Patent Application
20130085769
Not Applicable.
The present invention is in the technical field of Healthcare Fraud, Abuse and Waste Prevention and Detection. More particularly, the present invention uses non-parametric statistics and probability methods to calculate a score that mathematically describes normal, typical, acceptable or “good” healthcare provider, claim, beneficiary or healthcare merchant traits and behavior. The invention is intended for use by government, public sector healthcare payers and private sector healthcare payers, as well as any healthcare intermediary. Healthcare intermediary is defined as any entity that accepts healthcare data or payment information and completes data aggregation or standardization, claims processing or program administration, applies rules or edits, stores data or offers data mining software, performs address or identity analysis or credentialing, offers case management or workflow management or performs investigations for fraud, abuse, waste or errors or any other entity which handles, evaluates, approves or submits claim payments through any means. The invention uses historical databases to develop a score that summarizes peer group performance and compares current provider, beneficiary, claim or healthcare merchant transactions to the typical performance of their respective peer group to identify healthcare providers, beneficiaries, claims or healthcare merchants that exhibit normal or typical behavior. The invention can be applied within a plurality of healthcare segments such as
Hospital, Inpatient Facilities, Outpatient Institutions, Physician, Pharmaceutical, Skilled Nursing Facilities, Hospice, Home Health, Durable Medical Equipment and Laboratories. The invention is also applicable to a plurality of medical specialties, such as family practice, orthopedics, internal medicine and dermatology, for example. The invention can be deployed in diverse data format environments and in separate or a plurality of geographies, such as by zip code, county, metropolitan statistical area, state or healthcare processor region.
The invention's characterization models enable the collection and storage of legitimate historical healthcare data, such as claims data, that is tagged as normal, typical, good or acceptable as well as fraudulent, aberrant, incorrect, improper, wasteful, over-serviced, over-utilized and abusive, in order to validate the characterization score and to eventually provide a “feedback loop” to enable future weighted non-parametric based unsupervised models or parametric based “supervised” score model development. A supervised score model is here defined as a statistical score model that has a dependent variable and an unsupervised model is defined as a score model that does not have a dependent variable. The characterization score provides the probability, or likelihood, that any individual observation exhibits normal or typical behavior. Additionally, the characterization model outputs score reasons corresponding to why an observation scored as it did based on the specific variables with the highest probabilities of characterizing an observation as normal or typical. Once the characterization score model is developed on historical data, it is deployed in a production environment that scores current provider, beneficiary, claim or healthcare merchant transactions in order to estimate and rank the likelihood that the current transaction is “good”, “typical” or “normal”.
The present invention is in the technical field of Healthcare Fraud, Abuse and Waste Prevention and Detection. More particularly, the present invention is in the technical field of Healthcare Payment Fraud, Abuse and Waste Prevention and Detection where it pertains to provider, beneficiary or merchant healthcare claims and payments reviewed by government agencies, such as Medicare, Medicaid and TRICARE, as well as private commercial enterprises such as Private Insurance Companies, Third Party Administrators, Medical Claims Data Processors, Electronic Clearinghouses, and Claims Integrity Organizations that utilize edits or rules and Electronic Payment entities that process and pay claims to healthcare providers. More particularly, this invention pertains to identifying normal, typical, acceptable or “good” behavior by providers, beneficiaries or healthcare merchants in a plurality of healthcare segments, including:
Healthcare providers are here defined as those individuals, companies, entities or organizations that provide a plurality of healthcare services or products and submit claims for payment or financial gain in the healthcare industry segments listed in items 1-10 above. Healthcare beneficiaries are here defined as individuals who receive healthcare treatments, services or products from providers. Beneficiary is also commonly referred to as a “patient”. The beneficiary definition also includes individuals or entities posing as a patient, but are in fact not a legitimate patient and are therefore exploiting their role as a patient for personal or financial gain. Healthcare merchant is described as an entity or individual, not meeting the exact definition of a healthcare provider, but having the ability to offer services or products for financial gain to providers, beneficiaries or healthcare intermediaries through any channel, including, but not limited to retail store, pharmacy, clinic, hospital, internet or mail.
In particular, the present invention includes the description of provider related normal behavior, as well as individual medical claim, beneficiary or healthcare merchant normal behavior as a part of healthcare fraud, abuse and waste prevention and detection in the above referenced healthcare segments and markets. More particularly, the present invention uses non-parametric statistics and probability density methods to describe the characteristics of normal provider, beneficiary, claim or healthcare merchant behavior.
Existing fraud, abuse and waste prevention and detection analytical technology generally focuses on detecting or describing the behavior of “bad” or fraudulent, abusive or wasteful providers, beneficiaries, claims or healthcare merchants. However, there is much to be said in favor of describing the characteristics of a “good claim”, a “good provider”, a “good beneficiary” or a “good healthcare merchant”, rather than a constantly changing set of “bad guy” definitions or characteristics. In fact, the “good” behavior model is less complex to verify because it can be assumed a provider, beneficiary, claim or healthcare merchant is good until indicated bad, similar to statistical hypothesis-testing, where it is assumed a state of “NO Difference” exists unless “demonstrated” otherwise. In general, “typical”, “consistent” or “normal” behavior can be expected to occur at a far higher rate, rather than relatively smaller number of occurrences of rare, unstable and varied inconsistent or non-normal behavior. Normal behavior is more stable and predictable and there are a far larger number of typical or “good” providers, claims and beneficiaries to use in building the model.
The invention includes multi-dimensional capabilities that gauge the likelihood of normal patterns of behavior, including but not limited to, healthcare claims, providers, beneficiary (individual/patient) or healthcare merchant. The invention is a scoring model, which combines separate score model dimensions for claims behavior, provider behavior, beneficiary behavior and healthcare merchant behavior. Each dimension is a sub-model in itself, with further models created and segmented by additional dimensions, including but not limited to, provider specialty and geography, or patient or population illness burden or morbidity or disease state. Each sub-model provides a score, which summarizes the likelihood that either separately or combined, one or more of the dimensions has claim, provider or beneficiary characteristics with usual, normal or “good” behavior.
