Closing Care Gaps with Agentic Population Health on Databricks

Closing Care Gaps with Agentic Population Health on Databricks

1. Problem / Context

Closing care gaps is a daily grind for care managers. Lists are often exported once a month, quickly become stale, and fail to reflect real-time eligibility or recent encounters. Nurses and care coordinators spend hours reconciling payer portals, EHR problem lists, lab systems, and spreadsheets—time that should go to patient outreach. For mid-market health systems, ACOs, and regional health plans operating under HIPAA with lean analytics teams, manual processes limit performance on HEDIS and Stars, put shared savings at risk, and contribute to burnout.

Databricks offers a pragmatic foundation to modernize population health operations. By unifying claims and EHR data in a governed lakehouse and orchestrating agentic workflows, care teams can receive fresh, patient-specific tasks every week—prioritized, scripted, and auditable—so they spend less time hunting for who needs care and more time delivering it.

2. Key Definitions & Concepts

• Care gaps: Evidence-based services or screenings a patient is due for (e.g., diabetic eye exam, colorectal cancer screening). Closing them improves quality metrics and outcomes.
• HEDIS measures: Standardized quality measures widely used by payers and providers (e.g., CDC, BCS, COL). They drive Stars ratings and incentives.
• Agentic population health: A governed automation pattern where AI-enabled agents observe data, decide what to do (e.g., identify a gap and prioritize a patient), and act (e.g., create tasks with outreach scripts), with humans in the loop and full audit trails.
• Lakehouse (Databricks): A single platform where raw and curated data (claims, EHR problem lists, labs, membership) lands in Delta tables with versioning, permissions, lineage, and scalable compute for weekly refreshes.
• Human-in-the-loop (HITL): Nurses confirm outreach, capture dispositions, and resolve ambiguous cases. HITL is essential for clinical safety and compliance.

3. Why This Matters for Mid-Market Regulated Firms

Smaller and mid-sized healthcare organizations face the same regulatory and quality pressures as large systems—without the same bench of data engineers, MLOps specialists, or clinical informaticists. Contracts increasingly tie reimbursement to measure performance, and auditors expect repeatable, explainable workflows. A lakehouse-backed, agentic approach allows these firms to move beyond monthly spreadsheets and toward reliable weekly cycles: refreshed attribution, up-to-date gap logic, and tasks routed directly into the care management system. The result is higher quality scores, better outcomes, and stronger positioning for shared savings—achieved at a low incremental cost by leveraging existing data and automating the manual “last mile.”

Databricks offers a pragmatic foundation to modernize population health operations. By unifying claims and EHR data in a governed lakehouse and orchestrating agentic workflows, care teams can receive fresh, patient-specific tasks every week—prioritized, scripted, and auditable—so they spend less time hunting for who needs care and more time delivering it.

4. Practical Implementation Steps / Roadmap

1) Land and unify data on Databricks

• Ingest claims (medical, pharmacy), eligibility/membership, and EHR extracts (problem lists, encounters, vitals, labs) into Delta tables.
• Apply basic data quality checks (membership continuity, primary care attribution, ICD/CPT code plausibility) and create curated views for population health.

2) Start with three HEDIS measures

• Begin with high-yield, well-understood measures such as: diabetes A1c testing (CDC), breast cancer screening (BCS), and colorectal cancer screening (COL).
• Document inclusion/exclusion criteria and coding references as versioned logic (SQL notebooks or Delta Live Tables), so changes are controlled.

3) Build gap detection and prioritization

• Implement measure logic to flag gaps per patient and compute priority scores (e.g., number of overdue measures, risk scores, last contact date, next appointment).
• Generate patient-specific gap lists with recommended next best action (e.g., schedule screening, order lab, coordinate imaging).

4) Orchestrate an agentic workflow

• An agent compiles the per-patient gap list and assembles nurse-ready tasks, including context and outreach scripts (e.g., “Ms. Rodriguez is overdue for a colorectal screening; offer FIT kit mail-out or next available colonoscopy slot”).
• The agent posts tasks into the care management tool queue (or a shared worklist) with due dates, channels (phone, portal message, SMS), and relevant clinical context pulled from claims and EHR problem lists.

5) Integrate with the care management tool

• Use secure APIs or flat-file drops to push tasks; read back status to keep the lakehouse as the source of truth.
• Schedule a weekly refresh job so lists stay current after new claims adjudication or EHR updates.

6) Human-in-the-loop execution

• Nurses review tasks, confirm outreach, capture dispositions (completed, declined, unable to reach), and note clinical nuances.
• Edge cases escalate to care managers or physicians. All actions are recorded for auditability.

7) Monitoring and continuous improvement

• Track throughput (tasks created vs. completed), closure rates by measure, and time-to-close.
• Iterate scripts and prioritization rules based on nurse feedback and outcomes.

