Medalogix

The models are the entire company. There is no record system, no scheduling platform and no service layer, and the products reach clinicians by being embedded inside record systems other companies sell. What is bought is prediction and nothing else.

The task also resists any simpler method. Estimating that a hospice patient is within seven days of death, from assessment data, vital signs, medications and free text narrative, is not something a rule or a threshold does. The company states it compares each patient against millions of prior records, which is the mechanism rather than a marketing flourish.

The model predicts and stratifies; a clinical team decides what to do. Nothing is ordered, prescribed or withheld automatically, and the stated response to a high risk prediction is to schedule more nursing and social work contact.

That placement is well chosen, because the decision the prediction feeds is a staffing decision rather than a treatment decision. Held at B rather than A because the influence is stronger than the formal role suggests: a team told a patient is likely within a week of death will act on it, and nothing published describes what a clinician is shown about the model's confidence, or what happens when their own judgement disagrees.

One boundary deserves watching rather than grading. A model that predicts mortality could in principle inform admission or eligibility decisions rather than care intensity, and hospice eligibility is an area of active federal enforcement. Nothing suggests that use, and a buyer should establish it is contractually excluded.

More quantified than most. The company states the number of data points compared per hospice patient at over 800, the scale behind the home health product at more than 11 million visits, and names the input types specifically: the federally mandated assessment, vital signs, medications and clinician narrative notes. Naming narrative notes matters, because it says the model reads free text rather than only structured fields.

The performance claim is where precision is needed and is not supplied. Sources variously describe greater than 90 percent precision, close to 95 percent accuracy, and 90 percent accuracy in predicting death within seven days, and those are different statistics with different meanings on a prediction where the base rate matters enormously. No confusion matrix, no false positive rate and no definition of the prediction window is published.

The inputs are named specifically and the structural question about the corpus is unanswered. On inputs the disclosure is better than most: the federally mandated assessment, vital signs, medications and clinician narrative notes are each identified, and naming narrative notes matters because it tells a buyer the model reads free text rather than only structured fields, which changes what a de identification claim would have to cover.

Scale is quantified too, at more than eight hundred data points compared per hospice patient and more than eleven million visits behind the home health product. The structural question follows directly from that scale and it is the material term in the agreement rather than a side issue. The models are trained on assessment and visit data from agencies that bought the product, and the resulting comparison base is what makes the prediction work for the next agency.

So whether one agency's data contributes to models sold to its competitors is the thing a buyer needs answered, and it is not public in either direction. The content sharpens it: this data describes the last weeks of people's lives, gathered in their homes. Nothing else is named, with no model or model family, no hosting arrangement and no sub processor list located. Ask whether your data trains models served to others, whether that is severable, and for a sub processor list.

A decade of deployment, enterprise adoption at some of the largest providers in the sector, and a set of specific outcome figures, none of it published.

The figures are better attributed than most: a near 20 percent reduction in hospitalisation and 2.4 fewer visits per episode for the home health product, an 8.2 percent reduction in unwanted hospital transfers reported by an integration partner, and as many as 82 percent more visits above the national average during end of life care. Named customers include the third largest hospice provider in the country.

All of it is vendor or partner reported, with no baseline, period, control or peer review. This index does not treat scale of adoption as evidence of benefit, and the strongest claim, that patients receive more attention in their final days, is the one most in need of independent measurement, because it is also the claim most likely to be true and most useful if demonstrated.

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

The structural question is specific to how this company works. The models are trained on assessment and visit data from agencies that bought a product, and the resulting comparison base, described as millions of records, is what makes the prediction possible for the next agency. Whether an agency's data contributes to models sold to its competitors is the material term, and it is not public.

The content is also unusually sensitive: the data describes the last weeks of people's lives, gathered in their homes.

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

The arrangement has a wrinkle worth establishing. The models are delivered inside record systems sold by other companies, so an agency may be contracting with its record vendor rather than with this company directly, and which party holds which obligation is not obvious from outside. Ask who your agreement is actually with.

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.

Enterprise adoption at national providers and embedding inside two major record platforms both imply security assessment has occurred. Nothing about it is public.