Fraudulent, abusive or wasteful perpetrators typically change and adapt their behavior to avoid new techniques that are constantly being developed to detect and thwart their illicit behavior. Fraudsters, for example, by their natures, are continually plotting and scheming to find ways to beat the system while remaining anonymous and unpredictable. When building the fraud prevention models it is difficult to a-priori design and build models that effectively describe or predict all the different forms of “bad” behavior, and so, by definition, the model builders would have to wait to discover them after the fact. Thus the fraudster behavior characteristics are often transient, inconsistent and difficult to identify, and define. Therefore, fraud models intended to describe this “bad” behavior, behavior that is constantly changing, is like trying to describe a moving target or elusive quarry. Particularly when “bad” has so many different definitions, such as fraud, abuse, waste, over-servicing, over-utilization or any number of data entry error types. Additionally, there are generally a very small number of examples of any one type of “bad” behavior because it is estimated that less than 10% of all claims, providers, beneficiaries or healthcare merchants are considered to be “bad”, while the majority, (90%) of providers, claims, beneficiaries or healthcare merchants are not fraudulent or abusive. This disparity means there is available a much larger set of data for describing “normal” or “good” behavior than is available to describe “bad” behavior. The larger pool of more homogeneous data for describing “good” behavior also means there is more likely to be statistical model stability. The optimal approach to understanding and preventing healthcare fraud, abuse and waste is to build both types of models, those that detect “bad guys” and those that describe the good behavior.
The “good” behavior invention is designed and based on the concept that the majority of the submissions are “normal” or “NOT Bad” claims, providers, beneficiaries or healthcare merchants similar to statistical hypothesis testing where it is assumed that there is “no statistical difference” until demonstrated otherwise. Variables are created with the objective of describing “good” behavior. Historical data is used for score model development. Then, new incoming observations are scored based upon the historical data score model data. Non-parametric and probability statistics and mathematical techniques are used to build the score models and to mathematically describe the data distributions and test them for consistency. Rather than giving more points, on a score type scale, for odd, unusual or “bad” behavior, these “good” score models assign the most points for behavior that is centered about the mid-point of the data distribution under the assumption that providers, claims, beneficiaries or healthcare merchants that are nearest to the “midpoint” value of other similar providers, claims, beneficiaries or healthcare merchants that are “normal” or “not unusual”. Each variable in the “good” model is rescaled, or transformed, into an “intermediate” score so the mid-point value of that variable receives a maximum score and the outer values in the ends, or tails, of the distribution receive a minimum score. The midpoint of a distribution is here defined as a measure of a data distribution's central tendency, such as the median, for example. As characteristics deviate from the mid-point value, they receive fewer points so that those values that are outliers receive near zero points. These low scoring outliers with their low point values then define the “non-normal” distributional boundaries and identify the opposite of “good”, or “normal”, characteristics.
This final “good” score is a single number that represents the likelihood that a particular provider, claim, beneficiary or healthcare provider is “good” or “normal”. This single, scalar value is derived by combining the multiple, individual variable scores into one value. Once all the individual variable scores are calculated, the final, total score is calculated that combines all the individual, variable scores into the one overall score. It is important to point out that the final score is a single valued function, even though it may originate from either the high side or the low side of the distribution. In order to distinguish which side of the distribution the scored observation originates, the low side scores have a “negative” value (less than “0”) and high side of the distribution scores have a positive value (greater than 0). Therefore, scores closest to zero (0) indicate the most “normal” or “typical” values, while high positive and high negative values indicate characteristics that are farthest from the midpoint and, therefore, assumed to be least “typical” or normal.
It is important to build a score that characterizes “normal” behavior in order to complement a score that describes “bad”, or fraudulent, abusive or wasteful, behavior because the data rich “good” characteristics tend to be more stable and predictable over time. Using non-parametric measures, such as the Median and Percentiles, helps to ensure that the underlying distributional irregularities do not negatively impact the calculation of the score. Additionally, “good” characteristics, by their nature, have relatively strong central tendency with distributions that have a declining end, if skewed, and ends, or “tails”, if not skewed, away from the midpoint, regardless of the underlying data distribution shape, limitations and restrictions. This “tailing away” behavior allows for specification of potentially “bad” behavior at the ends of the distributions, both above and below the “good” or “typical” middle range of values. Conversely, the extreme end of the distribution behavior characteristics also add strong credibility to “good-ness” for those observations that fall close to the distribution central tendency or median. Bad, or unusual, behavior can occur at the “low” end of the distribution as well as the “high” data value end of the distribution. For example, a fraudulent provider may submit a bill for procedures that rarely occur in a particular specialty and therefore be at the low end of the data distribution. Or, a provider may have an unusually low amount of activity where high activity is generally considered “normal”.
In summary, this invention uses non-parametric statistical measures and probability mathematical techniques to calculate deviations of variable values from the Midpoint of a data distribution. It transforms the raw data values for each variable into a cumulative distribution function (CDF) and then combines all of the individual variable CDF values into a single scalar value that is a “good-ness” score. This “good” behavior modeling develops a score that describes “goodness” or normality, rather than simply characterizing fraudulent, abusive or wasteful behavior, or “badness”. The good model can be used to compliment a fraud, abuse or waste prevention or detection model to reduce false positive rates. The “good” score can be viewed as a measure of how likely it is that the data comes from a population of behavior characteristics representing a “good” or “normal” provider, claim, beneficiary or healthcare merchant. The optimal fraud, abuse or waste prevention or detection risk management program or system should include both a “good” behavior model and a “bad” behavior score model.
