Behavior-aware decision infrastructure for behavioral health revenue cycle. We capture what actually wins — and make it permanent.
RCM AI optimizes for volume — it handles the predictable cases faster. Behavioral health denials aren't predictable. They're high-variability clinical scenarios with significant payer discretion at the medical necessity level.
The result: facilities that fight denials intelligently — with the right evidence, assembled in the right sequence, referencing the right precedents — overturn 40+ points more than those who don't.
That gap is entirely recoverable. The asymmetry isn't clinical — it's informational.
Every denial, appeal, and outcome becomes a structured precedent object — versioned, queryable, outcome-labeled. Not a note. Not a PDF. A data asset the facility owns.
Each precedent improves the next decision. The corpus compounds — the same dynamic that makes case law more valuable over time than any single ruling.
No one has built a structured, longitudinal, cross-payer corpus of behavioral health denial decisions. It doesn't exist in the published literature.
We don't draft appeals — we build cases. The corpus is what makes that repeatable.
How does appeal decision quality degrade when key staff turn over? What's the half-life of institutional RCM knowledge, and can it be modeled? Behavioral health has the highest staff turnover of any care setting — this is the most acute version of the question.
Payer denial logic shifts without announcement. How quickly can a corpus-based system detect behavioral change in denial patterns — and how does that compare to a human team's detection latency? This is an ML problem with a real-world outcome.
Can precedent-based systems maintain overturn rates during workforce disruption (staff loss, mergers, system migrations)? This is the resilience question that ARPA-H's Resilient Systems ISO was written for — and it's untested in the RCM setting.
These questions require real denial data at scale — structured, longitudinal, cross-payer. That dataset is what we're building. It doesn't currently exist anywhere in the published literature.
Validation methodology input. Study design review. ~2 hrs/month in-kind. Publication opportunity on corpus analysis findings.
Named academic evaluator on the ARPA-H application. Budget line for formal evaluation deliverables and publication rights.
UMich as institutional sub-awardee. Larger scope — drift detection methodology, knowledge transfer modeling, corpus analysis.
ARPA-H-SOL-24-103 (Resilient Systems Office) funds health systems that maintain function under disruption. Stratum's thesis — that precedent-based infrastructure preserves decision quality through workforce instability — maps directly to the ISO's mandate.
We're preparing a solution summary and need an academic partner with credibility in AI/ML validation, health system research, and study design.
This isn't a primary revenue path for us — it's a validation accelerant. The goal is to fund the rigorous evaluation that makes the corpus credible at scale.
Named academic partner on solution summary. Scoped to validation and study design methodology. Submission target Q2 2026.
First-mover publication rights on denial pattern analysis. IRB-compatible de-identified structure by design. Dataset doesn't exist anywhere else.
Early access to the precedent corpus and platform. Inform design of the analytics and research export layer.
Nothing yet. The goal of this conversation is to find out if the research questions are interesting to you and whether there's a natural intersection with your current work.
If it's interesting, we can talk about what the right scope looks like. If ARPA-H is the natural vehicle, great. If a more informal research collaboration makes more sense, also fine.
No commitment needed. I'd rather find the right fit than move fast on the wrong one.
The methodology you've already developed for multi-site outcome pattern aggregation is exactly the kind of expertise this work needs. The research questions are new, but the infrastructure problem is one you've already solved in a harder setting.