From PDF to Platform: A Decade of Model Evolution

"What in the world do you want me to do with this?"

Elliott Wood didn't sugarcoat where this all started. Mosai's first clinical decision support product — built under the company's earlier name, Medalogix — was a hospitalization-risk model, but the delivery mechanism was almost comically manual:

"Our very first product that we ever provided was risk of hospitalization probability... an engineer would go run the model on a Friday, and they would print that into a PDF, and then they would send that PDF to the clinician." — Elliott Wood, Mosai

The result was predictable. A clinician would open an email, see a number with no context, and have no idea what to do with it. Kate Warnock, the panel's moderator, captured the moment perfectly when reflecting on that era: a piece of information handed over with, essentially, zero instructions attached.

That gap — a score without context — is exactly the problem the last decade of model development has been solving.

What changed: from a single number to a living signal

Rather than a static probability delivered once, Mosai's platform today works as a continuously updating layer sitting on top of the clinical workflow. A few of the shifts the panel described:

• Continuous rescoring, not a single snapshot. Every visit updates the picture, so a clinician sees a trend — is this patient improving or declining — rather than one number frozen in time.
• Narrative documentation, not just structured data. Elliott Wood described pulling in the clinician's own notes alongside structured EMR fields, because the richest signal often lives in what a nurse actually wrote down, not just what got checked off.
• Context alongside the score. Instead of a bare risk number, today's output tells a clinician why — is this a hospice-eligibility signal, a hospitalization risk, a care-plan gap — so the next action is obvious rather than a guessing game.
• A resourcing and capacity layer. As Elliott Wood put it, a large part of the underlying work is making sure "the resource availability and capacity allocated to a patient is aligned with what that patient needs" — while keeping the actual decision in a clinician's hands.

The common denominator across all of it: the model got more useful because it got closer to how clinicians actually think and work, not because it got more autonomous. 

The story that shows the model working as intended

Justin Searle grounded this evolution in a concrete example from BAYADA's home health team, and it's worth walking through in full because it shows every layer of the system doing its job in sequence.

A 75-year-old client with kidney disease and heart failure — a high-acuity, multi-condition profile Searle noted is typical rather than rare in this population — was flagged into a high-risk category. The trigger wasn't a dramatic event. It was a 13-pound weight gain over a short window, the kind of detail that's easy to miss buried in routine documentation but is exactly the sort of pattern a continuously updating model is built to catch.

"You're in effect trying to figure out... where do you triage and how do you put resources into the most vulnerable patient or the most pressing issue." — Justin Searle, BAYADA

Once flagged, the response happened fast: the care team looped in the client's cardiologist, got her medication adjusted, confirmed she could take her own daily weight measurements at home, and shifted from in-person visits to daily phone check-ins. At the next review, her weight had come back down, and the system moved her back out of the high-risk tier.

"That's a hospitalization saved. And that's more time — that's more moments for her in her home, that's more moments not in a stale, sterile hospital bed where nobody wants to be." — Justin Searle, BAYADA

Searle was clear about what the model actually did in that story — and, just as importantly, what it didn't do:

"It's intended to aid and serve as a pulse for our clinical managers to say, hey, check this out... you have a patient or a client for us that's been elevated into a high-risk category." — Justin Searle, BAYADA

The model surfaced the signal early. Every subsequent decision — which specialist to call, whether to adjust medication, how often to check in — stayed exactly where Part 1 said it should: with the clinician. 

The same shift, on the hospice side

Amy Hirsch described an equivalent transformation in hospice care, framed around something every hospice clinician knows intimately: how much time actually matters.

"Instead of a reactive, we are more proactive... you didn't have the technology or the AI as a tool to help you make judgments. As a clinician, you're doing note by note by note... by then, you've lost days and time." — Amy Hirsch, BAYADA

She shared a story from a weekend triage team reviewing signals the morning after a late-Friday hospice admission — the kind that comes in, in her words, "a little bit of a hot mess" straight out of the hospital. A signal surfaced that the patient needed extra visits immediately. The team mobilized additional support that same weekend, including a social worker. The patient's condition declined quickly, but they were able to stay out of the hospital and pass peacefully at home, surrounded by family.

"That system allowed us to do what was best for that patient... and was surrounded by their family. It was a really beautiful experience." — Amy Hirsch, BAYADA

That's the real measure of a decade of model refinement: not a more impressive-sounding accuracy statistic, but a family getting the days they had left, spent the way they wanted to spend them. 

Why this is hard to shortcut

It's worth naming directly why this matters competitively. There's no shortage of newer entrants claiming similar capabilities — risk scoring, predictive alerts, AI-powered insights. What's much harder to replicate quickly is the loop: years of real clinician feedback, across thousands of cases, refining not just the model's accuracy but its presentation, its timing, and its integration into an actual care team's workflow. Mosai's models didn't get here through a single well-funded product sprint. They got here through the exact chain in Justin Searle's patient story repeating itself, thousands of times, since 2013 — each cycle making the next signal a little more useful, a little more trusted, a little more actionable.

That trust doesn't happen automatically, though — it has to be built deliberately with every clinician who's asked to rely on a new signal. That's where Part 3 of this series goes next: how BAYADA turned skeptical, experienced nurses into champions, and how that same conversation ties directly into preventing clinician burnout.

This is Part 2 of a 4-part series recapping Mosai's live webinar with BAYADA. Next up: "Trust, Adoption, and Reducing Burnout" — how BAYADA built clinician confidence in AI-driven signals, and why efficiency gains have to be protected, not absorbed into higher caseloads.

→ Watch the full webinar on-demand

→ Read Part 1 of the series

→ Continue reading with Part 3