ScreenPoint Medical

The model is the entire product. A mammogram arrives, a risk score comes back, and marks appear on the suspicious regions. There is no imaging hardware, no reading service and no archive underneath, and the company sells nothing that would function without the model.

The capability is also not a faster version of something people already do. A radiologist reading a mammogram does not assign a calibrated population level risk score that can be used to route the examination; that is a new object, and the routing decision built on it is only possible because the score exists.

The highest autonomy grade in this index that is genuinely earned by evidence rather than by design intent.

The consequential act is not detection but triage. The score decides how many radiologists read a woman's mammogram, so for the majority of examinations the model removes the second human reader that the standard of care requires. That is real autonomy over a screening safeguard.

What makes it defensible is that every part of it is published and was tested before deployment. The score scale is stated, the calibration is stated, the exact threshold for extra high risk is stated at 9.8, the routing rule is stated, and the whole workflow was then run against standard double reading in a randomised controlled trial with a hard clinical endpoint. A buyer can see precisely where the automation begins and what happened when it did.

The physician remains the reader throughout. Nothing is reported, recalled or cleared without a radiologist.

The most complete operating disclosure encountered in this sweep, and the reason is that the trial literature forced it.

Published material states the output scale from 1 to 10, how it is calibrated so roughly a tenth of examinations fall in each band, the risk bands used for routing, the exact extra high risk threshold of 9.8, the regional risk scale from 1 to 98, and the intended use as a concurrent reading aid for both full field digital mammography and tomosynthesis. The peer reviewed papers even name the software version used, which almost nothing else in this index does and which matters because a score is only reproducible against a stated version.

The gap is upstream: no description of architecture, training data provenance or the populations the model was developed on. Given how thoroughly the operating characteristics are published, that absence stands out rather than blending in.

Nothing identifies any party in the chain: no model or model family, no hosting arrangement and no sub processor list was located in two passes, and no retention schedule, encryption detail or position on model improvement was found. The population makes the unanswered consent question sharper here than in most of this index, and it is worth stating precisely. The material is mammography images of asymptomatic women who attended a public screening invitation rather than sought care.

Someone who books an appointment about a symptom has entered a clinical relationship and expects their images to be used within it. Someone who responds to a screening invitation has agreed to a specific offer, framed as a public health service, and their consent is to being screened rather than to anything that might later be done with the resulting images. That is a narrower agreement than the clinical case, and it is given at population scale by people who will never be asked again.

Whether images processed by the system are retained, whether they contribute to model development, and what screening participants are told, was not established. Ask what is retained after a read, whether processed images train models, and what the screening programme's participant information actually says about the software.

The strongest evidence base in this index, and it is not close.

This is a randomised controlled trial inside a national screening programme, more than 105,000 women allocated one to one against the existing standard of care, registered, single blinded, published across three papers in The Lancet family, and reaching the endpoint that actually matters. Interval cancers, the ones that appear between screening rounds because screening missed them, are the hard test of any screening change, and the final analysis reported a 12 percent reduction meeting non inferiority, 27 percent fewer aggressive subtypes, sensitivity of 80.5 percent against 73.8 percent, and identical specificity.

The earlier analyses reported a 44 percent reduction in reading workload and a 29 percent increase in cancer detection with no rise in false positives, and the additional cancers were disproportionately small and node negative, which is what early detection is supposed to mean.

This index has treated randomised comparison against standard care as its top bar. This is the reference record for it.

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

The material is mammography images of asymptomatic women who attended a public screening invitation rather than sought care, which is a population that consented to screening rather than to anything else. Whether images processed by the system are retained or contribute to model development, and what screening participants are told, was not established.

Graded on an honest basis. No compliance documentation was located in this pass.

The frame fits awkwardly, as with the other European records here. The company is Dutch and much of its deployment is inside national screening programmes in Europe and elsewhere, where the governing instruments are European data protection law and national screening governance rather than United States health privacy law. A United States buyer should establish what applies to them specifically rather than assuming European arrangements transfer.

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.

Operation inside national screening programmes in more than 30 countries implies assessment by public health authorities, which is a demanding gate. Nothing about it was retrieved.

