The Model Park announcement went out across five channels today and there is nothing on murphi.ai about it, so every CTA currently points at a site that does not mention the product. This is written to be that page. It needs the accuracy question below settled before it publishes.
Most AI coding tools make one decision on your behalf, quietly, before you ever see a chart. They decide how much model to spend on it.
Usually the answer is the same for every chart, because that is simpler to build and simpler to price. And it is a defensible choice. It is just worth noticing that it is a choice, and that it is being made by the vendor rather than by the agency.
Two charts
A routine start of care for a knee replacement. One primary diagnosis, two stable comorbidities, a clear referral, orders that match the plan. A competent coder works through it in a few minutes and is not in any doubt at the end.
A recertification for a patient with heart failure, diabetes, early dementia and a wound that has changed twice this episode. The medication list moved mid-episode. The therapy frequency was adjusted after a hospitalisation. The sequencing question is genuinely hard, and two good coders might reasonably order the diagnoses differently.
These are not the same task with different inputs. They are different tasks.
Running both through the same engine means one of two things is happening. Either the engine is sized for the second chart, in which case the first one is consuming depth it does not need. Or it is sized for the first, in which case the second is getting less than the situation deserves.
The first case is a cost problem. The second is a quality problem. Agencies have been living with one or the other because the choice was not theirs to make.
What Model Park changes
Murphi Model Park moves that decision to the agency, on two axes.
The model. Prime, Pro, Premier, Platinum or Pinnacle, each available in multiple versions. Lighter models for charts where the answer is not in doubt, heavier ones where it is.
The scope. Coding on its own for simple charts. Coding and OASIS together for moderate ones. Coding, OASIS and Plan of Care for complex ones.
The two axes are independent, which matters more than it might appear. Chart complexity and review scope are different questions. A straightforward chart might still need the plan of care checked because the episode has had three order changes. A clinically complex patient might have documentation that is entirely clean. Being able to set those separately is the difference between a real control and a preset with three options.
From the time data is ingested, the coding report is ready in 3 to 7 minutes.
On accuracy, and what the number describes
The target accuracy for the model output is 85 to 95 percent, before a coder reviews it.
That is worth sitting with for a moment, because it is lower than the figure you will see elsewhere on this site, and the relationship between the two is the whole point.
Murphi's coding accuracy is reported as 98 percent with human review. That figure describes the finished work, after a certified coder has gone through the AI output and approved it. The 85 to 95 percent describes what the model produces before that step.
Both numbers are real and they measure different things. A vendor quoting only the second is describing a supervised workflow. A vendor quoting only the first is describing a model. Describing a product with one number and hoping nobody asks which it is has become common enough in this market that we would rather publish both.
The reason this matters operationally is that the gap between the two numbers is the coder's work, and the size of that gap is what determines whether the review is sustainable. If the model output is close, the coder is confirming. If it is far off, the coder is re-doing the work with extra steps, and no amount of accuracy reporting will disguise that from the person actually doing it.
Which is also the argument for letting agencies choose the model. The right model is the one where the gap is small enough that review is quick, on that kind of chart, in that agency's hands.
The part that does not change
Choosing a lighter model changes how much work the AI does first. It does not change who signs off.
A certified coder or clinician reviews and approves every change before anything is written back. That is true of the lightest configuration and the heaviest one, and it is not a setting.
This is worth stating plainly because the obvious worry about tiered AI is that the cheaper tier quietly comes with less oversight. It does not. The tier determines what the model attempts, not whether a qualified human checks it.
Findings also carry the evidence they came from, across all configurations. Every OASIS item Murphi scores comes with the patient's own words or the record it came from. That is what makes the review fast enough to actually happen rather than degrading into approval by habit, and it would be self-defeating to vary it by tier.
How to think about choosing
A few principles that hold up in practice.
Start by matching scope rather than model. Scope is the easier decision and it has the larger effect. Most agencies know which of their charts need the plan of care looked at, because it tends to correlate with things they already track: episodes with order changes, hospitalisations mid-episode, therapy added after start of care.
Let the mix follow your census, not your worst case. The instinct when configuring anything in healthcare is to set it for the hardest case and leave it there. That instinct is correct for clinical safety and expensive for everything else. The hardest charts should get the heaviest configuration. The point is that they are a minority of your volume.
Watch the review time, not just the accuracy. If coders on a particular configuration are spending longer than you expected, that is the signal to move up a model, and it will show before any accuracy figure does. Review time is the leading indicator. Accuracy is the lagging one.
Change one thing at a time. Moving the model and the scope together on the same chart type makes it impossible to tell which change did what.
Every chart gets the model it deserves
The underlying idea is not complicated. Attention should go where the difficulty is.
A coding operation that spends the same effort on a clean knee replacement as on a four-comorbidity recertification is not being careful. It is being undifferentiated, and the cost of that shows up in two places at once: paying for depth on charts that do not need it, and having nothing left to add on the charts that do.
Matching the model and the scope to the chart in front of you is how that gets fixed. And the coder still reads it at the end, which is the part that makes all the rest of it safe to do.
← All articles