How Azure and Databricks Enabled a Personalized Experience for Customers and Patients at CVS Health
2 likes930 views
CVS Health is focusing on enhancing customer personalization through data-driven insights to improve pharmacy and retail experiences. The challenges include adapting to diverse customer behaviors, HIPAA compliance, and leveraging technology for quick deployment of personalized services. Despite facing issues like data management, talent acquisition, and scalability, CVS is pursuing advanced machine learning techniques to refine its personalization strategies.
How Azure and Databricks Enabled a Personalized Experience for Customers and Patients at CVS Health
- 1. Bringing Personalization to Customers and Patients at CVS Health Michelle Un Raghu Nakka
- 2. Agenda Personalization at CVS Growth & Challenges Looking to the Future
- 3. CVS Health at a Glance 9,990+ retail locations 62M+ ExtraCare card members 2.5B prescriptions managed or filled 4.5M customers served by CVS Pharmacy stores daily Source: https://cvshealth.com/about/facts-and-company-information
- 4. Understanding CVS customer behavior Unique Customers Unique Data Unique Situations • Different types of customers means large number of micro-segments which poses challenges to understand overall behavior of customer • Pharmacy patient health needs are varied and sometimes unpredictable • Not a typical grocery store customer • Hard to predict customer behavior of a convenience shopper • Data sparsity and dimensionality leads into overfitting issues • Need to be able to respond to evolving retail and healthcare industries • Limited ways to influence or incentivize patient health behavior within the pharmacy setting • Situations like COVID-19 will instantly make static machine learning models either outdated or invalid Strict Data Use and Confidentiality Agreement o Due to patient confidentiality concerns and HIPAA restrictions, Front Store and Pharmacy data and Personalization efforts are kept completely separate o All use cases are reviewed by legal team to ensure conformance with laws and regulations o The team does share skills / technical expertise across separate environment to accelerate learning
- 6. Personalization at CVS Personalization is delivering the right individual experience in the right channel at the right time The Right Experience... In the Right Channel... At the Right Time... Tailored for Each Individual Customers Product Sales & Rewards Service Loyalty & Engagement Mobile Web In StoreReceipt Call Center Location Time Weather Schedule
- 7. Technology to support Personalization To move quickly, we leveraged existing technology where suitable and used the cloud to enable an agile analytics capability Structured Data Sources Files from Data Sources Data Warehouses Batch Mode Existing Sources Data Ingestion Data Layer Features Opportunities Model Offer Delivery Processing Layers Control-M Industrialized ‘Next-Best-Action’ Pipeline DevOps Orchestration Analytics and Compute DAG Measurements and Insights Microsoft Azure Test and Learning Experimentation
- 8. The Personalization Journey From idea to in-market in less than 6 months, but with constant need to pivot to meet future needs Built pilot in on-prem Hadoop environment 20192018 Scaled personalized offers to 100% Migrated from on-prem to Azure Databricks to enable scale-up 2020 Launched first personalized offers to 1% customers Reassess technology for future growth and use cases Scaled test and learn experimentation framework Introduced test and learn experimentation and doubled the number of personalized products Personalization journey begins
- 9. The Personalization Journey From idea to in-market in less than 6 months, but with constant need to pivot to meet future needs Built pilot in on-prem Hadoop environment 20192018 Scaled personalized offers to 100% Migrated from on-prem to Azure Databricks to enable scale-up 2020 Launched first personalized offers to 1% customers Reassess technology for future growth and use cases Scaled test and learn experimentation framework Introduced test and learn experimentation and doubled the number of personalized products Personalization journey begins Team Size
- 10. Impact of Personalization There are many use cases enabled through Personalization, which help to enhance customer experience, improve health outcomes, and drive value for CVS Health Example Use Case Desired ImpactExample Solution Determine the best clinical product to offer a patient given unique health needs • Increase prescription refill rate • Reduce patient gaps in therapy • Improve condition-level adherence Determine the best channel to send a clinical product to a patient Predict patient's propensity to accept an offer given outreach based off of patient behavior • Increase responsiveness to channel • Increase adoption of clinical products • Spend pharmacy time on only most critical clinical outreaches Test random assignment of channel, collect data, then optimize channel assignment using experimentation or modeling approach Present most relevant and efficient coupon offers to ExtraCare members Create microsegments based off of customer behavior to provide most relevant offers • Increase engagement of active customers • Re-engage with lapsed customers Improve interpretation of patient's responses Analyze past responses and improve interpretation through rule-based engine; explore and leverage natural language processing for real-time interpretation • Increase responsiveness to channel • Increase adoption of clinical products • Improve patient experience
