Federal Regulations) Sections Home Health Services Section Requirements for Home Health Services Section 484 Home Health Conditions of Participation Home Health Rules and Regulations Publications 100 XX Publication , Medicare General Information, Eligibility and Entitlement Manual, Chapter 1, Section 10.2, Home Health Services and Chapter 4, Section 30, Physician Certification Publication , Medicare Benefit Policy Manual, Chapter 7, Home Health Services Publication , Medicare Claims Processing Manual, Chapter 10, Home Health Agency Billing Publication , Program Integrity Manual (PIM) Home Health Eligibility To qualify for Medicare coverage of home health services, the patient must meet each of the following: Services billed from a Medicare participating Home Health Agency Be confined to the home (homebound) Be under the care of a physician Be in need of skilled services Be under a Plan of Care 2

3 State County Summary Provider Number of Number of Home Number of State Provider County Total Payment Patients Health Agencies Claims HARRIS $1,073,334,215 93, ,683 DALLAS $1,011,769,175 75, ,052 HIDALGO $866,242,700 54, ,491 TARRANT $483,479,760 50, ,154 BEXAR $444,553,066 49, ,302 LAHOMA $256,388,647 31, ,466 TULSA $167,193,549 22, ,786 PUSHMATAHA $81,676,418 6, ,216 DENVER $76,212,907 17, ,643 ARAPAHOE $71,196,049 15, ,366 EL PASO $29,105,530 6, ,106 NM BERNALILLO $63,050,653 11, ,879 NM DONA ANA $38,260,750 5, ,007 NM SAN JUAN $29,497,595 3, ,116 Growth Trend Trend by Payment Amount $250,000,000 $200,000,000 $150,000,000 $100,000,000 $50,000,000 $0 Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Payment NM Payment Payment Payment 3

4 Fraud Indicators Extensive Home Care Episodes Background: Home health care is usually temporary. The goal of home health care is rehabilitation. If a home health agency has a high percentage of beneficiaries receiving long term home care services, it may indicate fraud/abuse. Time Frame: April, 2006 March 31, 2009 Findings: In three years, Medicare paid around $4.2 billion for beneficiaries with five or more consecutive home care episodes. Out of the $4.2 billion, over $1.1 billion were paid for beneficiaries with twelve or more consecutive home care episodes. An initial threshold of 40 percent was established by the home health subject matter expert to flag aberrant providers. Consecutive Home Care Episodes Summary by State Provider Number of Home State Number of Home Health Episodes Payment Number of Patients Health Agencies (More than 5 Episodes, Less than 12 Episodes) $16,737,291 1, More than 12 Episodes $4,579, Subtotal $21,316,923 (More than 5 Episodes, Less than 12 Episodes) $38,965,259 2, NM More than 12 Episodes $11,890, Subtotal $50,855,557 (More than 5 Episodes, Less than 12 Episodes) $372,374,611 23, More than 12 Episodes $168,912,079 5,796 Subtotal $541,286,690 (More than 5 Episodes, Less than 12 Episodes) $2,564,022, ,193 1,565 More than 12 Episodes $985,801,179 28,029 Subtotal $3,549,824,046 Grand Total $4,163,283,216 4