This invention uses non-parametric statistics and mathematical probability techniques to analyze historical healthcare claims data and to create a “characterization template” or “characterization score model” based on historical data which can then be used to score current, incoming claims or claim payment information for the purpose of evaluating whether a claim, group of claims, provider, beneficiary or healthcare merchant is considered to exhibit “normal good behavior” or “typical good behavior” compared to the historical data and compared to relevant peer groups. The overall score along with “reason codes” which are generated to aid in explaining why an observation scored as it did, are then deployed in a “production environment” to estimate and rank new claims, providers, beneficiaries and healthcare merchants on their relative likelihood of being “good” or “typical” or “normal”.
The score development and deployment sequence of steps is outlined below:
Characterization Template. The preceding steps are a characterization “template”, which is used to process input provider, claim, beneficiary or healthcare merchant transactions for scoring, evaluation and reason coding. A characterization predictive modeling schematic is presented graphically, as shown in
The claim model dimension, for example, with further segmentation for healthcare segment or specialty group or geography, ascertains whether a specific claim has a likelihood of usual or normal behavior. A plurality of characterization variables can be used in the claim model to determine the likelihood of usual or normal behavior. Examples of characterization variables that may be included in a claims characterization model include, but are not limited to:
The provider or healthcare merchant model dimension, with further segmentation for healthcare segment, specialty group, geography or illness burden, ascertains whether a specific provider or healthcare merchant has a likelihood of usual or normal behavior. A plurality of characterization variables can be used in the provider or healthcare merchant model to determine the likelihood of usual or normal behavior. Examples of characterization variables that may be included in a provider or healthcare merchant characterization model include, but are not limited to:
The beneficiary model dimension, for example, with further segmentation for healthcare segment or specialty group or geography, ascertains whether a specific beneficiary has a likelihood of usual or normal behavior. Beneficiary demographics can also be used to provide further segmentation. A plurality of characterization variables can be used in the beneficiary model to determine the likelihood of usual or normal behavior. Examples of characterization variables that may be included in a claims characterization model include, but are not limited to:
Higher characterization score values for provider, claim, beneficiary or healthcare merchant characterization models indicate a higher likelihood that an observation is “normal” or “typical”. Lower score values indicate that an observation is abnormal or not typical.
“Good” provider, claim, beneficiary and healthcare merchant models, based upon the invention, are designed and created using a plurality of external data sources, for example, such as credit bureau, address or negative sanction files and historical healthcare data from past time periods, from 6 months, up to 3 years previously. Data is summarized, edited and “cleaned” by dealing with missing or incorrect information for each characteristic. In addition to the raw variables being used in the invention, a large number of variables are also designed and created, through transformations. Examples of transformations include the number of patients seen by a provider in one day, one week, one month or beneficiary co-morbidity and number of claims per patient in one month, 6 months and 1 year.
Characterization variables used to create models for each dimension in the invention are compared to peer group behavior, including but not limited to healthcare claims, providers, beneficiaries or healthcare merchants, to determine if their behavior is “typical” or normal of other participants in their peer group. A “peer group” is here defined as a group of members of the same dimension, including but not limited to healthcare claims, providers, beneficiaries or healthcare merchants. For example, a peer group for providers might be their medical specialty, such as pediatrics in a specified geography, such as one state or county.
Characterization score models from the invention are built using variables that can be used in a production environment when the model is deployed. Characterization variables used in the model must be adaptable to changing behavior trends or new conditions. For example, models in production must be able to calculate a score for a new provider, versus an existing provider.
Each claim, provider, beneficiary and healthcare merchant, model is designed, created and utilized to measure behavior and provide a final score to be used as a single number that represents “normal” or typical behavior. This final model score is then used in production by scoring new incoming claims as they are processed.
The “champion” or incumbent characterization model will continue to be used until another model is developed and enhanced and it then replaces the previous model in production. Predictive models are monitored, validated and optimized regularly. Models are optimized or redeveloped as experience is gained regarding the value of existing variables or the introduction of new models, model performance deteriorates or new information or new behavior patterns are identified, providing the opportunity for improvement.
The final model is then put into production in a model deployment process where it is used to score separate predictive model dimensions, including but not limited to, claims behavior, provider behavior, beneficiary behavior and healthcare merchant behavior. The model can be deployed on a “real time” or “batch mode” basis. Real time scoring occurs as a claim is received and processed by the payer. The score can also be calculated in “batch mode” where it is calculated on all claims received in regularly scheduled batches, for example hourly or daily batches.
This invention uses non-parametric statistics and mathematical probability techniques to analyze historical healthcare claims data and to create a “characterization template” or “characterization scoring model” based on current and historical data. This model can then be used to score current, incoming claims, providers, beneficiaries or healthcare merchants for the purpose of evaluating whether an observation (claim, for example), group of claims, provider, beneficiary or healthcare merchant is considered to exhibit “normal good behavior” or “typical good behavior” compared to the historical data and compared to relevant peer groups. Each of the variables on the incoming group of claims is converted to an estimate that the individual variable displays “normal” or “typical” characteristics or values. Some entities, in a general sense, refer to this mathematical estimate as a “profile”. The goal then is to build a characterization scoring model from historical medical claims from this valid data that define and describe the “normal” or “typical” claim, provider, beneficiary and healthcare merchant patterns of behavior. The entire score development and deployment sequence of steps are outlined below. The complete set of transformations that are involved in the data analysis and characterization score development and deployment process are performed in the following manner.
1. Non-Parametric Statistical Calculations
For continuous or discrete interval variables, perform the following steps individually for each of these characterization variables. Since it is hypothesized that at least the bulk of a variable's distribution is homogeneous, define a “reasonable” centralized homogeneous mid-fractional (equally-sized tails) data distribution area of containment of size β, initially, for example, β=0.80=the mid 80% of the distribution. This area is an estimate of the data distribution's homogeneous area for that variable. Then determine numerically the boundary percentiles for that β:
P-low, L=P%[(1−β)/2], and
P-high, H=P%[(1+β)/2], for that variable.
If (β=0.80, then P-low is the 10th percentile value of that particular variable and P-high is the 90th percentile value of that particular variable. Then calculate:
R=(H−L)/β; the projected (estimated) 100% range of the distribution
MP=(H+L)/2; the projected (estimated) symmetric distribution mid-point
LB=MP−R/2; the projected (estimated) 0th %-ile distribution lower bound
UB=MP+R/2; the projected (estimated) 100th %-ile distribution upper bound
The estimated midpoint of a distribution is a non-parametric measure of a data distribution's central tendency (its centrality). It can be measured by several different statistics such as the arithmetic mean, the median, and (as developed here) the projected mid-point. If the data are symmetric and everywhere homogeneous in nature, these various measures will tend to be equal to the same value. If, however, the distribution is heterogeneous, bimodal or skewed, or if it contains significant outliers, then these central tendency measures will return differing values, depending upon the degree and scope of the heterogeneity and skewness. Skew is defined here as the degree of asymmetry of the distribution around the selected measure of centrality (often the mode for homogeneous distributions), where the body or tail on one side of the distribution is consistently longer than that of the other. It is important to note here the difference between overall homogeneous skewness of a distribution (the result of a naturally skewed physical process), as compared to the effect significant outliers can have on the “apparent skewness” of a fundamentally homogeneous un-skewed distribution. One of the primary challenges facing an outlier detection system is to be able to distinguish accurately, consistently, and with precision, the naturally skewed but well-behaved distribution (the desired “good guys” data distribution format) from the distribution that also contains maverick outliers (the undesired “bad guys”).
2. Distribution Mid-point Deviations—G-Values
As an observation is processed, each of the raw data characterization variables is converted into a standardized positive or negative deviation from the distribution midpoint using non-parametric statistical techniques. This deviation calculation is calculated by subtracting the variable's midpoint value from the raw data value and then dividing this difference by R/2, half the projected 100% range of the distribution, for each observation for each variable. The result of this calculation is termed the “G-Value” for a particular, individual x-variable for each observation in the data. The G-Value calculation for each x-variable observation is:
g[x]=ω
x·(x−MP)/(R/2), or
g[x]=ω
x·[2·(x−MP)/R]
where
ωx=optional variable importance weighting factor for variable x(0≦ωx≦1)
Note: hereafter ωx=1 to simplify notation and analyses (excepting binaries as noted)
β=assumed mid-homogeneity Span; [0.5≦13<1]
L=P%[(1−β)/2]; H=P%[(1+β)/2]
MP=(H+L)/2; R=(H−L)/β
Choosing an optimal value for the β is important in that it optimizes the trade-off between the potential for false-positive and false negative conclusions as well as the potential total percent of estimated “good” or “typical” or “normal” observations. A false-positive is defined as reaching the conclusion that a variable value is a maverick outlier when it naturally is not, whereas a false-negative is reaching the conclusion that a variable value is a member of the homogeneous, typical distribution when it is not. Unless it is practical to examine every observation individually and in detail, false positives and false negatives are unavoidable outcomes of large-volume decision-making models.
3. Cumulative Distribution Function Estimates—H-Values and Lambda
In order to better understand non-zero variable G-Values and their likelihood, each value is transformed into a Cumulative Distribution Function (CDF) value using a simple sigmoid logistic transfer function. A CDF is here defined as a mathematical function that describes the probability that a variable's value is found at a value less than or equal to the distributional value. The general sigmoid function Sig is of the form
Sig[t]:=1/(1+e−t), −∞<t<∞
Note that for real values oft, Sig[t] is everywhere positive and ranges from 0 to 1. These sigmoid transformed variables are identified as H-Values, and these H-Values are made a function of g[x] above. The specific formula used for H-Value computations is:
H[x≦x
0]=1/(1+e−λ·g[x])
where g[x] is the G-Value described and calculated above, and Lambda, λ, is a scaling coefficient that adjusts the H-equation so that g[x]=1 which provides an H-Value probability of (1+β)/2, and g[x]=−1 provides an H-Value probability of (1−β)/2. The following calculations are made to determine the λ value:
For g[x]=1
H[g[x]=1]=1/(1+e−λ·1)=(1+β)/2
e
−λ=2/(1+β)−1=(1−β)/(1+β)
λ=Ln[(1+β)/(1−β)]
Note the desired symmetry for g[x]=−1:
H[g[x]=−1]=1/(1+e−λ·1)=(1−β)/2
e
λ=(1+β)/(1−β)
λ=Ln[(1+β)/(1−β)]
For example, for {β=0.9}→λ=2.944. Note that λ is a function only of β, not x, and so the determination of β (and thus the associated λ) should be made individually for each variable based on the assumption of the scope of homogeneity and shape of that variable's distribution.
4. Transformation to T-Values. For decision-making and variable value inlier identification it is desired to have a maximum “good guy” score at the individual variable's distribution midpoint (MP) and have progressively smaller scores away from that MP, in both the positive and negative direction from the MP. To accomplish this maximum value at the midpoint, the G-Values are transformed into associated T-Values as follows.
T[g[x]]=2/(1+e|λ·g[x]|); 0<T[g[x]]<1, |y|→absolute value of y, therefore, |λ·g[x]|=absolute value of λ·g[x].
