Iodine Software

The engine came first and everything else was built around it. The company launched its prediction product in 2015 and its emulation approach in 2016, well before general purpose language models, and the models have been the differentiator throughout rather than a layer added to workflow software.

The task itself resists rules. Reading every chart continuously and predicting which ones contain a documented clinical picture that does not match the recorded diagnoses is a judgement problem, and the company positions explicitly against the legacy rules based and retrospective systems that preceded it.

One qualification. Two acquired businesses, a physician query platform and a documentation integrity vendor, brought workflow tooling that is ordinary software. That surrounds the engine rather than substituting for it.

Two human decisions sit downstream of every prediction, which is a genuinely well structured oversight chain. The system surfaces and prioritises a case; a documentation specialist reviews it and decides whether to raise a query; the treating physician answers that query and their answer, not the model's, changes the record.

The company also states that its utilisation management product provides transparency and reasoning behind predictions, which is the right disclosure for a tool asking a specialist to spend their scarce time on this chart rather than that one.

Held at B because the prioritisation itself is consequential and unexamined in public material. A case the model ranks low may simply never be reviewed, and nothing published describes what share of charts are surfaced, or what happens to the ones that are not.

Substantially more disclosed than the category norm. The engine is named, the methodological approach is named and described as emulating clinical reasoning rather than applying rules, the architecture is characterised as a hybrid of generative models, large language models, natural language processing and machine learning, and the training base is quantified at 1.5 billion medical concepts across millions of admissions.

Quantifying the corpus is the disclosure that matters most here, because a prediction engine of this kind is only as good as the inpatient data behind it and a reader can weigh the claim.

What is absent is performance. No precision, recall or false positive rate is published for the predictions, which is the number a documentation team would actually work from, since a queue full of unproductive suggestions is the failure mode of the whole category.

The corpus is quantified and the question that quantification raises is unanswered. The engine is named, the approach is described as emulating clinical reasoning rather than applying rules, the architecture is characterised as a hybrid of generative models, language models and machine learning, and the training base is stated at 1.5 billion medical concepts across millions of admissions.

Quantifying a corpus is the disclosure that matters most for a prediction engine of this kind, because the product is only as good as the inpatient data behind it and a reader can weigh the claim rather than accept an adjective. It also makes the provenance question unavoidable, and this record carries the largest chart level exposure located anywhere in this index.

The product reads every inpatient chart continuously across close to five hundred hospitals, and the company describes the resulting dataset as among the largest inpatient clinical collections in the country and treats it as a competitive asset. So the corpus and the customer base are the same thing, and whether a given hospital's charts contribute to the asset sold to the next hospital, and on what terms that hospital agreed, is the material term of the relationship.

Nothing public answers it in either direction. No model provider, hosting arrangement or sub processor list was located either. Ask whether your charts join the corpus, whether that is severable, and what happens to your contribution at termination.

Strong third party corroboration and no clinical study. Best in KLAS for clinical documentation integrity in two consecutive years is an independent ranking drawn from a large body of provider interviews, and this index treats that as real evidence of delivered value rather than marketing. Deployment across more than 1,000 hospitals and health systems, a figure stated by the acquiring company at closing, supports it.

The outcome claims are financial and substantial: 1.5 billion dollars in additional appropriate reimbursement annually across the customer base, and more than 2.1 billion dollars recovered in 2024. Those are vendor reported, unaudited, and measured in dollars rather than in documentation accuracy, so they establish that the product moves money rather than that it makes records more truthful. The word appropriate is carrying weight in that sentence and nothing published tests it.

One further form of validation is worth naming plainly: a listed acquirer paid 1.25 billion dollars for this business after diligence. That is a market judgement rather than a clinical one, and it is not evidence the product improves documentation, but it is a substantial independent bet on the asset.

Graded on an honest basis. No retention schedule, encryption detail or training position was located in this pass.

The scale is worth stating because it is the largest chart level exposure of any record in this index. The product reads every inpatient chart continuously across close to 500 hospitals, and the company describes the resulting dataset as among the largest inpatient clinical collections in the country and treats it as a competitive asset. Whether customer charts contribute to that asset, and on what terms each hospital agreed, is exactly the question and it is unanswered publicly.

Graded on an honest basis and flagged for re verification. No compliance statement or agreement posture was located in this pass.

Agreements plainly exist and have been negotiated at scale with hundreds of hospitals; none of the terms is public. The specific clause a buyer should ask about is whether their chart data may be used to improve models sold to other hospitals, given how central the aggregate dataset is to the company's own account of its advantage.

