Inference of Product Quality by using RFID-enabled Traceability Information
A study on the US pharmaceutical supply chain
Tatsuya INABAKeio University/Auto-ID Lab. Japan
IEEE RFID 2009@Orlando
May 2, 2009Page 2
Agenda
• Background
• Requirements of the model
• Case study : the US pharmaceutical Industry– Why US pharma supply chain?– Model– Evaluation– Observation
• Discussion and Future research
May 2, 2009Page 3
Background
• Trend companies provide traceability info to consumer– Simple effect
• Products w/ traceability info > Products w/out traceability info
• Consequences of too much (raw) traceability info– Too complex to understand– As a result, avoid “intuitively” low quality products (e.g., long
travel history)– Cause problem on SCM practice
• Current premise: products on shelf are sold• Problem: products with long travel history may stay on the shelf
May 2, 2009Page 4
Background (cont’d)
• New problem under fully traceable world!– Try to promote RFID to make SCM more efficient.
BUT…unwanted side effects??
• Necessity of model to provide a simple index– Out from complex traceability information– Easy to understand quality and easy to choose products
• Enable companies to operate supply chain as it is– Not item-level inventory management but SKU-level inventory
management
May 2, 2009Page 5
Requirements of the Model
• Incorporate elements that affect product quality– Company names, conditions, IDs, contractual info
• Stochastic, Not deterministic– Causal relationship is known deterministic, but not in this case– Elements stochastically impact quality of products
• E.g., Difference b/w quality of fruits shipped by air and by sea
Quality Index = f (names of companies, conditions of storage,IDs of other products, contractual information, . . .)
Use Bayesian Network as a stochastic model and apply the model to the US pharmaceutical supply chain
May 2, 2009Page 6
Bayesian Network (BN)
• BN is a stochastic model with acyclic graph– Consists of variables (nodes) and arrows that connect variables
(parent node to child node)– Direction of arrow show casual relationship– The extent of impact from a parent node to a child node is
expressed as a set of probabilities called Conditional Probability Table (CPT)
C D
EB0.8 0.2
b
eb
0.6 0.4
0.01 0.990.2 0.8
ebee
b
EB P(A | B, E)
A
CBT
May 2, 2009Page 7
Why US Pharma Supply Chain?
• Example of fully traceable world
• Item level tagging to prevent drug counterfeiting– Studies regarding ILT to prevent counterfeiting by government
and industry– Studies to mandate exchanging drug traceability information
(called pedigree) in some states– Virtually all the individual prescription drug traceability info is
available in those states
May 2, 2009Page 8
Why US Pharma Supply Chain? (cont’d)
• Example of complex traceability information
• Complex drug supply chain– Tier-structured wholesaler– Three 1st tier wholesalers buy drugs from manufactures and 2nd
tier wholesales, the same trade in 2nd tier wholesalers – Repackaging process makes the supply chain more complex– Companies avoid to buy long travel history drugs (intuitive
purchase decision!)
May 2, 2009Page 9
Illustration of US Pharma Supply Chain
• Complicated layered trade of the industry
ManufacturerManufacturer
Tier-1 wholesaler
Tier-1 wholesaler
Tier-2 wholesaler
Tier-2 wholesaler
Tier-3 wholesaler
Tier-3 wholesaler
RetailerRetailer
Repackaging process makes the situation more complex…
May 2, 2009Page 10
Model: Elements to Affect Quality
• From current counterfeit cases– Involvement of 2nd and 3rd tier wholesalers– Involvement of small repackagers
• From pedigree doc exchange process assessment– Electronically exchange pedigree w/exception
• Small retail pharmacies are allowed to use paper-based pedigree– Product return from paper-based pedigree retail pharmacies
• Malicious parties tamper paper based pedigree and return fake drug with tampered pedigree
May 2, 2009Page 11
Model: Development of BN Graph
• Selection of nodes– Nodes are selected based on observed important characteristics
of real drug trades
• Graph development– Graph is developed based on the real counterfeit cases from
both government and industry reports
Repackage Paper-based pedigree
Tier-2Wholesaler
Tier-3Wholesaler
Return(paper-based)
Return(non paper-based)
Product Fraud
May 2, 2009Page 12
Evaluation
• Simulation with a hypothetical scenario– Real traceability data is not available– Traceability data is generated from a hypothetical scenario– Data of fraudulent products is included based on the counterfeit
reports
• Two sets of data for CPT and evaluation– A set of traceability data is needed for CPT calculation– The other set of traceability data for evaluation
May 2, 2009Page 13
Hypothetical Scenario for Evaluation
• Fraudulent cases assumed in the evaluation– Pattern A: Through repackage process by tier-2/3 repackager– Pattern B: Through return process of retailers that use paper-based
pedigree– Pattern C: Through tier-3 wholesaler handling
Pattern CPattern C
ManufacturerManufacturer
Tier-1 repackager
Tier-1 repackager
Tier-2/3 repackager
Tier-2/3 repackager
Tier-1 wholesaler
Tier-1 wholesaler
Tier-2 wholesaler
Tier-2 wholesaler
Tier-3 wholesaler
Tier-3 wholesaler
RetailerRetailer
Retailer(Paper-pedigree)
Retailer(Paper-pedigree)
RetailerRetailer
Pattern APattern A
Pattern BPattern B
May 2, 2009Page 14
0
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Simulation Result (Pattern A)
• Inferred probability ranges 6 – 99 %• Fake drug’s highest authenticity probability is less than 90%• “Products whose authenticity level is above 90 % is authentic.”