5 Outlier Payment Claim Background: Home Health outlier payment is identified by the value code 17. Additional PPS reimbursement is based on the number of visits, often in the $10,000 range. PPS assumes the cost of care exceeds the threshold dollar amount resulting in an outlier payment. On July 30, 2009 CMS proposes to cap outlier payments at 10 percent per agency and target total aggregate outlier payments at 2.5 percent of total HH PPS payments for calendar year An initial threshold of 50 percent was established by the home health subject matter expert to flag aberrant providers. Home Health Outlier Summary by County Provider State Provider County Outlier Payment Outlier Patients Outlier Claims Total Payment Total Patients Total Claims Ratio Claims (Outlier Claims / Total Claims) Ratio Pats (Outlier Patients / Total Patients) Ratio Payment (Outlier Payment / Total Payment) DALLAS $248,218,141 5,945 31,672 $996,815,929 74, , HIDALGO $185,583,072 6,395 33,311 $865,246,001 54, , BEXAR $115,709,588 4,052 18,585 $443,424,195 49, , TARRANT $76,372,653 2,749 10,706 $481,189,362 50, , HARRIS $55,210,502 2,631 7,693 $798,415,325 78, , LAHOMA $28,603,974 1,619 4,527 $255,233,556 30, , TULSA $12,322, ,007 $167,094,612 22,447 68, DENVER $4,286, $75,635,586 17,272 28, ARAPAHOE $2,039, $71,182,890 15,887 25, NM DONA ANA $1,405, $38,260,750 5,816 15, NM EDDY $854, $13,865,001 2,014 6, Home Health Outlier Summary by State Provider State Number of Home Health Agencies (with 50% or more Outlier Payment Claims) 6 4 Outlier Payment $1,008,391 $9,478,029 Total Payment (for the Home Health Agencies with 50% or more Outlier Payment Claims) $1,711,589 $17,784,424 Findings: There were over 100 home health agencies that had 50 percent or more outlier payment claims. Out of these providers the majority concentrates in the State of Texas. The total outlier payment made in ZPIC Zone 4 states was over $270 million in a three year time period. 117 $263,880,584 $410,093,209 Grand Total 127 $274,367,004 $429,589,221 5

6 Single Referring Part B Physician Background: This indicator is to look for home health referring providers who referred the majority of the patients to home health agencies. An initial threshold of 40 percent of claims was established to flag aberrant providers. Time Frame: April 1, 2006 March 31, 2009 Findings: Thirty one home health agencies had 40 percent or more of their patients referred by an individual Part B physician. The home health payment referred by these physicians was over $28 million. Some of these agencies shared one common Part B Physician who is currently under investigation. HIPPS Code Indicators 1 st position prior to 2008 always an H, effective 1/1/2008 numbers 1 5, dependent on episode and therapy services 2 nd position Clinical Severity Domain 3 rd position Functional Status Domain 4 th position Service Utilization Domain 5 th position prior to 2008 data validity flag, effective 1/1/2008 alpha/numeric related to supplies Power of a HIPPS Code Therapy HIPPS Code PPS Reimbursement Increase in $$ 1BGKW (0-13 therapy) 2BGKW ( therapy) 3BGKW (0-13 therapy) 4BGKW ( therapy) 5BGKW (20+ therapy) $2, $4, $2, $5, $7, $2, $2, Range-$1,800 $1,800-$2,200$2,200 6

7 HIPPS Code First Position Therapy Visits Background: Medicare made changes to the Home Health Prospective Payment System (HHPPS) effective January 1, The first position of a HIPPS code with a value 5 signifies 20 or more therapy visits. An initial threshold of 40 percent of claims was established to flag aberrant providers. Time Frame: February 1, 2008 January 31, 2009 (1 Year) Findings: 4 home health agencies were selected for further investigation in that they had a percentage as high as 87 percent of billing the HIPPS code with a value 5 in the first position. The total payment for this HIPPS code billed by these 4 providers was over $4.5 million in a one year time period. HIPPS Code Second Position Clinical Severity Background: The second position of a HIPPS code with a value C signifies the most severe clinical condition. An initial threshold of 40 percent of claims was established to flag aberrant providers. Time Frame: February 1, 2008 January 31, 2009 (1 Year) HIPPS Code Third Position Functional Status Background: The third position of a HIPPS code with a value Hsignifies the lowest level of functional status. An initial threshold of 40 percent of claims was established to flag aberrant providers. Time Frame: February 1, 2008 January 31, 2009 (1 Year) 7

8 Static Diagnosis Code Background: This indicator is to look for home health agencies billing the same diagnosis codes for the majority of their beneficiaries. An initial threshold of 50 percent of claims was established to flag aberrant providers. Time Frame: April 1, 2006 March 31, 2009 Exact Five Visits per Claim Background: Medicare makes a low utilization payment adjustment (LUPA), if at the end of an episode there have been four or fewer services. The home health agency will be compensated based on a per visit amount, which is much less than an episode rate. This indicator looks for home health agencies with a high percentage of beneficiaries with exact five visits per claim. An initial threshold of 40 percent of claims was established to flag aberrant providers. Decision Tree 8