- 11. Example Use Case: Pharmacy Problem • One barrier to medication adherence that many patients face is picking up medication on a regular basis, due to a variety of barriers, such as forgetfulness, cost, access, side effects, etc. Example Solution • Understand medication adherence profile and barriers to medication adherence • Use machine learning to determine what products and services may help a patient based on their specific needs (i.e. automated refills, reminders and notifications) • Use test and learning experimentation to determine the best channel, timing, and messaging to promote and package the service to the patient Offer automatic refill service to patient Offer the service to the patient through a SMS Emphasize the convenience offered by the service in the messaging
- 12. Example Use Case: Retail Front Store CouponOffer at Register Coupon influences the next store trip Presence of offer expands basket size during the trip Problem • Growing the engagement of most valuable customers. Re-engage lapsed customers and engage all active product shoppers Example Solution • Personalize communication, provide relevant and exciting offers using data from customer profile
- 13. Results driven through Personalization Data and insights driven strategy in retail pharmacy help to improve to medication adherence Pharmacy RESULTS 1.6% overall improvement in adherence PREDICTION Optimized Offers Refill rate reminder timing Preferred messaging/ effectiveness Prescriber outreach effectiveness KEY PROFILE ELEMENTS Adherence likelihood Medication profile Cost profile and cost sensitivity Offer and response history Prescriber relationship Source: Enterprise Analytics, 2019; Image Source: CVS Health Creative Resource Library, accessed 2019
- 15. Enabling Speed to Market Cloud-based environment enabled us to get to market quickly, expand the platform with the growth of the team, and rapidly introduce new improvements to the application Ease of Use Collaboration • More flexibility to spin up and down different types of clusters to meet emerging project needs and changing team size • Not restricted to physical hardware constraints • Less focus on Spark tuning due to performance and autoscaling features • Databricks workspace centralizes assets to make development easier, including interactive notebook, cluster management and performance monitoring, metastore • Less worrying about infrastructure support and more focus on analytics • Easy to integrate new Azure services or supported services into our environment • Easy to onboard new users and teams to integrated workspace environment through Databricks • Easy to collaborate using Databricks notebooks 3Ease of Use2Scalability1
- 16. Growth & Performance Challenges With quick growth, we encountered many challenges along the way, many of which we continue to tackle Handling Big Data Cost Management Evolving Technology • Size of data causes breakdown of production pipeline while adding new features • Growing team has varying exposure to Spark so use cases outpace optimization • Retail and healthcare industry not evolving as fast as cloud technology industry • New solutions need to be developed within the constraints of the retail and healthcare industry (i.e. security, compliance, archival regulations) • Different clusters for ETL, model training, and feature creation • Code optimization and sampled data for more efficient cluster usage • Utilize features like cluster policies, pools, etc. Finding Talent • Difficult to find cross-functional skillsets with data science, engineering, and DevOps background • New / changing technology means limited pool of talent • Hiring talent with Python helped bridge the gap between data science and engineering • Increase training and opportunities to explore new technology to grow the skillset • Increasing size of team leads to high cost • Cost visibility can be fragmented and difficult to monitor cost • Some optimization from Delta, but comes with challenges • Focus on partition sizing and cluster usage as core concepts • Assess alternate options that meet specific requirements • Advocate for features to be added where needed Solutions & Areas for ExplorationChallenge
- 17. Feature Engineering Machine Learning Journey CVS is making strides towards leveraging machine learning for Personalization but with this comes a unique set of challenges Model Training Implement Solution • Automate model training pipeline, making it generalizable across many use cases • Modernize architecture that expands algorithm options Collaboration • Blended roles and responsibilities • Hand-off between data science and data engineering can be clunky • Increase cross-training and encourage ML Engineer skillset • Integrated development and deployment teams in an Agile environment • Micro-segmentation and disparate use cases require many models to be trained and implemented in parallel • Robust model selection process • Create centralized features which can be leveraged across many use cases • Explore deep learning and reinforcement learning • Exploring optimization software like IBM CPLEX • Explore tools like Kubeflow and MLFlow to support model management and tracking Solutions & Areas for ExplorationChallenge • Complex feature engineering requires a lot of compute power • Optimization Layer constraints outside of model increases complexity • Model deployment is a manual process • Limitations to each modeling approach
- 18. Looking to the Future
- 19. Finding the Right Tool for the Job As future use cases evolve, there is a need to continually explore and new tools and technology to existing tech stack Real-Time Use CasesMachine LearningStorage and Compute Orchestration Azure Cosmos DB Azure Event Hubs
- 20. Questions?
- 21. Feedback Your feedback is important to us. Don’t forget to rate and review the sessions.