page rank for anomaly detection

© Hortonworks Inc. 2015

PageRank for Anomaly Detection Hadoop Summit – SF Data Mining Meetup San Jose, 2015 Ofer Mendelevitch, Hortonworks


About Us

Ofer Mendelevitch Director, Data Science @ Hortonworks Previously: Nor1, Yahoo!, Risk Insight, Quiver blog: hHp://hortonworks.com/blog/author/ofermend/

Joint work with Jiwon Seo Ph.D Candidate @ Stanford SoNware Engineer @ Pinterest Designed SociaLite (w/ professor Monica Lam)


What is this talk about?

• Why is fraud detection important in healthcare? • The Medicare-B dataset

• Our approach: Similarity and PageRank

• Implementation: Apache Pig and SociaLite

• Some Results


Fraud prevention is important in healthcare

Recovery rates are still low, e.g., 3-4% Source: https://fullfact.org/wp-content/uploads/2014/03/The-Financial-Cost-of-Healthcare-Fraud-Report-2014-11.3.14a.pdf

$0

$500

$1,000

$1,500

$2,000

$2,500

US EU

$2,270

$940

$171

$71

Healthcare Expenditures (Billions)

Fraud

Non fraud


Example fraud cases in healthcare…

• A doctor billing too often for most expensive office visits

http://www.dallasnews.com/investigations/20140515-medicare-data-reveals-unusual-billing-patterns-by-nearly-80-texas-doctors-medical-practitioners.ece

• Medical supply stores paid off local doctors to prescribe motorized wheelchairs worth $7500 but instead provided scooters worth $1500

http://blog.operasolutions.com/bid/388511/Data-Science-As-the-Panacea-for-Healthcare-Fraud-Waste-and-Abuse


What are some fraud patterns?

• Billing for services that were not actually performed

• Performing unnecessary services • Using stolen patient IDs to submit claims • Unbundling: billing each stage of a procedure as if it is performed separately

• Upcoding: billing for more expensive services than were actually performed

• Billing cosmetic surgeries as necessary repairs • Etc…


Most healthcare providers have some type of system in place to identify such fraud • Rules based: – Business rules catch known fraud paHerns

• Machine-learning based: – Automated learning catches difficult to characterize fraud paHerns

• What are “good features” in the model that increase the accuracy? – Claim features, e.g. total amount – Provider features, e.g., total payment last year – Pa`ent features, e.g., current set of diagnoses


Why PageRank for fraud detection?

• Most approaches apply supervised learning – Graph algorithms not as widely-‐used

• The main idea: – Produce new “features” for the exis`ng model – Specifically, a score per provider reflec`ng its degree of anomaly rela`ve to a medical specialty


Our Dataset

• Medicare-B – real world public healthcare dataset – Released by CMS (US Centers for Medicare and Medicaid Services) in 2014 – Includes provider payment informa`on for 2012 – 9.5M records; 880K+ providers; 5616 CPT (procedure) codes

• We will only use 4 fields: – NPI: provider ID – Specialty: e.g. Internal Medicine, Den`st, etc – CPT code: medical procedure code – Count: # of procedures performed (normalized)


Example rows from the dataset 1003000126 ENKESHAFI ARDALAN M.D. M I 900 SETON DR CUMBERLAND 215021854 MD US Internal Medicine Y F99222 Initial hospital care 115 112 115 135.25 0 199 0 108.11565217 0.9005883395 1003000126 ENKESHAFI ARDALAN M.D. M I 900 SETON DR CUMBERLAND 215021854 MD US Internal Medicine Y F99223 Initial hospital care 93 88 93 198.59 0 291 9.5916630466 158.87 0 1003000134 CIBULL THOMAS L M.D. M I 2650 RIDGE AVE EVANSTON HOSPITAL EVANSTON 602011718 IL US Pathology Y F88304 Tissue exam by pathologist 226 207 209 11.64 0 115 0 8.9804424779 1.7203407716 1003000134 CIBULL THOMAS L M.D. M I 2650 RIDGE AVE EVANSTON HOSPITAL EVANSTON 602011718 IL US Pathology Y F88305 Tissue exam by pathologist 6070 3624 4416 37.729960461 0.0012569747 170 0 28.984504119 5.6268316462 1003000134 CIBULL THOMAS L M.D. M I 2650 RIDGE AVE EVANSTON HOSPITAL EVANSTON 602011718 IL US Pathology Y F88311 Decalcify tissue 13 13 13 12.7 0 39 0 7.8153846154 4.2806624494

We use only 4 fields: NPI, specialty, CPT code and count:

1003000126, Internal Medicine, Initial hospital care (F99222), 115 1003000126, Internal Medicine, Initial hospital care (F99223), 88 1003000134, Pathology, Tissue exam by pathologist (F88304), 209 1003000134, Pathology, Tissue exam by pathologist (F88305), 4416 1003000134, Pathology, Decalcify tissue (F88311), 13