9 Predictive Modeling Predictive models depend on a property known statistically as stationarity, meaning that its statistical properties do not change over time. Training Data: Predictive Modeling starts with a training data set. The training data set consists of cases (also known as observations, examples, instances, or records). Associated with each case is a vector of input variables (also known as predictors, explanatory variables) and a target variable (also known as a response, outcome, or dependent variable). Score Data: Predictive Modeling ends with a score data set. The score data set has the same structure as the training data set, but lacks a target variable. Predictions: The bridge between the training and score data is the predictions generated by the analysis. The predictions represent the best guess for the unknown target variable score data, based on the known input variables and associations between the input variables and target variable learned from the training data. Model Essentials Predict New Cases: the simplest type of prediction is the decision. Decisions usually are associated with some kind of action (such as classifying a home health agency as potentially fraudulent or not fraudulent). For this reason, decisions are also known as classifications. Select Useful Inputs: to select an independent set of inputs that are correlated with the target. Optimize Complexity: Underfitting versus. Overfitting Data Splitting For honest assessment of model performance, data splitting is performed. A portion is used for fitting the model, that is, the training data set. The remaining data is separated for empirical validation. The validation data set is used for monitoring and tuning the model to improve its generalization. The tuning process usually involves selecting among models of different types and complexities. 9

10 Decision Trees Decision Trees addresses each of the modeling essentials described earlier. Cases (home health agencies) are scored using prediction rules. A split search algorithm facilitates input selection. Model complexity is addressed by pruning. Decision Tree Prediction Rules: The rules are arranged hierarchically in a tree like structure with nodes connected by lines. The nodes represent decision rules, and the lines order the rules. Decision Tree Split Search Split Search starts by selecting an input for partitioning the available training data. For a selected input and fixed split point, two groups (branches) are generated. A Pearson Chi Square statistic is used to test if the proportion of zeroes and ones in the left branch is significantly different than the proportion in the right branch. A large difference in outcome proportions indicates a good split. The statistic is converted to a p value. For large data sets, these p values can be very close to zero. For this reason, the quality of a split is reported by logworth = log(chi squared p value). SAS Enterprise Miner uses a logworth of approximately 0.7. The process repeats until there are no more allowed splits whose logworth exceeds the thresholds. The resulting partition of the input space is known as the maximal tree. It is likely that the maximal tree will fail to generalize well on an independent set of validation data. Decision Tree Split Search 10

11 Pruning a Decision Tree To avoid potential overfitting, many predictive modeling procedures offer some mechanism for adjusting model complexity. For decision trees, this process is known as pruning. The general idea of pruning is to select the simplest model with the best validation performance. Misclassification measures the fraction of cases where the decision does not match the actual target value. Decisions require low misclassification. Pruning a Decision Tree Conclusion In three years, Medicare paid around $4.2 billion for beneficiaries with five or more consecutive home care episodes. Out of the $4.2 billion, over $1.1 billion were paid for beneficiaries with twelve or more consecutive home care episodes. There were over 100 home health agencies that had 50 percent or more outlier payment claims. Out of these providers the majority concentrates in the State of Texas. The total outlier payment made in ZPIC Zone 4 states was over $270 million in a three year time period. Based on the fraud indicators and decision tree rules, an initial selection of 24 home health agencies currently are under investigation. The total Medicare payment at risk is over $365 million. Investigative Findings: These providers were found to bill for a large percentage of their patients who were either not home bound, not in need of a skilled level of care, or have been educated to self administer insulin injections for an extended period of years. Due to the high volume of fraud found in our states, a separate task order was initiated for this effort. 11

12 References Applied Analytics Using SAS Enterprise Miner 5 Code of Federal Regulations CMS Manual Publication , , , Contact Information Holly Pu Chief Statistician Health Integrity, LLC 9240 Centreville Road Easton, MD puh@healthintegrity.org Telephone: Fax: