1. Reduce variation in cost and utilization
2. Monitor a member's health via:
   • Adherence and Opioid Use
   • ER Utilization and Readmissions
   • Comorbidities for health risk assessment
3. Identify potentially inappropriate medical care
4. Minimize medical waste
5. Provide transparency to physicians

• WHIO is a non-profit organization dedicated to improving the quality, affordability, safety and efficiency of health care in Wisconsin
• It is the only statewide voluntary All-Payer Claims Database in the nation.
• It currently includes more than 4.8M individuals
• It contains ~$110B in billed charges

• Claims data
  • Procedures, Diagnoses, Costs, Quantities, Providers
• Eligibility data
  • Enrollment, Demographics
• Pharmacy data
  • Prescriptions, Prescribers
• Unfortunately, it does not include EMR data

• Defining Adherence
• Why does adherence matter?
  • Health implications
  • Financial implications
• Measuring adherence
  • Proportion of Days Covered (PDC)
  • Medication Possession Ratio (MPR)
  • Which is better?

• "The extent to which patients take medications as prescribed by their health care providers"\(^{1}\)
  • Fill medication in a timely manner\(^{2}\)
  • Implies active engagement and collaboration, a therapeutic alliance between the patient and provider
    
    
• Compliance
  • Suggests passively following a doctor's orders

Correlation between continuation of medication regimen and post-MI survival

Kaplan-Meier survival curve comparing patients discontinuing use of all medications at 1 month with patients continuing use of 1 or more medications among patients discharged with all 3 medications (log-rank test, \(p<.001\))\(^{8}\)

• Unintentional
  • Memory issues
  • Complex medication regimen - forgetfulness
  • Cost
• Intentional
  • Fear of side-effects
  • Lack of beliefs of necessity
  • Depression

• 50% of the US population is prescribed medication for chronic conditions
• Of those taking medication, only 50% are taking it as directed\(^{10}\)

• "Pharmaceuticals have the effect of improving or maintaining an individual's health..adhering to a drug regimen for a chronic condition such as diabetes or high blood pressure may prevent complications..taking the medication may [also] avert hospital admissions and thus reduce the use of medical services."\(^{11}\)
• Every additional dollar spent on medicines for adherent patients with congestive heart failure, high blood pressure, diabetes and high cholesterol generated $3-$10 in savings on ER visits and IP hospitalizations.\(^{12}\)

Difference in Annual Spending of Adherent Patients vs. Non-adherent Patients\(^{13}\)

• Total healthcare spending reached $1 trillion in the same period\(^{14}\)
• One estimate attributes $317.4 billion in annual spending to unnecessary medical costs to treat avoidable complications\(^{15}\)
  • ER Visits
  • Hospitalizations
  • Extra Tests

• Most measures are indirect and are based on pharmacy claims data
  • Medication Possession Ratio (MPR)
    • Inconsistently defined numerator and denominator
  • Proportion of Days Covered (PDC)\(^{16}\)
    • More sensitive
    • Consistently-defined measurement criteria
  • Both overestimate actual adherence
    • Patient may fill the prescription, but still fail to take medication

• Most studies suggest an 80% threshold
  
  
• Specific conditions (HIV/AIDS) may require higher adherence (\(\geq\) 95%)\(^{17}\)

• Increased cost
• Increased side effects
  • Diminished immune response
  • Anaphylaxis

• Typically expressed as a percentage
• General estimate of how often patients take medication as directed
• Defined as: \[\frac{Number~of~days~on~which~medication~was~available}{Number~of~days~in~the~study~period}\]
• Limitations
  • Cannot differentiate a change in therapeutic regimen from non-adherence
    • Decreased dose
    • Termination of therapy due to adverse events or clinical improvement

• Code samples available online in Python, R, SAS
• SQL Server UDF
  • Pros:
    • Adherence can be calculated within calling query
    • Patients can be identified based on adherence ratio
    • Streamlined workflow
  • Cons:
    • Efficiency
    • Scalability

• Data is aggregated by patient and by medication
• Study period pre-defined by a begin and end date, typically one year.
  • Exceptions
    • Coverage begins or ends in the middle of the study period
    • Patient starts a new medication regimen in the middle of the study period
    • Patient terminates medication for \(\geq\) 60 days
      • Treat as a termination of or change in therapy vs. non-adherence

• For each medication, create an array for each day in the study period (adjusted for exceptions)
  • The total number of days covered by the medication is determined by the number of days supplied by each prescription refill.
  • Populate the array based on the availability of medication for that date.
  • If a prescription was filled just prior to the start of the study, some medications will be available to the patient before the first drug fill within the study.
    • Look-back period considers this amount and allocates it appropriately to days at the start of the study period.

• Patient takes atovastatin (Lipitor) once a day
• Prescriptions is for 30 tablets
• Study period from 01/01/2014 to 12/31/2014
• A fill occured on 12/31/2013, which makes 29 tables available at the start of January
• Days in study: 365
• Days of medication: 359
• Adherence Ratio (PDC): 98.4%

• Best subset selection
• Forward Stepwise Regression
• Backward Stepwise Regression
• Ridge Regression
• Lasso
• Many others

• Tests all possible combinations of predictors (\(2^{p}\))
• This can be good, and this can be bad.
• Typically, this is not a smart choice when the number of parameters is \(\geq\) 10 because you end up with more than 1024 possibilities
• Imagine if you have 40 predictors, this gives 1.099511610^{12} possibilites and can be computationally expensive

• Regularization is a statistical method which penalizes many predictors (overfitting)
• Two methods include Regularization & Lasso
• Lasso \[\sum_{i=1}^N\left(y_{i}-\beta_0-\sum_{j=1}^p\beta_jx_{ij}\right)^{2} + \lambda\sum_{j=1}^p\mid\beta_j\mid\]
• This is a good, relatively newer method that will choose which variables are valuable for us.

# Set up the response variable - this must be a vector
y = reg_df$Average.PDC
# Set up the regressor matrix to pass into glmnet
x = model.matrix( Average.PDC ~.-1, data = reg_df )

# Fit the Lasso
fit.lasso = glmnet(x, y)

# Now let's cross-validate the model and make our model selection
cv.lasso = cv.glmnet(x, y)
plot(cv.lasso)

# `glmnet` selected model coefficients
coef(cv.lasso); coef(cv.lasso, s = "lambda.min")

• Patients who:
  • are familiar with the importance of a regular medication regimen because of other conditions
  • have established an ongoing relationship with a primary care provider
  • and have incorporated this pattern into their daily routine (with years of repetition)

will likely demonstrate higher rates of medication adherence when starting a new medication

Slide 4 - Healthcare Spend is Rapidly Growing

Slide 5 - National Variation

Slide 7 - Wisconsin Health Information Organizatoin

glmnet - Regularization