Our approach – the steps

• Step 1: Data Preparation/cleansing

• Step 2: Compute similarities, build graph

• Step 3: Compute PageRank, identify anomalies


Step 1: Data cleansing

1003000126 ENKESHAFI ARDALAN M.D. M I 900 SETON DR CUMBERLAND 215021854 MD US Internal Medicine Y F99222 Iniàl hospital care 115 112 115 135.25 0 199 0 108.11565217 0.9005883395 1003000126 ENKESHAFI ARDALAN M.D. M I 900 SETON DR CUMBERLAND 215021854 MD US Internal Medicine Y F99223 Iniàl hospital care 93 88 93 198.59 0 291 9.5916630466 158.87 0 1003000134 CIBULL THOMAS L M.D. M I 2650 RIDGE AVE EVANSTON HOSPITAL EVANSTON 602011718 IL US Pathology Y F88304 Tissue exam by pathologist 226 207 209 11.64 0 115 0 8.9804424779 1.7203407716

10030126 Internal Medicine Iniàl care(F99222) 115 10030126 Internal Medicine Iniàl care(F99223) 88 10030134 Pathology Tissue exam(F88304) 209 Filter columns, data

cleansing

• Extract needed data fields from dataset – NPI (Naònal Provider ID), Specialty, CPT (procedure) code, count – For count, we chose: “bene_day_srvc_cnt” (number of dis`nct Medicare beneficiary per day services)

• Re-compute “specialty” due to data quality issues


Specialty Lookup: NPI and NUCC datasets

• Problem: – Some “specialty” values are inaccurate or not specific enough

• Solution: pre-processing step – NPI data: maps NPI to specialty code – NUCC data: maps specialty code to taxonomy


Step 2: build graph by similarities

10030126 Internal Medicine Ini`al care(F99222) 115 10030126 Internal Medicine Ini`al care(F99223) 88 10030134 Pathology Tissue exam(F88304) 209

• Two providers are “similar” if they have the same “procedure code patterns”

• We use “Cosine Similarity” – Each provider represented as vector of 5949 CPT codes


Example: similar providers

• NPI1

• NPI2

CPT 93042 99283 99284 99285 99291

Count 280 29 265 410 28

CPT 99283 99284 99285 99291

Count 118 151 270 37

CPT Descrip`on

93042 Rhythm Ecg report

99283 Emergency dept visit (1)



99291 Cri`cal care first hour


Computing similarity at large scale…

• Number of providers: ~880,000 • 880K * 880K = 77,440,000,000 similarity computations • Each one a “dot product” between vectors of length 5949 (but sparse)


How do we address scalability?

• Our Implementation: – Heuris`cs:

– Only compute similarity between NPI1 and NPI2 if they share their most important CPT codes

– Filter out NPIs with less than 3 CPT codes

– Use Apache PIG on a Hadoop cluster (with UDFs) to compute in parallal

• Alternatives: – DIM-‐SUM (map-‐reduce or Spark) – Locality Sensi`ve Hashing (DataFu)


PIG code: compute similarity GRP = group DATA by npi parallel 10; PTS = foreach GRP generate group as npi, DATA.(cpt_inx, count) as cpt_vec; PTS_TOP = foreach PTS generate npi, cpt_vec, FLATTEN(udfs.top_cpt(cpt_vec)) as (cpt_inx: int, count: int); PTS_TOP_CPT = foreach PTS_TOP generate npi, cpt_vec, cpt_inx; CPT_CLUST = foreach (group PTS_TOP_CPT by cpt_inx parallel 10) generate PTS_TOP_CPT.(npi, cpt_vec) as clust_bag; RANKED = RANK CPT_CLUST; ID_WITH_CLUST = foreach RANKED generate $0 as clust_id, clust_bag; ID_WITH_SMALL_CLUST = foreach ID_WITH_CLUST generate clust_id, FLATTEN(udfs.breakLargeBag(clust_bag, 2000)) as clust_bag; ID_WITH_SMALL_CLUST_RAND = foreach ID_WITH_SMALL_CLUST generate clust_id, clust_bag, RANDOM() as r; ID_WITH_SMALL_CLUST_SHUF = foreach (GROUP ID_WITH_SMALL_CLUST_RAND by r parallel 240)

generate FLATTEN($1) as (clust_id, clust_bag, r); NPI_AND_CLUST_ID = foreach ID_WITH_CLUST generate FLATTEN(clust_bag) as (npi: int, cpt_vec), clust_id; CLUST_JOINED = join ID_WITH_SMALL_CLUST_SHUF by clust_id, NPI_AND_CLUST_ID by clust_id using 'replicated'; PAIRS = foreach CLUST_JOINED generate npi as npi1, FLATTEN(udfs.similarNpi(npi, cpt_vec, clust_bag, 0.85)) as npi2; OUT = distinct PAIRS parallel 20;

