When caring for Medicaid patients with complex needs, care teams face a significant challenge: deciding which intervention to prioritize when someone has multiple overlapping physical, mental, and social needs – from housing assistance and transportation barriers to medication management and behavioral health support.
Until now, these critical decisions have relied primarily on individual judgment and generic protocols. This manual decision-support approach not only makes it difficult to ensure the best possible outcomes for patients, but also limits our ability to learn from outcomes across entire patient populations.
To address this challenge, Waymark's team studied historical intervention outcomes from multidisciplinary care teams, comparing outcomes of similar Medicaid patients who received different intervention sequences. Our recent peer-reviewed study in JMIR AI demonstrates that Waymark's reinforcement learning approach prevents one emergency department visit or hospitalization for every eight patients who receive our algorithm-recommended interventions.
The Challenge: Managing complex physical and social needs in tandem
Care coordinators and community health workers supporting Medicaid populations navigate extraordinary complexity daily. Unlike physicians who follow evidence-based protocols for specific conditions, these team members must quickly prioritize among competing urgent needs and without standardized guidance.
Without appropriate tools, choosing which intervention to prioritize – for example, focusing on an antipsychotic medication versus ensuring adherence to a blood pressure management protocol – becomes highly subjective. National guidelines typically only address one condition at a time, but patients receiving Medicaid benefits often have multiple chronic conditions plus social needs that all interact. These sequential decisions have long-term consequences, and today's intervention choice directly affects a patient's future state and outcomes.
The Solution: Building and learning from a unique data foundation
Over several years, Waymark developed a fundamentally different approach to understanding care management decisions. We created a structured data collection system where our multidisciplinary care teams capture the full context of their decision-making process, instead of just documenting actions taken at any given time.
Our teams document:
- All available intervention options, not just the one chosen
- Their assessment of patient risk as they weighed different choices
- What's happening in the patient's life between visits: social needs, symptom changes and environmental factors
- The reasoning behind each choice they made for their patient's health
This structured format creates a unique dataset showing how provider choices relate to patient outcomes – information that cannot be easily derived from typical electronic health records or claims data.
With this foundation, Waymark's next-best-action model uses reinforcement learning to recommend the next-best-action to care teams—specifically, a SARSA (State-Action-Reward-State-Action) algorithm. This is fundamentally different from the AI tools most people are familiar with, like large language models that can "hallucinate" or generate plausible-sounding but potentially dangerous recommendations.
Our next-best-action model learns exclusively from these thousands of real care management decisions and their actual outcomes, identifying which intervention sequences consistently led to better results when providers faced similar choice sets for similar patients.
Results: Better, more equitable patient outcomes
This structured approach to capturing how care teams think through their options and make choices for each patient creates an irreplaceable foundation. The model acts like a seasoned care coordinator who has worked with thousands of patients over many years, recognizing patterns that individual team members couldn't possibly learn from experience alone. Unlike human memory, it never forgets a lesson learned and applies these insights consistently and equitably across all patients.
This unique dataset, built from years of structured documentation of provider thinking, available options, and patient outcomes, is what makes our model so effective and difficult to replicate. This tool demonstrated particular strength in recognizing complex medical-social interactions that standard approaches often miss. For instance, it identified when poor housing conditions were exacerbating respiratory issues and recommended housing interventions alongside medical care. It recognized when depression symptoms were actually related to underlying substance use that needed to be addressed first. These connections emerge from analyzing patterns across thousands of members – and these are insights no individual care team member could recall or reference alone.
When we applied this reinforcement learning model trained on our unique structured data to real-world care management, the results validated our approach. Our study, conducted across 3,175 Medicaid beneficiaries in Virginia and Washington, found significant improvements:
Clinical Impact:
- 12 percentage point reduction in emergency visits and hospitalizations
- For every 8 patients receiving interventions, one acute care event is prevented
- Greatest benefit for highest-risk patients, with only 5 patients in this group needing care to prevent one acute event
Equity Improvements:
- 28.3% reduction in gender-based outcome disparities
- 37.1% reduction in race/ethnicity-based outcome disparities
Transforming Medicaid care management at scale
With over 80 million Americans in Medicaid programs requiring care management services, the potential impact is substantial. Reinforcement learning augments it, providing evidence-based intervention recommendations while preserving care team autonomy to override suggestions based on individual circumstances.
The next-best-action system continuously learns and improves, with built-in monitoring to detect and prevent any potential bias, ensuring sustained fairness across all demographic groups. Every recommendation comes with clear explanations, helping care teams understand not just what to do, but why. By learning from thousands of real-world care decisions, we've created a system that prevents acute care events, reduces disparities, and recognizes the complex interplay between medical and social needs.
For health plans and providers seeking to improve outcomes for their most vulnerable members while reducing costs, our next-best-action model isn't an AI tool. It's a proven, peer-reviewed solution built on a foundation that many organizations simply don't have.