The resulting triangular T transform has a maximum value of 1 at g[x]=0, which occurs when x=MP, and tails off quickly toward zero as absolute value g[x], |g[x]|, becomes large. The T-Value is a pyramid shaped distribution rather than sigmoid, but tends to follow the distributional shape of the H[g] tails, as illustrated in
5. Binary Variables—T-Transformation. If they are present in the data as contributors rather than simply discriminators, binary variables can also be included in this T-Value transformation process. Numeric binary variables are defined here as data that have only two numeric indicator values, such as zero and one (b i: 0, 1). Because of this restriction they are much simpler to address mathematically than are wide-ranging interval variables. Specifically, assume that the occurrence of an event or condition ξ, is coded 1 and its non-occurrence 0. Examination of the N observations in the data indicates that ξ occurs in Phi, φ, proportion of the N observations and does not occur in (1−φ) proportion of the N. Since binaries have only two dimensionless numeric states, for example zero and one (0,1), the T transformation can be done directly, rating the binaries as follows:
bi=1: T[1]=ωbφ
bi=0: T[0]=ωb·(1−φ)
where, Omega, ωb, (0≦w≦1) is an importance-weighting constant for that binary variable, similar to Omega, ωx for the g-computation. Note then that for all such binary variables:
T[1]+T[0]=ωb·φ+ωb(1−φ)=ωb
6. Summarization—Combining T-Values into one Overall T-Score. Observations almost always consist of more than one variable and these variables together, as a composite, may indicate overall observational goodness or normalcy, so a summary statistical measure is needed that allows for meaningful accumulation of individual variable measures, such as combining all the T-Values. All of the T-Values associated with each of the individual variables for one observation, for example a particular provider, claim, beneficiary or a healthcare merchant, can then be combined into one overall T-Score, that represents the likelihood that this particular transaction or observation, provider, claim, beneficiary or healthcare merchant is, overall, “normal” or “a good guy”. This transformation is referred to as the observation's “ΣT-Score” (Sigma-T-Score or Sum-T) calculation. This ΣT calculation accumulates the individual variable T-Scores into a single summary score, where summary values near 1 indicate overall homogeneous good-guy behavior, and scores approaching zero indicate suspicious non-homogeneous (possible outlier) behavior. As such the ΣT score will be the final fraud, abuse and waste score for an observation. The ΣT-score is defined by control-coefficients φ (phi) and δ (delta), by the following formula.
ΣTφ,δ=[Σt=1,kωt·Ttφ+δ]/[Σt=1,kωt·Ttφ]
Where, ωt is the weight for variable Tt, (0≦ωt≦1, use of ωt is optional);
δ (delta) is a power increment, such as 1, 1.2, 1.8, 2.1, 3.0, etc. (Note that for this invention φ and δ are both positive, but do not need to be integers.) The ΣT-Score is the summary estimate of all of the variables and how they combine to create an overall estimate of “goodness” for each observation (T-Score).
7. Reason Code List. The next step in the score development and evaluation process is to create a reason code list that explains why observations with the lowest ΣT scores scored as they did, based on the component individual variable T-Scores. The characterization variable associated with the smallest T-Value for that observation is the primary, number one variable, and therefore reason, that contributed negatively (i.e., is the least “normal” or “typical”) to the overall “good-guy” score for the provider, claim, beneficiary or healthcare merchant being scored. The variable with the second smallest T-Value is the next highest negative contributor, etc.
8. Score Deployment. The final step in developing a scoring model is to deploy it so it can be used to score a large block of new, incoming transactions. Each of the variables on the incoming claims is converted to a G-Value and, ultimately a T-Value that indicates the individual variable's values that express “typical” or “normal” characteristics of this variable. These individual variable estimates are then combined into an overall ΣT score, which is an estimate of the overall degree to which the claim, provider, beneficiary or healthcare merchant associated with that particular observation, or group of variables, is typical or normal and acceptable, i.e., an inlier. The individual T-Value and overall ΣT scores (along with necessary reason codes) are part of the deployed “production environment” which scores new claims, providers, beneficiaries and healthcare merchants on their relative likelihood of being “good” or “typical” or “normal”.
9. Summary. While this invention may be embodied in many different forms, there are described in detail herein specific preferred embodiments of the invention. This description is an exemplification of the principles of the invention and is not intended to limit the invention to the particular embodiments illustrated. The present invention is a “Normal-Behavior Characterization Scoring Model” that is designed to focus primarily on normal or typical values of the variable distributions for each observation scored by the model. The individual T and summary ΣT scores are hereby defined as the values that represent the likelihood that one or more of the claims, provider, beneficiary or healthcare merchant characterization variables, are likely to represent normal or typical behavior. The highest scores are closer to the midpoint of the variable values or the observation's data distribution, while the lowest scores are farther from the distribution midpoint.
Referring now to
Referring now to
The overall scoring process is shown in
Referring now to
From the Historical Data Summary Statistics Data Security Module 105 the data is sent to the Raw Data Preprocessing Module 106 where the individual claim data fields are then checked for valid and missing values and duplicate claim submissions. The data is then encrypted in the Historical Data Summary Statistics External Data Security Module 107 and configured into the format specified by the Application Programming Interface 108 and sent via secure transmission device to an external data vendor's Data Vendor Data Security Module 109 for un-encryption. External Data Vendors Module 110 then append(s) additional data such as Unique Customer Pins/or Universal Identification Device (UID) to assign proprietary universal identification numbers, to append, for example, the Social Security Death Master File, Credit Bureau such as credit risk scores and/or a plurality of other external data and demographics, Identity Verification Scores and/or Data, Change of Address Files for Providers or Healthcare Merchants, including “pay to” address, or Patients/Beneficiaries, Previous provider, healthcare merchant or beneficiary fraud “Negative” (suppression) files or tags (such as fraud, provider sanction, provider discipline or provider licensure, etc.), Eligible Beneficiary Patient Lists and Approved Provider or Healthcare Merchant Payment Lists. The data is then encrypted in the Data Vendor Data Security Module 109 and sent back via the Application Programming Interface in Module 108 and then to the Historical Data Summary Statistics External Data Security Module 107 to the Appended Data Processing Module 112. If the external database information determines that the provider, healthcare merchant or patient is deemed to be deceased at the time of the claim or to not be eligible for service or to not be eligible to be reimbursed for services provided or is not a valid identity, at the time of the original claim date, or any other reason not considered here today, the claim is tagged as “invalid historical claim” and stored in the Invalid Historical Claim Database 111. These claims are suppressed “negative” files for claim payments and may or may not be used in calculating the summary descriptive statistical values for the “Good Provider” characterization score. They may be referred back to the original claim payer or processor and used in the future as an example of fraud. The valid claim data in the Appended Data Processing Module 112 is reviewed for valid or missing data and a preliminary statistical analysis is conducted summarizing the descriptive statistical characteristics of the data.