No device authorisation and none claimed.

The exemption is worth understanding because of what the model actually predicts. Software that supports a clinical decision sits outside the device definition where the clinician can independently review the basis and is not intended to rely primarily on it, and this product is positioned as informing how often a team visits rather than what treatment a patient receives. So a mortality prediction operating at national scale carries no regulatory oversight of its accuracy, because the decision it feeds is framed as staffing.

The sector's live regulatory exposure sits elsewhere, in hospice eligibility and length of stay enforcement, which is why the boundary noted on the autonomy axis matters.

Nothing published on subgroup performance, monitoring or error handling, and this is the record where that absence carries the most weight of any in this index.

The mechanism is allocation. A hospice has a finite number of nurses and social workers, and the model determines whose final week receives the extra visits. A patient the model does not flag is not merely unmeasured; they are less likely to have someone present when they die. Performance will not be uniform: a model learns the modal trajectory of decline, and patients whose deterioration follows an atypical course, or whose records are thinner because they were admitted late or documented less thoroughly, are the ones it will miss. Those are not randomly distributed patients.

The second gap is disclosure to families. The system generates a prediction that a person will die within a week. Nothing published states whether families are told, what they are told, or who decides.

Three different performance claims circulate for this product and they are three different statistics, which is the finding rather than the absence of one. Sources variously describe better than 90 percent precision, close to 95 percent accuracy, and 90 percent accuracy in predicting death within seven days.

Precision and accuracy answer different questions, and on a prediction of a rare event within a short window the base rate dominates both, so a buyer cannot reconcile them and cannot act on any of them. No confusion matrix, no false positive rate and no definition of the prediction window is published, and no warranty, indemnity or remediation commitment was located. The output makes the missing number specific.

A model predicting death within seven days drives whether a hospice team is deployed to a bedside, so a false positive means a family is told to prepare for a death that does not come, and a false negative means the visit is not scheduled for one that does. Both are irreversible in their own way and neither is countable from anything published.

This index has recorded elsewhere that this product allocates presence at a patient's death, and that allocation is only as defensible as the error rates behind it. Ask for the confusion matrix, the false positive rate at the operating threshold, the prediction window definition, and what a clinician is shown about the confidence of an individual prediction.

Distribution is the strategy, and it is a good one. Rather than asking agencies to adopt another application, the predictions surface inside the record systems clinicians already work in, including an exclusive partnership for the hospice product with the platform that states it serves around 37 percent of United States home health and hospice providers, and an integration that shows risk directly in another major post acute record workflow.

This is the fourth instance in this sweep of a vendor selling intelligence into somebody else's software rather than competing with it. The trade is real: reach and adoption in exchange for the customer relationship and, often, the commercial terms. Held at B because no interface standard or technical integration detail is published.

Not described. No hosting model, region or retention position was located.

The delivery path is the more interesting fact: because the models run inside partner record platforms, an agency's data may travel through the record vendor rather than directly, and the effective architecture depends on which partner they use. Nothing published sets that out.

Nothing published: no price, no mechanism and no unit of sale, and the partner distribution makes it harder still, since an agency buying through its record vendor may not see this company's pricing at all.

The unit matters in a sector with thin margins and fixed episode payment. Per patient, per episode, per agency and per seat behave very differently when a home health episode pays a fixed amount for 30 days regardless of how many visits are made. That is also the tension a buyer should name out loud: the home health product reports fewer visits per episode, which improves margin, while the hospice product drives more visits at the end of life, which costs money. Establish how each is priced and whether the vendor shares in either direction.

Deliberately confined to one setting and complete within it. Home health, hospice and palliative care are covered across the whole patient journey, from hospitalisation risk on admission through visit by visit management to the final days of life, and the company's own positioning is that this sector is all it does.

Reach is national and spans agency sizes, with enterprise adoption at some of the largest providers and availability to small agencies through the record platforms. Specialty coverage follows the population rather than a discipline, since home health and hospice patients arrive with every diagnosis. Nothing extends to facility based care, primary care or the hospital.