5. Governance, Compliance & Risk Controls Needed

• HIPAA and PHI safeguards: Encrypt data at rest and in transit; enforce role-based access with Unity Catalog; apply row/column-level masking for sensitive fields (e.g., mental health, substance use); maintain signed BAAs with all vendors.
• Auditability: Version all measure logic, notebooks, and Delta tables. Keep immutable audit logs of tasks generated, scripts presented, and nurse dispositions. Preserve weekly snapshots so any closure decision can be reconstructed.
• Human oversight by design: Require nurse confirmation before patient contact; flag uncertain cases for review. Document standard operating procedures for outreach and escalation.
• Model and rule governance: Even if early phases rely on deterministic HEDIS rules, treat prioritization models as governed assets—register versions, monitor drift, and retain explainability.
• Vendor and platform resilience: Favor open formats (Delta) and documented APIs to avoid lock-in. Ensure SLAs for weekly refresh jobs and define a fall-back mode (e.g., last known list) if a run fails.

6. ROI & Metrics

Focus on measurable, near-term impacts:

• Measure closure rate lift: Compare closure rates for the three target HEDIS measures pre/post program. Many mid-market teams see a double-digit percentage lift when tasks are refreshed weekly and scripted for outreach, especially in the first two quarters.
• Cycle time reduction: Track time from identification to outreach completion. Weekly refreshes reduce the lag from weeks to days, raising the probability of closing gaps before year-end.
• Labor efficiency: Quantify nurse hours reclaimed from list reconciliation. Aim for a higher ratio of “patient touches per nurse hour” by minimizing time spent assembling context.
• Cost per closed gap: Divide all-in operating costs by the number of closures; use this to prioritize measures and channels (e.g., portal messages vs. phone calls).
• Financial impact: Tie closure lifts to Stars bonuses or shared savings. Even modest lifts on high-weight measures can materially improve plan ratings and value-based payments.

Example: A regional ACO begins with CDC, BCS, and COL. With weekly Databricks refreshes and agent-generated tasks, the team increases completed screenings by 12–18% over a quarter. Nurse time spent compiling lists drops by half, and the cost per closed gap falls below traditional call blitzes—resulting in improved quality scores and tangible shared savings.

7. Common Pitfalls & How to Avoid Them

• Boiling the ocean: Starting with 10+ measures dilutes focus. Begin with three that have clear logic and high impact.
• Stale or incomplete data: Without a scheduled weekly refresh and clear attribution, tasks miss the mark. Automate refresh pipelines and maintain membership continuity tables.
• Weak identity resolution: If patients can’t be matched across claims and EHR, outreach fails. Implement deterministic and probabilistic matching with human review for exceptions.
• Unscripted outreach: Leaving nurses to craft messages each time reduces consistency and throughput. Provide tested scripts per measure and channel.
• Missing audit trail: Regulators and payer partners expect traceability. Version logic and capture every task and disposition.
• No HITL checkpoint: Fully automated contact risks errors and trust. Require nurse confirmation and provide escalation pathways.
• Tool silos: If tasks don’t land in the care management system, adoption lags. Integrate early and close the loop with status feedback.

30/60/90-Day Start Plan

First 30 Days

• Discovery: Inventory data sources (claims, EHR problem lists, labs, membership), care management tools, and existing quality workflows.
• Data checks: Land sources in Delta tables, perform basic quality profiling, and document coding reference sets for the three starter HEDIS measures.
• Governance boundaries: Define PHI access roles, audit requirements, and HITL checkpoints; set up Unity Catalog and logging.

Days 31–60

• Pilot workflows: Implement measure logic, generate initial gap lists, and push tasks into a pilot queue for a single care team.
• Agentic orchestration: Add the agent that compiles patient context and outreach scripts; enable weekly refresh jobs.
• Security controls: Validate access policies, audit logs, and de-identification in non-prod; run a small-scale outreach campaign.
• Evaluation: Collect nurse feedback on task quality, scripts, and prioritization; refine before broader rollout.

Days 61–90

• Scale: Expand to additional clinics or lines of business; add channels (portal, SMS if approved) and refine prioritization weights.
• Monitoring: Stand up dashboards for closure rates, cycle time, and cost per closure; track data pipeline health and job SLAs.
• Stakeholder alignment: Review ROI with finance and compliance; publish SOPs and finalize governance artifacts. Plan the next set of measures.

9. (Optional) Industry-Specific Considerations

• Medicare Advantage vs. Commercial: Weight measures and outreach channels according to Stars impact and member preferences.
• Community and SDOH: Incorporate language preferences, transportation flags, and clinic hours to improve reach and reduce no-shows.
• Network alignment: Share aggregated, de-identified insights with affiliated providers to coordinate scheduling capacity for screenings.

10. Conclusion / Next Steps

Agentic population health on Databricks turns care-gap closure from a monthly scramble into a weekly, governed routine. By starting with three HEDIS measures, unifying claims and EHR problem lists, and integrating directly with the care management tool, mid-market organizations can boost quality scores, improve outcomes, and capture shared savings at low incremental cost—while preserving clinical oversight and auditability.

If you’re exploring governed Agentic AI for your mid-market organization, Kriv AI can serve as your operational and governance backbone. As a governed AI and agentic automation partner, Kriv AI helps teams with data readiness, MLOps, and workflow orchestration so you can move from pilot to production with confidence. For regulated healthcare organizations with lean teams, Kriv AI’s pragmatic approach helps turn agentic workflows into reliable, measurable results.