Visualizing nationwide variation in medicare Part D prescribing patterns

Abstract

Background: To characterize the regional and national variation in prescribing patterns in the Medicare Part D program using dimensional reduction visualization methods.

Methods: Using publicly available Medicare Part D claims data, we identified and visualized regional and national provider prescribing profile variation with unsupervised clustering and t-distributed stochastic neighbor embedding (t-SNE) dimensional reduction techniques. Additionally, we examined differences between regionally representative prescribing patterns for major metropolitan areas.

Results: Distributions of prescribing volume and medication diversity were highly skewed among over 800,000 Medicare Part D providers. Medical specialties had characteristic prescribing patterns. Although the number of Medicare providers in each state was highly correlated with the number of Medicare Part D enrollees, some states were enriched for providers with > 10,000 prescription claims annually. Dimension-reduction, hierarchical clustering and t-SNE visualization of drug- or drug-class prescribing patterns revealed that providers cluster strongly based on specialty and sub-specialty, with large regional variations in prescribing patterns. Major metropolitan areas had distinct prescribing patterns that tended to group by major geographical divisions.

Conclusions: This work demonstrates that unsupervised clustering, dimension-reduction and t-SNE visualization can be used to analyze and visualize variation in provider prescribing patterns on a national level across thousands of medications, revealing substantial prescribing variation both between and within specialties, regionally, and between major metropolitan areas. These methods offer an alternative system-wide and pattern-centric view of such data for hypothesis generation, visualization, and pattern identification.

Keywords: Healthcare variation; Machine learning; Medicare; Prescribing; t-SNE.

Fig. 1

Overall features of 2013 Medicare Part D prescribing patterns data set. a. Distribution of percentage of providers prescribing each of 2892 unique drugs, sorted by percentage of providers prescribing. b. Same as A except for 197 unique drug classes. c. Distribution of number of claims for each of 2892 unique drugs, sorted by number of claims. Note that the unique drug order is not necessarily the same as in a. d. Same as b except for 197 unique drug classes. e. Distribution of drug prescription diversity across all providers sorted by number of unique claims. Numbers of providers prescribing more than 100 and 300 unique drugs are annotated on plot. f. Distribution of number of claims across all providers sorted by claims per provider. Number of providers making more than 10,000 and 25,000 claims are annotated on plot. G = Gini index

Fig. 2

Distribution of Medicare Part D providers across states. a. Share of providers by state (as a percentage of the total number of providers) plotted against share of Medicare Part D enrollees by state (as a percentage of the total number of enrollees nationwide) are shown by black circles and fit to a line (gray dashed line); green line is slope of one. A similar plot based on a data subset of high-claims providers (> 25,000 claims resulting in 2062 providers) is shown superimposed as open triangles colored by their relation to the corresponding data from the full data set. Some states are annotated. b. Comparison of the provider composition by state for the full data set (left) and the high-claims data set (right). Ribbons connecting the two join corresponding states

Fig. 3

Low-dimension embedding of providers using t-SNE and PCA. 2-D density plots in low dimensional space created using t-SNE (upper) or PCA (lower) of 227,573 Medicare Part D providers, each with ≥ 1000 prescription claims in 2013 organized by a the 227,573×2791 drug claims matrix or b the 227,573×195 drug class claims matrix. Number-of-claims data per provider by drug or drug class is scaled by the total claims per provider to express the prescribing pattern as a composition prior to t-SNE

Fig. 4

Array of t-SNE plots each highlighting providers of a specific specialty. Each 2-D density plot (grey) is the same as shown in Additional file 5: Figure S4A, and represents the set of 227,573 Medicare Part D providers ×2791 drug claims. Included providers had ≥ 1000 prescription claims in 2013. The plot is a heatmap, with densities representing increased numbers of providers. Provider specialties are shown in red to emphasize their collocation by prescribing pattern, and are labeled by NPPES self-reported specialty designation. Note the separation of provider clusters, even to the extent that subspecialties (annotated in blue) are distinguishable within the specialty cluster (e.g. Cardiology and Cardiac Electrophysiology