– Percentage of false positive (Judge fake drug as authentic): 0%– Percentage of false negative (Judge authentic drug as fake): 2.4 %
Inferred authenticity
Authentic
90 %
Fraudulent
May 2, 2009Page 15
0
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Simulation Result (Pattern B)
• Inferred probability ranges 6 – 99 %• Fake drug’s highest authenticity probability is less than 70%• “Products whose authenticity level is above 70 % is authentic”
– Percentage of false positive (Judge fake drug as authentic): 0%– Percentage of false negative (Judge authentic drug as fake): 1.1 %
70 %
Inferred authenticity
Authentic
Fraudulent
May 2, 2009Page 16
0
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Simulation Result (Pattern C)
• Inferred probability ranges 6 – 99 %• Fake drug’s highest authenticity probability is less than 80%• “Products whose authenticity level is above 80 % is authentic”
– Percentage of false positive (Judge fake drug as authentic): 0%– Percentage of false negative (Judge authentic drug as fake): 2.8 %
80 %
Inferred authenticity
Authentic
Fraudulent
May 2, 2009Page 17
Observation
• Benefit of the model– Successfully infer the authenticity level– Enable to decide threshold with zero % false positive and small
% false negative• Need an operation to check the drug authenticity with other method
whose authenticity level is below threshold (e.g., ingredient analysis)
• Beneficial to both consumers and companies– No need to be selective as far as inferred probability is above a
certain level– Companies could run their supply chain with SKU level
granularity
May 2, 2009Page 18
Discussion and Future Research
• Effectiveness of the stochastic model– Be useful for consumers when choosing products– Help companies to run their supply chain with product traceability
info
• Future research– Approach would be reasonable but operation needed to carefully
designed to provide product quality indices• Model• Data to be used• Patterns to infer the product quality
– May need regulation and standards for quality indices providing process
May 2, 2009Page 19
Summary
• In this research…
– Identify a possible problem that is caused by increasing traceability information to both consumers and companies
– Propose and evaluate a model by using Bayesian Network to provide a simple quality index out from traceability information so that consumers can judge product quality with the simple index not traceability information
May 2, 2009Page 20
References
[1] M. Roberti, “EPC reduces out-of-stocks at Wal-Mart,” 2005[2] RFID Journal, “Tesco deploys Class 1 EPC tags,” 2005[3] R. Koh, E. W. Schuster, I. Chackrabarti, and A. Bellman, “Securing the pharmaceutical supply
chain,” Auto-ID Labs White Paper Series, 2003[4] P. Jones, “Networked RFID for use in the food chain,” in ETFA ’06. IEEE Conference on
Emerging Technologies and Factory Automation, 2006., 2006[5] B. Patkai, L. Theodorou, D. McFarlane, and K. Schmidt, “Requirements for RFID-based
sensor integration in landing gear monitoring - a case study,” Auto-ID Labs white paper 2007[6] M. Ogawa and M. Umejima, “Traceability system for disclosure – a case of consumer
empowerment strategy,” Auto-ID Labs White Paper Series, 2005[7] United States Food and Drug Administration, Combating anticounterfeit drugs A Report of the
Food and Drug Administration, 2005[8] S. Russell and P. Norvig, Artificial Intelligence : A Modern Approach, 2002[9] United States Department of Health and Human Service, Industry: Development &
Distribution, Stakeholder Meeting, 2004[10] EPCglobal Inc., Pedigree Ratified Standard, 2007[11] Norsys Corporation, http://norsys.com/[12] The United States Food and Drug Administration Med Watch[13] Pfizer, http://www.pfizer.com/[14] Eli Lilly, http://www.lilly.com/[15] CVS Pharmacy, http://www.cvs.com/.
May 2, 2009Page 21
Questions