A copy of claim data is sent from the Appended Data Processing Module 112 to the Claim Historical Summary Statistics Module 115 where the individual values of each claim are accumulated into a claim characterization score using calculated variables by industry type or segment, provider, healthcare merchant, patient, specialty and geography. Examples of individual claim variables include, for example, but are not limited to: fee amount submitted per claim, sum of all dollars submitted for reimbursement in a claim, number of procedures in a claim, number of modifiers in a claim, change over time for amount submitted per claim, number claims submitted in the last30/60/90/360 days, total dollar amount of claims submitted in the last 30/60/90/360 days, comparisons to 30/60/90/360 trends for amount per claim and sum of all dollars submitted in a claim, ratio of current values to historical periods compared to peer group, time between date of service and claim date, number of lines with a proper modifier, ratio of amount of effort required to treat the diagnosis compared to the amount billed on the claim.
Within the Claim Historical Summary Statistics Module 115, a plurality of historical descriptive statistics are calculated for each variable for each claim by industry type, specialty and geography. Calculated historical summary descriptive statistics include measures such as the median and percentiles, including deciles, quartiles, quintiles or vigintiles. Examples of historical summary descriptive non-parametric statistics for a claim would include values such as median number of procedures per claim, median number of modifiers per claim, median fee charged per claim.
The historical summary descriptive statistics, for each variable in the characterization score model, are used by G-Value Normalization Module 214 in order to calculate normalized variables related to the individual variables for the scoring model.
A copy of the data is sent from the Appended Data Processing Module 112 to the Provider and Healthcare Merchant Historical Summary Statistics Module 116 where the individual values of each claim are accumulated into claim characterization score variables by industry type, provider, healthcare merchant, specialty and geography.
Within Provider Historical Summary Statistics Module 116, a plurality of historical summary descriptive statistics are calculated for each variable for each Provider and Healthcare Merchant by industry type or segment, specialty and geography. Calculated historical descriptive statistics include measures such as the median, range, minimum, maximum, and percentiles, including deciles, quartiles, quintiles and vigintiles for the Physician Specialty Group.
The Provider and Healthcare Merchant Historical Summary Statistics Module 116 for all industry types and segments, specialties and geographies are then used by the G-Value Standardization Module 214 to create normalized variables for the scoring model.
A copy of the data is sent from the Appended Data Processing Module 112 to the Patient, or beneficiary, Historical Summary Statistics Module 117. A plurality of historical summary descriptive statistics are calculated for the individual values of the claim and are accumulated for each claim characterization score variable by industry type or segment, patient, provider, healthcare merchant, specialty and geography for all Patients, or Beneficiaries, who received a treatment, or supposedly received a treatment
The Patient Historical Summary Statistics 117 for all industry types, specialties and geographies is then used by the G-Value Standardization Module 214 to create normalized variables.
Referring now to
The data is then sent via a secure transmission device to the Score Model Deployment and Validation System Application Programming Interface Module 203 and then to the Data Security Module 204 within the scoring deployment system for un-encryption. Each individual claim data field is then checked for valid and missing values and is reviewed for duplicate submissions in the Data Preprocessing Module 205. Duplicate and invalid claims are sent to the Invalid Claim and Possible Fraud File 206 for further review or sent back to the claim payer for correction or deletion. The remaining claims are then sent to the Internal Data Security Module 207 and configured into the format specified by the External Application Programming Interface 208 and sent via secure transmission device to External Data Security Module 209 for un-encryption. Supplemental data is appended by External Data Vendors 210 such as Unique Customer Pins/Universal Identification Descriptors (UID) Social Security Death Master File, Credit Bureau scores and/or plurality of other external data and demographics, Identity Verification Scores or Data, Change of Address Files for Providers, Healthcare Merchants or Patients/Beneficiaries previous provider, healthcare merchant or beneficiary fraud “Negative” (suppression) files, Eligible Patient and Beneficiary Lists and Approved Provider or Healthcare Merchant Lists. The claim data is then sent to the External Data Vendors Data Security Module 209 for encryption and on to the External Application Programming Interface 208 for formatting and sent to the Internal Data Security Module 207 for un-encryption. The claims are then sent to the Appended Data Processing Module 211, which separates valid and invalid claims. If the external database information reveals that the patient or provider is deemed to be inappropriate, such as deceased at the time of the claim or to not be eligible for service or not eligible to be reimbursed for services provided or to be a false identity, the claim is tagged as an inappropriate claim or possible fraud and sent to the Invalid Claim and Possible Fraud File 206 for further review and disposition.
A copy of the individual valid current claim or batch of claims is also sent from the Appended Data Processing Module 211 to the G-Value Standardization Module 214 in order to create claim level variables for the characterization score model. In order to perform this calculation the G-Value Standardization Module 214 needs both the current claim or batch of claims from the Appended Data Processing Module 211 and a copy of each individual valid claim statistic sent from the-Claim Historical Summary Statistics Module 115, Provider and Healthcare Merchant Historical Summary Statistics Module 116 and Patient Historical Summary Statistics Module 117. The G-Value Standardization Module 214 converts raw data individual variable information into non-parametric values. When using the raw data from the claim, plus the statistics about the claim data from the Historical Claim Summary Descriptive Statistics file modules, the G-Value Standardization Module 214 creates G-Values for the scoring model. The individual claim variables are matched to historical summary claim behavior patterns to calculate the current individual claim's historical behavior pattern of a peer group of claims. These individual and summary evaluations are non-parametric, value transformations of each variable related to the individual claim. The calculation for the G-Value transforms the raw data value for each x, into a dimensionless scaled variable g[x], where wx is an assigned importance-weighting constant (0≦w≦1) for variable x. The initial value for wx is unity (1.0) unless it is known or believed that the variable should have less weight, then the value of wx is less than 1.0.
g[x]=w
x·(x−MP2x)/IPRx
Note the characteristics of g[x]. It can range from −∞ to +∞, and has a value of zero at the projected β midpoint MP2x. If the data are naturally skewed, using MP2 for the measure of centrality instead of Q2x tends to compress the longer leg of the g[x] distribution and extend the shorter leg of g[x].