Fig. 5

Representative prescribing patterns corresponding to different regions of t-SNE plot. Left: t-SNE plot as shown in Additional file 5: Figure S4A with 20 different regions labeled as A through T. Right: Heat map showing prescribing patterns. Columns are individual providers, 10 randomly selected from each of the 20 regions. Each row represents a drug. The drugs shown are the union of the top eight most frequently prescribed in each region. Increasing gray density corresponds to the percent of claims for a particular drug made by a provider relative to their total claims, with white denoting no claims. Prescribing volume (total claims) and diversity (number of unique drugs prescribed) are shown above the heat map as bar graphs. Note region N, which is enriched for providers with a high volume of opioid analgesic claims

Fig. 6

Array of t-SNE plots of providers annotated for fraction of claims for each of eight heart/circulation related drugs. The t-SNE plots were created from the set of 227,573 Medicare Part D providers ×2791 drug claims. Included providers had ≥ 1000 prescription claims in 2013. The color for each provider corresponds to the percentage of claims for the indicated drug relative to the provider’s total claims. Gray is 0%, the maximum scale (red) is 15% of total claims. Note the high volume of prescriptions within within both the cardiology and internal medicine areas

Fig. 7

Array of t-SNE plots each highlighting providers of a specific specialty. These t-SNE plots are derived from the dataset of 227,573 Medicare Part D providers ×195 drug classes. Even with dimension reduction from 2791 individual medications to 195 medication classes, t-SNE plots produced clear groupings of specialties and subspecialties. This plot removes potential bias introduced by prescribing of generic versus brand name medications, and thus is a better representation of prescribing variation across specialties due to patient populations and practice patterns

Fig. 8

Array of t-SNE plots of providers annotated for fractions of claims for each of six cardiac drug classes The t-SNE plot layout was generated using the dataset of 227,573 Medicare Part D providers ×195 drug classes. The 195 drug classes include all medications (generic and brand name) collapsed into the the indicated class. The color for each provider corresponds to the percentage of claims for the indicated drug relative to the provider’s total claims. Gray is 0%, the maximum scale (red) is 15%. This dimension reduction and visual representation eliminates differences due to formulary, or generic versus brand name medication prescribing patterns. Note, for example, the high percentage (red areas) of beta blockers prescribed in cardiology and nephrology (oral preparations) and opthomology (eye drops)

Fig. 9

Unsupervised hierarchical clustering by drug class. Provider clusters obtained by hierarchical agglomerative clustering using a Euclidean distance measure and centroid criteria. a) Cumulative distribution of provider size over 605 clusters. b) Provider specialties within each cluster were tallied and the number of providers in the dominant specialty plotted against cluster size. The lines indicates where 100% (red), or 30% (gray) of providers in the cluster are the same medical specialty. c) t-SNE visualization of provider prescribing pattern variation for Family Medicine providers by United States Federal Region. Each plot represents a 2D density histogram

Fig. 10

Distribution of provider prescribing patterns by census region. Providers with ≥ 1000 claims (n=227,573) were divided into subsets by census region (lower figures within regional pairs). For comparison, a random sample of equivalent size was taken from the entire data set such that the providers in each random sub-sample did not overlap with any of the others (upper figures). This allows visual comparison of regional provider distributions with a random national sample of equivalent size

Fig. 11

Variation of prescribing pattern by core-based statistical areas. a. Multidimensional scaling (MDS) of 52 CBSAs based on 532 drugs that have over 100,000 claims (across 50 states and Washington DC). Data were expressed as number of claims for a particular drug in a particular CBSA per number of enrollees in that CBSA. CBSAs are specified by IATA airport code. Magenta dots indicate MDS performed on a randomly permuted data sets where the data corresponding to the CBSA providers were shuffled, preserving the number of providers for each CBSA. b. Comparison of two CBSAs of similar sizes: Oklahoma City OK vs. Rochester NY. Dots represent individual drugs and axes are the number of claims per enrollee in log scale (for the respective CBSAs). Dashed lines indicate 5-fold differences in the per-enrollee numbers of claims. Drugs beyond these regions are indicated. c. Comparison of Houston TX and Dallas-Fort Worth Texas CBSAs that might be expected to have similar profiles as an internal control. d. MDS plot of 52 CBSAs based on 198 drug categories, similar to part A. e. Comparison of prescribing patterns in Boston MA and Miami FL based on drug categories. f. Houston TX vs. Dallas-Fort Worth TX based on drug categories