Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

LeAnna Kent - Using Network Analysis to Detect Kickback Schemes Among Medical Providers

157 views

Published on

Using Network Analysis to Detect Kickback Schemes Among Medical Providers

Published in: Technology
  • Be the first to comment

  • Be the first to like this

LeAnna Kent - Using Network Analysis to Detect Kickback Schemes Among Medical Providers

  1. 1. Detecting Kickback Schemes Using Network Analysis LeAnna Kent 1
  2. 2. About Me | 2
  3. 3. Challenge Estimated 20% of all medical fraud is attributed to kickback schemes† Time and cost of investigations means we don’t have sufficient known cases for supervised modelling | 3 † https://www.gao.gov/assets/680/674771.pdf
  4. 4. | 4 What is a Kickback? • Something of value being offered in return for a favorable action • In health care: • Referrals • Prescribing the use of specific drugs/equipment • As a result: • A large percentage of a doctor’s patients will see the same doctor and/or receive the same drug or equipment
  5. 5. | 5 What are Networks? • Represent relationships between entities • Start with nodes (doctors) • Connect nodes based on associations (shared patients)
  6. 6. Connecting Providers | 6 𝐽 𝐴, 𝐵 = 𝐴 ∩ 𝐵 𝐴 ∪ 𝐵 = 𝐴 ∩ 𝐵 𝐴 + 𝐵 − 𝐴 ∩ 𝐵
  7. 7. Connecting Providers | 7 𝐽 𝐴, 𝐵 = 𝐴 ∩ 𝐵 𝐴 ∪ 𝐵 = 𝐴 ∩ 𝐵 𝐴 + 𝐵 − 𝐴 ∩ 𝐵 +
  8. 8. | 8 Analyzing the Network
  9. 9. | 9 Analyzing the Network
  10. 10. | 10 Analyzing an Egonet
  11. 11. | 11 Analyzing an Egonet
  12. 12. Subgraph Strength Formula | 12 𝑆 𝑔 𝑛 = 𝑤 − ∆ 𝑛 𝑛 𝑤 𝑛 𝑝 𝑤 𝑛 weight of alter node 𝑛 𝑤 avg weight of alter nodes ∆ 𝑛 change in 𝑤 𝑛 alter nodes used 𝑝 maximum possible edges 𝑤 𝑛 weight of alter node 𝑛 𝑛 alter nodes used 𝑝 maximum possible edges 𝑆 𝑔 𝑛 = 𝑤 − ∆ 𝑛 𝑛 𝑤 𝑛 𝑝 ∆ 𝑛 change in 𝑤 𝑤 avg weight of alter nodes 𝑆 𝑔 𝑛 = 𝑤 − ∆ 𝑛 𝑛 𝑤 𝑛 𝑝 𝑆 𝑔 𝑛 = 𝑤 − ∆ 𝑛 𝑛 𝑤 𝑛 𝑝 𝑆 𝑔 𝑛 = 𝑤 − ∆ 𝑛 𝑛 𝑤 𝑛 𝑝
  13. 13. Example | 13
  14. 14. Algorithm Results | 14 Original Graph: Resulting Graph: • 6k nodes • 600k edges • Median degree of 60 • 1 connected component • 6k edges • 93% stopped after first node • Maximum degree of 16 • 1,316 connected components Run Time: < 2 min
  15. 15. Interpreting Results Each egonet’s subgraph now has a score, by which they can be ranked | 15
  16. 16. Interpreting Results Each egonet’s subgraph now has a score, by which they can be ranked When providing ranked list, remove any perfect subgraphs | 16 27% of nodes’ resulting egonets were redundant
  17. 17. Interpreting Results Each egonet’s subgraph now has a score, by which they can be ranked When providing ranked list, remove any perfect subgraphs Can pair with supplemental information such as patients or dollars involved | 17
  18. 18. Conclusions • Built graph where nodes were physicians, and edges were built based on shared patients and weighted by the Jaccard index • Focused on the egonet of each node independently • Applied algorithm to select an egonet’s subgraph with the relative strongest edges • Ranked subgraphs for investigation based on subgraph strength score | 18

×