The T-Value Sigmoid Transformation Module 215 converts the G-Value normalized variables into estimates of the likelihood of being a normal behavior pattern.—It is important to have a single measure of likelihood of observing a large but legitimate value for each variable that will be a part of the characterization scoring model. Therefore, the T-Sigmoid Transformation Module 215 converts the G-Values in G-Value Standardization Module 214 to a sigmoid-shaped distribution that approximates a traditional cumulative distribution function (CDF).
This T-Value provides an estimate that the raw data value for this observation has a normal or typical pattern of behavior. All characterization variables and their corresponding T-Values are then sent from the T-Value Sigmoid Transformation Module 215 to the Sum-T Score Calculation Module 216. At this point there is a collection of n-different T-Value values for each of the “n” variables in the score model. Each characterization variable measures a different characteristic of the individual claim, or batch of claims, the Provider, the Healthcare Merchant and the Patient. These characterization variable values, T-Values, that are estimates of being a normal or typical pattern, can then be aggregated into a single value, Sum-T. This Sum-T function is used to obtain one value, a score, which represents an estimate of the overall likelihood that the current observation reflects a normal or typical pattern of behavior. Because it is necessary to have a maximum “Good” score at the distribution midpoint, and because the H-value sigmoid function is a continuous “S” shaped CDF curve, it is necessary to transform the H-values to a value that increases as it approaches the distribution midpoint and then decreases after it passes the midpoint and moves toward zero as it continues to increase away from the distribution midpoint. In order to accomplish this peak valuation at the distribution midpoint, the H-values are transformed to T-values by using the same sigmoidal logistic transfer function, except the absolute value of the term [−λ·g] is used and the numerator is “2” instead of “1”. Absolute value is here defined as the numerical value of a without regard to its positive or negative sign. This formula results in increasing values from the lowest percentile up to the distribution midpoint and then decreasing values from the midpoint to the highest percentile value. The individual T-Values can be thought of as individual variable “scores” or “intermediate” characterization scores. Note that with all the above steps completed, there is now a characterization score determined solely in terms of the predetermined λ-value and the computed g[x] for that observation's variable value. The formula for the T-value is:
T[g]=2/(1+exp[|λ·g]|)
All of the transformed “T-Values” for the characterization variables for one observation, for example a particular provider, healthcare merchant, claim or beneficiary, are then combined and transformed by combining them into one “overall” value, a characterization score, that represents the likelihood that this particular transaction or observation, provider, healthcare merchant, claim or beneficiary, is “normal”, “typical” or “good”. This T-value transformation can be done using one of two methods. The first method begins by identifying and listing all the possible percentile decimal values between 0.005 and 0.995, in increments of 0.01, with added delimiters of 0 and 1. The list of percentiles begins a tabular “process-of-location” where the variables, for each observation, can be “fit” into a percentile rank as a value in the cumulative distribution function. Each variable for each observation is “transformed” to create a “standardized” percentile, or a common percentile rank value, so that all the variable percentile values can be compared to one another. These “standardized” percentile values are then combined, using, for example, a geometric mean, to calculate a final, single, overall “Good” characterization score for that observation. In order to further screen for “good” or normal behavior, the variability of the T-Values can be used to measure “normal behavior consistency” using, for example the geometric standard deviation, or the minimum of the ratio of the low T-Value to the maximum T-Value. These statistics will provide a measure of the variability and consistency of the T-Values and provide an indication whether a group of variables are tightly centered about the average T-Value or have a wide dispersion about the T-Value measure of central tendency. The “Sum-T” calculation converts, for a set of “T-Values”, into a single summary variable that represents the likelihood that the “sum” of the “T-Values” represents a “good” provider, claim, beneficiary or healthcare merchant. The “Sum-T” value then represents the overall likelihood that this observation is typical or normal. This calculation combines the T-Values for an individual observation and summarizes their combined values into one number to represent overall “good-ness”. These individual observation “Sum-T” scores can then be summed and aggregated to compare the relative performance, or normal behavior, among a plurality of different healthcare segments, or dimensions such as geographies or across multiple provider specialties. The formula for the Sum-T is:
Sum−T: ΣTφ,δ=[Σt=1,kωt·Ttφ+δ]/[Σt=1,kωt·Ttφ]
where ΣT, Sum-T, is the summary estimate of all of the normalized score variable estimates for the characterization variables for one observation, which is the “score” for this observation, wt is the weight for variable Tt, φ (Phi) is a power value of Tt, such as 1, 2, 3, 4, etc. and δ (Delta) is a power increment which can be an integer and/or decimal, such as 1, 1.2, 1.8, 2.1, 3.0, etc. The score, ΣT, Sum-T, will have a high value, near 1.0, if any or all of the individual variable “T-Values” have high values near 1.0, thereby indicating that at least one, and perhaps more, of the variables for that observation have a high likelihood of being normal or typical. The second method, to transform the T-Values, calculates one value, termed the “Sum-T”, which is another approach to calculate the overall score. This technique is a generalized procedure that calculates one value to represent the overall values of a group of numbers or probabilities. It converts, for a set of k numbers, such as probabilities p1, p2, . . . , pk, for example, into a single generalized summary variable that represents the values of these numbers with emphasis on larger probabilistic values. This calculation then isolates the higher T-Value variable values and gives them more emphasis or weight in the calculation. In the fraud detection models, it effectively ranks the overall risk of an outlier variable being present for an individual observation. The Sum-T is the final fraud score and it is defined for control-coefficients φ and δ, as follows:
Sum−T: ΣTφ,δ=[Σt=1,kωt·Ttφ+δ]/[Σt=1,kωt·Ttφ]
Note that phi φ and delta δ do not need to be integers. For this invention the numerator powers are always greater than the denominator powers for the Sum-H function. Smaller φ values emphasize the smaller individual values over the larger ones, and larger φ values emphasize the larger individual values over the smaller. These estimates can then be used to compare the relative performance, or risk, among different geographies and across multiple provider specialties.