Recorded honestly and provisionally: the dedicated trust and security search this index requires was not run in this pass, and no attestation was encountered incidentally.

A vendor reading complete inpatient records at nearly 500 hospitals has passed a great many institutional security reviews, so the absence here reflects what was retrieved rather than what exists.

No device authorisation and none claimed, which is conventional for this category and worth examining rather than accepting.

The product is sold as revenue cycle software, and yet what the engine actually does is infer, from the clinical evidence in a chart, that a patient probably has a condition which has not been documented. That is a clinical inference. It is exempt from device regulation because it is directed at a documentation specialist rather than at a treating clinician, and because its stated purpose is administrative.

The consequence a buyer should hold is that a model emulating clinical judgement at this scale operates with no regulatory oversight of its clinical accuracy, because the claim it makes is financial. The real regulatory exposure sits in federal false claims enforcement, where the question is whether the resulting documentation was supported.

Nothing published on evaluation, monitoring, error rates or subgroup performance, and this category has a structural issue that makes the omission matter more than usual.

Documentation integrity is asymmetric by design. The engine is built to find conditions that are clinically evident but not recorded, and finding them raises documented severity, which raises reimbursement. There is no equally resourced counterpart function hunting for conditions recorded but not clinically supported, because nobody pays for that. A model tuned to the first task and not the second will drift in one direction, and the drift is invisible in any single chart.

The company's pre bill product is framed as right sizing reimbursement, which implies correction in both directions, and that framing is credited here. What is not published is the distribution: how often the system's suggestions reduce a documented diagnosis rather than add one. That single number would settle the question.

Two passes located no precision, recall or false positive rate for the predictions, no evaluation methodology, no published limitations and no warranty, indemnity or remediation commitment. The missing figure is not a general accuracy number but the specific one a documentation team would work from, and its absence is the failure mode of this whole category rather than a disclosure preference.

The product surfaces suggested documentation queries for clinical staff to pursue, so the cost of a wrong suggestion falls on a specialist's time and, further down, on a clinician being asked to clarify something that did not need clarifying. A queue full of unproductive suggestions is how these products fail, and precision is the number that predicts it.

Nothing published states it, and nothing states the rate at which suggested queries are ultimately agreed by the physician or result in a documentation change, which is the same measure seen from the other end. The scale makes the absence more consequential rather than less: this runs continuously across close to five hundred hospitals, so a precision problem is not a local inconvenience.

The corpus point on the other axis compounds it, because a system whose value rests on a proprietary comparison base should be able to show what that base buys in measured terms. Ask for precision at the operating threshold, the physician agreement rate on surfaced queries, and what the vendor commits to when a query is unfounded.

Integration is a precondition rather than a feature. The product reads the complete chart continuously and in real time, which requires a live feed of notes, results, orders and medications rather than a nightly extract, and it returns work into the queues that documentation and utilisation teams already use.

The acquired physician query platform matters here too, because closing the loop means delivering a question to a treating physician inside their workflow and capturing the answer back into the record. Held at B because no named record vendor certification, interface standard or write back mechanism was located.

Not described. No hosting model, region, retention schedule or customer controlled option was located.

The workload implies substantial ongoing data movement, since continuous review of every inpatient chart at hundreds of hospitals is not a periodic batch job. Where that processing happens, and whether any of it can run inside the hospital estate, is not published.

The mechanism is named and the number is not, which is the split this index sees repeatedly. Third party product documentation describes an enterprise, outcomes based pricing model, with no rate published and a quote required.

Outcomes based is the important phrase and it deserves the same scrutiny this index applied to contingency pricing in payment integrity. If the vendor is paid as a share of reimbursement recovered, the vendor's revenue rises with documented severity, which is the same direction the product already pushes. That is not improper and it is a real alignment question on a product whose output determines what a hospital bills.

Ask three things: whether the fee is a share of recovery or a fixed licence, what happens commercially when a documentation change is later reversed on audit, and whether the utilisation management and pre bill products price on the same basis as documentation integrity.

Deep in the inpatient hospital and absent outside it. Coverage spans three connected functions, documentation integrity, utilisation management and pre bill review, which together cover most of the middle of the revenue cycle, and the buyer is a health system finance, health information management or case management leader.

Breadth comes from the fact that the engine is specialty agnostic: it reads whatever is in the chart, so a general medicine admission and a cardiac surgery admission are the same kind of problem to it. Reach is close to 500 hospitals. Held at B because everything is inpatient, with nothing addressing ambulatory documentation, and because coverage appears to be United States only, which is inherent given that the work exists because of how this reimbursement system codes severity.