Embedding Predictive Analytics in All Aspects of a Health Plan Ian Blunt Health Datapalooza 3/27/2019 HIGHMARK.COM
Overview of Highmark Health Highmark is one of the largest integrated • • Aiming for transformational change in healthcare, delivery and financing networks healthcare delivery and financing in the nation We will be focusing on the Health Plan • • We are a national healthcare brand, we are today a leading edge technology platform, we are • One of nation's 10 largest health a healthcare provider. insurance organizations • Employs more than 40k people nationwide • Geographic Service Area includes and serves over 50 million Americans in all Pennsylvania, West Virginia and Delaware 50 states • Over 5 million active members • Our Mission: To create a remarkable health experience, freeing people to be their best. Customer Engagement & Insights is a • ~200 person organization focused on • Our Vision: A world where everyone advanced analytics, data science and embraces health member engagement 2
Our journey into advanced analytics Beyond 2016 2019 Enrollment and claims, Wide range of datasets, Full interoperability with Data and partners’ systems read-only growing API ecosystem, push to frontline systems interoperability Retrospective, descriptive, Predictive, near-time, Seamless prescriptive Analytics manual automated analytics Mainly SAS, many staff in SAS plus full open source Scouting for new same role 10+ years stack (R, Python, H2O, technologies, deep AI Tools and talent Keras etc). Staff rotation integration, leading talent and skills matrix culture Evidence driven “ Gimme ‘the data’ ”, varying Partnership for insight, Insights embedded within definitions caused swirl and standard data definitions, business, innovation center, culture data “in the DNA” distrust increasing self-serve, robust evaluation 3
Embedding predictive analytics in all aspects of a health plan Highmark proactively identifies Clients understand the value when a member would benefit from a Highmark’s programs create for them clinical intervention through robust evaluation Clients receive tailored Highmark meets the member where recommendations to improve quality they are in their journey using and lower cost for their employees personalized messaging Customer based on their risk profile Engagement & Insights Highmark continuously engages the Predictive analytics embedded in member (via multi-channel) to the underwriting process achieve desired outcome Providers Provider staff is well trained on the Highmark Machine learning provides customized recommendations around lower efficiency care analytic suite and can easily generate their own episodes actionable insights Share predictive analytics with strategic partners to enable joined-up intervention with members 4
Maximizing available data produces more impactful analytics Modeling retention in our Medicare Advantage direct pay members • Common to predict likelihood of direct-pay member Clinical Network retention – Highmark was using basic age and tenure Demographic Claim activity outreach disruption engagement regression model in early 2010s • We wanted to gain the maximum insight from our retention model, by predictive power and influential factors Social Household Enrollment Access determinants membership • Dramatically increased the scope of data we trained the model on by chasing processes and feature engineering • Our predictive power now matches leading published Health Provider Product program Complaints retention work in other industries, yielded new business affinity change engagement insights, shaping retention strategy • Data sources unlocked can be used for other modeling Customer Marketing Morbidity and projects Stars metrics service calls responses frailty Comprehensive data assembly and creative feature Base Added to 2017 Added to 2019 engineering are a powerful asset in driving value from 2015 model model model predictive analytics 5
Inserting analytics into operational processes drives value Using advanced analytics to route proven health advocacy in near-time • Started a dedicated health advocacy team to guarantee effective and Extreme need efficient care for member with most complex needs (analytics included from model Incoming start, w/ senior support) authorization • Needed to predict which cases are most likely to become complex as IP request requests arrive, can’t delay workflow • Developed a extreme need predictive model, trained on everything we know about a member’s care history Trigger rules • Allied to business rules which respond swiftly to specific pieces of information, all validated by outcomes measurement • Leverages sophisticated off-platform model to directly drive the routing in- system – required cultural change • Members with complex care needs now receiving dedicated support from a multi-disciplinary team from day one through to post-acute case UM care ICT manager management in the community from our integrated care center of excellence, leading to dramatic reductions in readmissions and length of stay Old process Embedding predictive analytics in the operational workflow New process requires strategic alignment, senior support, cultural change, 6 and pragmatism
Scale insight processes with machine learning Automatically detecting drivers of cost variation within treatment episodes No • Analyzing of the drivers cost variation within similar catheter procedure episodes produces actionable insight to drive efficiency, ? but using clinical expertise is resource intensive • Combine clinical expertise with machine learning techniques to expose causes of cost variation rapidly and Single at scale Less than physician • 3 caths? Automated feature engineering 000s of metrics for each ? episode • Machine learning algorithms to test every possible combination to “learn” which ones best separate the N N N N episodes into less or more efficient Avg Avg Avg Avg • The characteristics which led to them being classified O/E O/E O/E O/E Std Dev Std Dev Std Dev Std Dev gives questions to ask SMEs and clinicians • Actionable insights are turned into recommendations to impact workflow at the appropriate level (medical policy, A X X B health system, or specific facility/providers) More efficient Mixed Less efficient Finding new, ambitious use cases for predictive analytics can scale up 7 expert processes to impact value, if done sensitively
Using predictive analytics to understand our own effectiveness Quick, cost effective evaluation of clinical programs 80% • Essential to know value from clinical programs, studies can be slow, expensive Prevalence 60% • Retrospective matched controls method - 40% Potential control quick, cost effective, uses existing data, and Matched control 20% practical in real-world settings Intervention • Match “control group” members that did not get 0% the intervention to the intervention group on a member-by-member basis using predictive algorithms - represents what would have happened without the intervention • Compare any set of outcome metrics among the two groups using a simple statistical test • Informs scale up/down of programs, improves their design, influences routing Predictive analytics techniques deployed to measure value form clinical programs, builds culture of evidence-based decision making and feeds into 8 what other predictive analytics are used to trigger
Conclusion Just a few examples of our analytics transformation, lots more we could discuss, and even more work still to do While predictive models are easy to develop in isolation or as one-offs, building them into the workflow to be routinely used to actually impact the way we care for our members requires a different mindset to overcome technical and cultural barriers Achieve adoption Build confidence Taking machine and value by Relentless seeking Sustained senior through human learning into new injecting predictive out of data support for integration, and areas of the analytics direct into sources held by or analytics being a setting an the workflow, even business with accessible to the core part of wider expectation of ambitious use if that means organization strategy evidence-based cases challenge and decision making trade-offs What's next? Continue to explore new use cases where machine learning can add value, while exploring and developing ever more advanced analytic approaches (such as deep learning) 9
Recommend
More recommend