The individual T-score value and the individual T-Values corresponding to each variable are then sent from the T-Sigmoid Transformation Module 216 to the Score Reason Generator Module 217 to calculate score reasons for why an observation score as it did. The Score Reason Generator Module 217 is used to explain the most important characterization variables that cause the score to be highest for an individual observation. It selects the characterization variable with the highest T-Value and lists that variable as the number 1 reason why the observation scored high. It then selects the characterization variable with the next highest T-Value and lists that characterization variable as the number 2 reason why the observation scored high, and so on.
A copy of the scored observations is sent from the Score Reason Generator Module 217 to the Score Performance Evaluation Module 218. In the Score Performance Module, the scored distributions and individual observations are examined to verify that the model performs as expected. Observations are ranked by characterization score, and individual claims are examined to ensure that the reasons for scoring match the information on the claim, provider, healthcare merchant or patient. The Score Performance Evaluation Module details how to improve the performance of the Normal Behavior score model given future experience with scored transactions and actual performance on those transactions with regard to normal and abnormal performance. This process uses the Bayesian posterior probability results of the model for the T-Values of the model variables
p[V|T]=p[normal−claim|acceptable−T−Value]
p[V′|T]=1−p[V|T]
p[V|T′]=p[abnormal−claim|unacceptable−T−Value]
p[V′|T′]=1−p[V|T′]
To determine their values calculate the prior conditional and marginal probabilities
These last two conditionals are represented by distributions obtained from Module 224, the Feedback Loop of actual claim outcomes, one for the normal claims and one for the abnormal claims, and p[V] is a single value for the current version of the Feedback Loop in Module 224. These values can be determined directly from summarizing the data obtained from actual results, based on the normal/abnormal determinations. The results would be presented in the form of two relationships—the probability of misclassifying a normal claim and the probability of misclassifying an abnormal claim, based on the selected critical Tcritical value. The decision rule assumes that a claim is normal unless indicated to be abnormal and is stated as “Assume Claim Normal, then if T>T−boundary assign as abnormal”.
The advantages of the present invention include the following, without limitation.
Fraudulent, abusive and wasteful perpetrators typically change and adapt their behavior to avoid new techniques that are constantly being developed to detect and thwart their illicit behavior. When building the fraud prevention models it is difficult to a-priori design and build models to predict all the different forms of “bad” behavior, and so, by definition, the model builders would have to wait to discover them after the fact. Perpetrators of fraudulent, abusive and wasteful behavior, by their nature, are continually plotting and scheming to find ways to beat the system while remaining anonymous and unpredictable. Thus their behavior characteristics are often transient, inconsistent and difficult to detect, and define. Therefore, fraud models intended to describe this “bad” behavior, behavior that is constantly changing, is like trying to describe a moving target or elusive quarry. Additionally, there are generally a very small number of examples of any one type of “bad” behavior because only about 5% to 10% of all claims, providers or beneficiaries are considered to be “bad”, while the majority, (90-95%) are not fraudulent, abusive or wasteful. This disparity means there is available a much larger, more stable set of data for describing “normal”, “typical” or “good” behavior. The larger pool of more homogeneous data for describing “good” behavior also means there is more likely to be statistical model stability. The final “good” characterization score is a single number that represents the likelihood that this particular provider, claim or beneficiary is “good”, “typical” or “normal”. This single, scalar value is derived by combining the multiple, individual variable scores into one value for each observation.
The above disclosure is intended to be illustrative and not exhaustive. This description will suggest many variations and alternatives to one of ordinary skill in this art. All these alternatives and variations are intended to be included within the scope of the claims where the term “comprising” means “including, but not limited to”. Those familiar with the art may recognize other equivalents to the specific embodiments described herein which equivalents are also intended to be encompassed by the claims. Further, the particular features presented in the dependent claims can be combined with each other in other manners within the scope of the invention such that the invention should be recognized as also specifically directed to other embodiments having any other possible combination of the features of the dependent claims. For instance, for purposes of claim publication, any dependent claim which follows should be taken as alternatively written in a multiple dependent form from all prior claims which possess all antecedents referenced in such dependent claim if such multiple dependent format is an accepted format within the jurisdiction (e.g. each claim depending directly from claim 1 should be alternatively taken as depending from all previous claims). In jurisdictions where multiple dependent claim formats are restricted, the following dependent claims should each be also taken as alternatively written in each singly dependent claim format which creates a dependency from a prior antecedent-possessing claim other than the specific claim listed in such dependent claim below (e.g. claim 3 may be taken as alternatively dependent from claim 2; claim 4 may be taken as alternatively dependent on claim 2, or on claim 3; claim 6 may be taken as alternatively dependent from claim 5; etc.).
This completes the description of the preferred and alternate embodiments of the invention. Those skilled in the art may recognize other equivalents to the specific embodiment described herein which equivalents are intended to be encompassed by the claims attached hereto.
This application incorporates the entire contents of each of the following patent applications, utility patent application Ser. No. 13/074576, filed Mar. 29, 2011; provisional patent application 61/319554, filed Mar. 31, 2010, and provisional patent application 61/327256, filed Apr. 23, 2010.
| Number | Date | Country | |
|---|---|---|---|
| 61561561 | Nov 2011 | US | |
| 61319554 | Mar 2010 | US | |
| 61327256 | Apr 2010 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 13074576 | Mar 2011 | US |
| Child | 13617085 | US |
