Using machine learning to unlock value across the healthcare value chain

• ML is a set of techniques that can identify patterns in data to make informed predictions or locate anomalies

• ML can generate value through enhanced physician decision making, reduced waste and fraud, and improved administrative efficiency

• ML is not a magic wand—it needs expert monitoring to maximize model impact

Machine learning (ML) 101

Challenges Potential organizational solutions

Payers and providers need to overcome a few challenges to best leverage ML techniques

Lack of artificial intelligence (AI) strategy

• Investor-like approach to funding AI use cases• Dedicated AI leader with strong Analytics Interpreter

bench

Fragmented data storage and tech limitations

• Shared, inexpensive computing resources • Investment in data infrastructure and storage systems• Strong data governance for quality and traceability

Tight talent market

• Home-grown data science and translator talent• Shared service model across the organization for

AI/ML

Disbelief in AI’s potential

• “Quick-win” use cases leveraging existing datasets

To discover more insights from

McKinsey’s Healthcare Analytics & Delivery, visit

https://healthcare.mckinsey.com/topic/healthcare-analytics

SOURCE: CMS 2016 US healthcare expenditure estimate; proprietary McKinsey models; team analysis

Potential use cases across healthcare

Inform clinical decisions on effective care management and preventive care solutions

At-risk patient prediction

Help primary care physicians identify optimal specialists based on patient needs and preferences

Patient referral

Analyze long-term patient / member journeys through learned prediction of preferences and clinical events (e.g., use individual clinical data to predict preventable catastrophic events)

Improved patient / member retention

Payment services streamlining

Improve claim accuracy and reduce administrative inefficiencies (e.g., during claims submission and processing)

Payers and providers could reduce fraud,

waste, and abuse

(FWA) by $20-30Bn from the

US healthcare value chain by implementing

ML

1 Selected portion of algorithm presented (from perspective of Payer claims editing system)CPT: Current Procedural Terminology (published by the American Medical Association); PTP: Procedure to Procedure; HCPCS: Healthcare Common Procedure Coding System; E&M: Evaluation and Management

Case example: decision model for FWA detection

HCPCS Q0091 (Pap smear)

NO YES

CPT 99395/99396on same claim

CPT 99395/99396on same claim

YES NO YES

Recover

NO

Do not recover … more rulesCPT 36415On the same claim

Do not recover

Is E&M procedure on the same claim

… more rules Global surgery 090 on same day by same provider

NO YES

RecoverDo not recover

Global surgery 090 on same day by same provider

YES

Recover

NO

Do not recover

YES NO NO

NO PTP edit present in claim with no

modifier allowed YES

YES

Modifier 59 Modifier 59

HCPCS G0101on same claimDo not recoverYES NO

NO

Do not recover

RecoverDo not recoverGlobal surgery 090 in next 3

days at same hospitalIs PPO claim

YES NO YES NO

YES

HCPCS G0101on same claim

YES

Recover

NO

Do not recover

Using ML techniques, we designed an algorithm to identify claim recovery cases for unbundling of services – the snapshot below describes an ML algorithm that can be converted into rules to implement in a payer claims editing system1

Supervised learning (e.g., ML classifiers such as neural nets, random forests): Algorithms use training data and feedback from experts to identify relationships from a set of inputs to a given output

Unsupervised learning (e.g., K-means clustering, hierarchical clustering): Algorithms explore input data and identify patterns of behavior. This helps in cohort analysis and anomaly detection

Machine learning techniques

… more rules

YES NO