The final morning of The 16th Annual HEDIS® & Quality Improvement Summit in Baltimore paired two sessions around one question every quality team faces as Electronic Clinical Data Systems (ECDS) and digital measurement take hold, which is how to review whole populations with the staff you already have. A quickfire panel moderated by Reid Kiser of Kiser Healthcare Solutions brought real results and misses from Marcos Vazquez of Ultimate Health Plans, Jimmy WuRohe of Clarest Health and Mariah Osborn of UPMC Health Plan. Tamara Hess and Katie Gries of Blue Cross of Idaho followed with a case study on building one chart repository for both HEDIS® and risk adjustment.
Here are five takeaways from both sessions, with what each one means for your plan.
1. Size the opportunity first, then match the tool to the task
Jimmy WuRohe, Chief Product Officer at Clarest Health, started with a funding gap many quality leaders recognize, since risk adjustment teams under the CFO win fast budget approvals thanks to a direct line to revenue, while quality teams under the CMO often struggle to fund technology. His fix is to quantify the opportunity before you evaluate any vendor price. If your plan is weighing a move from 3 Stars to 4 Stars in medication adherence, calculate the revenue and quality bonus payment value first, using a $1 million opportunity as the example. A vendor charging $100,000 to capture half of the value is a strong return, while a vendor charging $750,000 and claiming all of the value does not earn the investment.
WuRohe also told plans to ask every vendor why this technique and why AI for this problem, since traditional rule sets, standard machine learning or legacy NLP often deliver the same outcome at lower cost with better auditability. Reid Kiser reinforced the point by running identical prompts through paid versions of ChatGPT, Microsoft Copilot and Google Gemini, where two tools returned closely aligned analyses and the third returned inaccurate results. His advice was to treat generative AI like an entry-level intern who needs expert review and sign-off.
2. NLP bridges to digital quality, and its errors follow patterns
UPMC Health Plan, represented on the panel by Mariah Osborn, Director of Medicare STARs, uses natural language processing (NLP) as a transitional bridge while not every provider is FHIR-enabled, extracting evidence from unstructured charts to support year-round HEDIS® review. The side benefit is intelligence about your network, since NLP shows which provider systems keep quality documentation in unstructured notes and so which to prioritize for FHIR connectivity, data feeds and supplemental data contracting. UPMC judges reliability in three stages in place of one accuracy score: whether the engine reaches the chart repository, whether the engine isolates the exact clinical phrasing, and whether the evidence meets the measure specification.
Both sessions named repeat errors worth building your validation around. UPMC saw NLP flag cancer screening mentions under administrative headers like open care gap lists, and miss hospital-administered medications for Osteoporosis Management in Women Who Had a Fracture (OMW) because the pipeline had no connection to the hospital chart repository. Blue Cross of Idaho saw its engine struggle with Transitions of Care (TRC) discharge summaries, capture 2025 dates of service for HbA1c tests while ignoring 2026 data, pair a June 2025 service date with a June 2026 lab value on health information exchange summaries, and pull blood pressure readings from colonoscopy and surgical notes, which do not close a HEDIS® gap.
3. One chart, one request: Blue Cross of Idaho’s shared repository
Before the project, Blue Cross of Idaho ran quality operations out of 18 folders holding 84 subfolders, with separate portals, SFTP sites and faxes, while risk adjustment and quality asked providers for the same charts on separate timelines. Staff reviewed 20,000 to 25,000 charts in the prospective cycle and another 20,000 to 25,000 during retrospective HEDIS® audit season. Katie Gries said the manual approach stopped being sustainable once ECDS meant reviewing whole populations.
The answer was one web-based platform where providers see risk gaps, quality gaps and full rosters, message abstractors directly and use NLP to suggest gap closures. Quality led the effort with Risk Adjustment, embedded IT, compliance and cybersecurity at the table, so executives heard one enterprise ROI case. When the CFO ruled out adding headcount, the team ran time studies on abstraction in its legacy HEDIS® software and extrapolated the staffing ECDS would require. Tamara Hess described a first-year rule of two humans in the loop, with a primary abstractor evaluating every NLP-processed chart and a secondary abstractor verifying the result, after a February soft launch in Medicare Advantage and a May full launch.
The platform now covers about 65,000 attributed members and has passed 12,000 submissions against a projection of 8,800, with chart review turnaround under 48 hours and rejection feedback delivered inside the portal. Rural clinics went from 233 charts through late October last year to more than 1,000 this year, and Blue Cross of Idaho reported measure performance trending 5% to 10% higher across every measure, with breast cancer screening at the highest rate in the plan’s 80-year history. Pushing 4,000 risk adjustment charts through the NLP engine returned a 50% hit rate for open HEDIS® gaps in 25 minutes, work the presenters said would take abstractors hundreds of hours, though the engine flags risk adjustment conditions only when an open suspect already exists.
4. Predictive and agentic AI scale outreach and reporting
WuRohe showed why history alone misleads. Prior-year adherence data captures only 20% to 40% of the members who will show adherence risk this year, so historical targeting misses up to 80% of eventual failures while spending outreach on members who are not at risk, and Clarest’s machine learning models improved on historical targeting by more than 2x. In Comprehensive Medication Reviews, where clinicians have about 10 minutes and rules engines produce roughly 40 recommendations per chart, dynamic models rank the top targets while humans validate the final recommendation.
Marcos Vazquez, Director of HEDIS & Stars Operations at Ultimate Health Plans, grew membership from 7,300 in 2025 to about 23,000 this year without matching nurse staffing, so his team added agentic AI for routine member outreach and freed registered nurses for intensive clinical management. A 3,000-member campaign takes five or six coordinators a full business week by phone, while agentic voice AI reaches everyone and returns aggregated responses in about an hour. Plan for about 10% of members to ask for a person, roughly 300 callers in a campaign of 3,000, and require the voice agent to identify itself as AI, avoid clinical advice and transfer emergencies immediately. Vazquez also showed a ChatGPT-written prompt feeding Microsoft Copilot to turn raw gap data into an interactive dashboard in 2 minutes and 3 seconds, work he said takes an analyst three to four hours in Excel.
5. Governance starts on day one and continues after launch
Osborn told attendees implementations fail when only technical staff run them, so HEDIS® subject matter experts and abstractors belong at kickoff to define specifications and clinical sufficiency, while IT and data governance teams design pipelines for NCQA and CMS audit requirements. Finding non-compliant data formats late forces plans to discard finished builds and start over. Vazquez added third-party vendor risk assessments before signing any member-facing or provider-facing tool, verified security certifications to protect protected health information, and staged pilots before enterprise launch.
Governance also covers drift, when a tool meeting its benchmarks slowly loses accuracy as populations, guidelines, documentation practices or data feeds change. The panel’s protocol is to send the vendor corrective findings for retraining and reserve termination for cases where retraining fails. WuRohe pressed plans to demand customer-facing dashboards tracking pipeline health and extraction accuracy continuously in place of quarterly retrospective reports, with proactive alerts for member safety flags in medication regimens and for over-extraction in chart review, which creates regulatory audit exposure.
The morning ended with measure-by-measure work groups, where attendees traded barriers, strategies and lessons learned on specific HEDIS® measures.
Keep building your quality and Stars strategy with peers at The 17th Annual RISE Star Ratings Master Class, December 14-16, 2026, at the Fairmont Scottsdale Princess in Scottsdale, AZ.
HEDIS® is a registered trademark of the National Committee of Quality Assurance (NCQA) and any reference thereto by RISE Health does not imply any endorsement by NCQA of RISE Health and its offerings.