Things to highlight: • Using “replicated” joins (map-side joins) where possible • Handling Data Skew • Using Python UDFs to compute similarity, break large bags, etc


Step 3: Personalized PageRank

Run Personalized PageRank with SociaLite

• Compute specialty-centric “Personalized PageRank” for each node (provider)

• Anomaly candidate: high score but wrong specialty

0.025

0.3 0.092

0.095

0.15

0.2

0.002

0.005

0.02

0.01

0.012

0.2


PageRank – a quick overview •  Random walk over the graph •  Start from any (randomly selected)

node •  At each step, walker can:

– Move to an adjacent node (probability d = 80%) –  Randomly jump (or “teleport”) to any node in the graph

(probability 1-‐d = 20%)

All doctor names are fictitious

Dr. Miller 9.6 %

Dr. Jones 9.6 %

Dr. Lam 9.6 %

Dr. Ng 12.5 %

Dr. Cheng 6.7 %

Dr. Das 12.0 %

Dr. Seo 9.2 %

Dr. Page 6.6 %

Dr. Ortega 12.0 % Dr. Padian

12.1 %


Personalized PageRank Focused on a given specialty

•  Random walk over the graph •  Start from any random node IN THE

SPECIALTY GROUP •  At each step, walker can:

– Move to an adjacent node (probability d = 80%) –  Randomly jump (or “teleport”) to any node OF THE GIVEN SPECIALTY GROUP



Dr. Miller 1.6 %

Dr. Jones 1.6 %

Dr. Lam 1.6 %

Dr. Ng 3.3 %

Dr. Cheng 4.6 %

Dr. Das 20.7 %

Dr. Seo 15.7 %

Dr. Page 11.8 %

Dr. Ortega 20.7 %

Dr. Padian 18.2 %


Personalized PageRank Focused on a given specialty

•  Random walk over the graph •  Start from any random node IN THE

SPECIALTY GROUP •  At each step, walker can:

– Move to an adjacent node (probability d = 80%) –  Randomly jump (or “teleport”) to any node OF THE GIVEN SPECIALTY GROUP



Dr. Miller 20.1 %

Dr. Jones 20.1 %

Dr. Lam 20.1 %

Dr. Ng 23.2 %

Dr. Cheng 5.8 %

Dr. Das 2.2 %

Dr. Seo 1.7 %

Dr. Page 0.9 %

Dr. Ortega 2.2 %

Dr. Padian 3.9 %


Personalized PageRank with SociaLite

`Rank(int npi:0..$MAX_NPI_ID, int i:iter, float rank).` `Rank(source_npi, 0, pr) :- Source(source_npi), pr=1.0f/$N.` for i in range(10): `Rank(node, $i+1, $sum(pr)) :- Source(node), pr = 0.2f*1.0f/$N ; :- Rank(src, $i, pr1), pr1>1e-8, EdgeCnt(src, cnt), pr = 0.8f*pr1/cnt, Graph(src, node).`

Initialize PageRank value of source providers to be 1/N In each iteration: •  Teleport to source providers (w/ probability 0.2) ; •  Random walk to one of neighbors (w/ probability 0.8)


What’s so cool about SociaLite?

• PageRank in 3 lines of code

• Python integration

• You don’t have to “think like a node”. Declarative language – “looks like” the formula


Using the PageRank scores?

Rules Fraud Model Claim Generate

Features

PageRank Scores

Decision

Provider Patient Amount Date, time Etc…

Paènt informaòn

Provider Informaòn Etc…

Feature 1 Feature 2 … Feature N

PR Feature 1 PR Feature 2 … PR feature M


Example result #1: Ophthalmology

Found internist with high score, but these CPT codes: •  Internal eye photography •  Cmptr ophth img optic nerve •  Echo exam of eye thickness •  Cptr ophth dx img post segmt •  Revise eyelashes •  Ophthalmic biometry •  Eye exam new patient •  Eye exam established pat •  After cataract laser surgery •  Eye exam & treatment •  Eye exam with photos •  Cataract surg w/iol 1 stage •  Visual field examination(s)


Example result #2: Plastic Surgery

Found Otolaryngologist with high score, but these CPT codes:

•  Skin tissue rearrangement (multiple variants) •  Biopsy skin lesion


Thank you!

Any Ques`ons?

Ofer Mendelevitch, [email protected], @ofermend Code available here: hHps://github.com/ofermend/medicare-‐demo/ Blog post series: hHp://hortonworks.com/blog/using-‐pagerank-‐detect-‐anomalies-‐fraud-‐healthcare/

page rank for anomaly detection

Data & Analytics

fraud patterns

fraud prevention

fraud rules

fraud detection important

example fraud cases

healthcare providers

doctor billing

medicareb dataset