Authorised on both sides of the Atlantic and across both mammography modalities. The published intended use, as a concurrent reading aid for physicians interpreting screening full field digital mammography and digital breast tomosynthesis, is United States labelling language, and deployment across more than 30 countries including European national screening programmes implies European conformity marking.

Held at B rather than A because specific clearance numbers and dates were not retrieved in this pass and should be confirmed before the grade is quoted.

One point of precision worth carrying: the authorised use is concurrent reading support. The workload reduction demonstrated in the trial comes from using the score to route examinations to single reading, which is a screening programme policy decision rather than something the label authorises on its own. A buyer adopting the triage workflow is adopting a protocol, not just a device.

More performance is published here than for almost any record in this index, and the specific subgroup question that matters most in mammography is still missing.

What exists: sensitivity, specificity, recall rate, positive predictive value, and the type, grade and stage of cancers detected, all from a randomised trial. That is a great deal.

What is absent is performance by breast density. Dense breast tissue reduces mammographic sensitivity for everyone and is the central confounder in this field, it varies by age and by ancestry, and women with dense breasts are precisely those most likely to have a cancer missed. A triage system that routes low scoring examinations to a single reader needs to be shown not to score dense examinations low for the wrong reason. No such breakdown was located.

Generalisation is the second gap, and it is inherent rather than a failure. The trial ran in four sites in southwest Sweden. The screening population, the equipment, the reading standard and the recall culture there are not those of every programme adopting the product. Declarations of interest in the published work were made openly, which is the system functioning as intended.

The operating characteristics are published to a level nothing else in this index matches, which is what a clinician actually needs to use a score responsibly. Published material states the output scale, how it is calibrated so that roughly a tenth of examinations fall in each band, the risk bands used for routing, the exact extra high risk threshold, the regional risk scale, and the intended use as a concurrent reading aid across both standard mammography and tomosynthesis.

Publishing the calibration and the exact threshold means a reader knows what a given score represents in the population and where the line for action sits, rather than being handed a number and left to infer both. One further practice deserves specific credit because it is almost unique here: the peer reviewed papers name the software version used.

A score is only reproducible against a stated version, and a paper that omits it describes performance of something the reader cannot identify, which is a quiet and widespread failure in this field. The gap is upstream. No description of architecture, training data provenance or the populations the model was developed on was located, and given how thoroughly the operating side is published that absence stands out rather than blending in. No warranty, indemnity or remediation commitment attaches. Ask for the development population and for performance across breast density categories.

Integration is with imaging infrastructure rather than the electronic health record, which is correct for the setting, and it is specific: the system writes its extra high risk flag directly into the reading worklist so the radiologist sees the prioritisation where they already work.

That matters more than it sounds. A screening programme reads in high volume batches, and a score that arrives anywhere other than the worklist changes nothing about how the day runs. Held at B because no interface standard, archive vendor certification or reporting system integration was located.

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

Deployment across more than 30 countries, much of it inside publicly run national screening programmes, means residency has been settled repeatedly and often under national rules that require images to remain in country. None of those arrangements is published.

Nothing published: no price, no mechanism, no unit of sale.

The omission is more consequential here than usual because the buyer is frequently a public screening programme spending public money, and because the economic case is unusually calculable. A 44 percent reduction in reading workload converts directly into radiologist hours at a known rate, and any programme can compute what that is worth to it. A published price per examination would let a health system model the whole thing against its own screening volume in an afternoon. Ask for it per examination and establish whether the price differs between the digital mammography and tomosynthesis pathways.

One organ, one purpose, and an unusually wide reach within it. The scope is breast cancer screening across both mammography modalities, with the company also positioning toward risk assessment beyond detection.

Deployment spans more than 30 countries and over 12 million processed mammograms, reaching national screening programmes, university hospitals and imaging groups, which is a broader institutional spread than most records here.

Held at B rather than higher because the focus is deliberately singular. Nothing extends to diagnostic breast imaging outside screening, to other modalities such as ultrasound or magnetic resonance, or to any other organ.