Pediatric murmur referral, to FDA submission

Worked example — open, and readable line by line

Pediatric murmur referral, to FDA submission

A hypothetical device, planned against real public records. It is here so a stranger can check what this tool does before describing anything of their own.

What this plan assumes about the device

What it does: A clip-on stethoscope and phone app that flags possible heart murmurs in children
Who uses it: A primary-care nurse
Where: A community clinic
Population: Children 2 to 17
Decision it supports: Refer for an echo, or not
Market: United States (FDA)
Next milestone: FDA submission

Estimated cash to FDA submission

$1.4M to $7.06M

at least $2.46M, modeled subtotal: 1 package still unpriced

The range is $5.66M wide. That width is not caution: $5.41M of it is one line, pivotal clinical performance study. That is where what nobody has settled yet actually sits.

A planning range, not a statistical interval. It moves the quantities — how many participants, how many annotation hours — and holds every price at its base, so each end is one coherent plan rather than a pile-up of worst cases. What a wrong price would do is the separate figure below. The range says how much the plan moves with the assumptions, not how likely any figure is.

Prices are held at their base. If all 4 of the rates that could move were at the top of their ranges at once — not a forecast, but the whole price question asked in one go — the same work would be at least $4.6M rather than at least $2.46M. Both are modeled subtotals: 1 package is still unpriced.

Staff effort

at least 15 person-months, and 8 packages never recorded what it takes

Staff time, counted separately from cash. It sits inside the burn and is never billed again as a cost.

Elapsed

48 months to the milestone.

Weakest basis anywhere in this total: our planning assumption. A total is never reported as stronger than its softest input — one published fee does not make a plan sourced.

1 package could not be priced. The total above is a subtotal until it is settled.

Annotated pediatric heart-sound dataset access — needs a quote you can settle this

Licensing terms for a pediatric heart-sound corpus depend on which corpus and on the intended commercial use. Nothing read in this pass prices it, and a placeholder here would be the invented number this planner exists to refuse.

What holds the date

Annotated pediatric heart-sound dataset access
Clinical annotation and annotation QA
Device feasibility study
Pivotal clinical performance study
Submission preparation
FDA 510(k) user fee

Elapsed time comes from the dependency path, not from adding the activities up. Work with slack can be shortened without the date moving, which is worth knowing before committing to a trim.

The roadmap — 10 priced, 1 unpriced, 1 deferred

Work Modeled cost Quantity Basis Slack

Clinical annotation and annotation QA $72K 400 clinician-hour our planning assumption 0m

Work	Modeled cost	Quantity	Basis	Slack
Clinical annotation and annotation QA	$72K	400 clinician-hour	our planning assumption	0m
Quantity — our planning assumption. Neither the hours nor the rate is published; both are planning placeholders. Rate — our planning assumption. A placeholder.
Analytical and bench performance evaluation	$57K	3 person-month	our planning assumption	29m
Quantity — our planning assumption. A planning placeholder. K252284 reports a non-inferiority design against its predicate but does not state the sample size, so the effort behind it is not observable. Rate — our planning assumption. A placeholder. Public wage tables give base salary by occupation and not loaded startup cost, and no wage source was read in this pass.
Pivotal clinical performance study	$1.83M	269 participant	borrowed from a comparable case	0m
Quantity — borrowed from a comparable case. A transfer from other people's studies, never an observation of this device. 120 / 269 / 916 participants: the 10th, 50th and 90th percentiles of 12 registered multi-site diagnostic studies of FDA-regulated devices, with actual counts and actual dates. Correction, twice: the second one was an outlier problem. The figure before this one read 53 / 265 / 1,534 and came from 37 studies — every multi-site diagnostic study, rather than every multi-site diagnostic study of a device. 25 of the 37 were not device studies, and the biggest enrolled 17,294 people into a health-systems access program. Filtering on the registry's own FDA-regulated-device flag left the median almost unchanged (265 to 269) and pulled in both tails at once: the top fell 40% and the bottom doubled. The figure before that one was 51 / 88 / 254, from the auscultation family, which is dominated by single-site observational work — feasibility-shaped, not pivotal. Twelve studies is thin, and it is thinnest exactly where it is quoted highest: with n=12 the 90th percentile is simply the second-largest study. Read 916 as 'about as big as the larger ones', not as a percentile. A scoped protocol from a specialist should replace all three numbers. Rate — borrowed from a comparable case. The low end is sourced and the rest is judgment, and those are different things. Akpinar 2019 gives CA$4,797 per enrollee for a device trial once the device itself is removed — clinical services 2504, management 1901, ethics 392 — and the device is 74% of their total, which for a phone app and a clip-on stethoscope is not a cost at all. Converted at the Bank of Canada 2018 average of 0.7718 USD/CAD and moved to 2026 by BLS CPI-U medical care (x1.2223), that is US$4,525. It is a floor rather than a middle: those are Canadian public-system prices, and Wenzl 2018 finds US cardiac implant prices two to six times German ones. The base and high apply 1.5x and 3x for US commercial pricing — deliberately below the 2-6x that paper found for devices, because it measured devices and this is services, and read across as a direction rather than a multiplier. That uplift is an expert assumption and a scoped supplier quote should replace it. The previous figure here was an unsourced 600/1100/2200, which the sourced floor alone exceeds fourfold. Sources: `BLS-CPI-MEDICAL`, `BOC-FXUSDCAD-2018`, `CTG-AI-MURMUR-SET`, `CTG-DIAGNOSTIC-PIVOTAL`, `PAPER-AKPINAR-2019`, `PAPER-WENZL-2018`
FDA 510(k) user fee	$6,517	1 submission	measured, from a named source	0m
Quantity — measured, from a named source. One submission per milestone, at the published small-business rate. Rate — measured, from a named source. Published MDUFA fee for 1 Oct 2025 to 30 Sep 2026. Exact, not a range — FDA charges what it charges. Sources: `FDA-MDUFA-FY2026`
Device feasibility study	$75K	11 participant	borrowed from a comparable case	0m
Quantity — borrowed from a comparable case. 4 / 11 / 40 participants: the 10th, 50th and 90th percentiles of 253 registered DEVICE_FEASIBILITY studies with actual counts and dates. This is the early study — does it work, is it safe enough to study properly — and 208 of the 253 ran at one site. An order of magnitude smaller than the pivotal that follows it, which is the whole reason a submission program needs both lines rather than one. Rate — borrowed from a comparable case. The low end is sourced and the rest is judgment, and those are different things. Akpinar 2019 gives CA$4,797 per enrollee for a device trial once the device itself is removed — clinical services 2504, management 1901, ethics 392 — and the device is 74% of their total, which for a phone app and a clip-on stethoscope is not a cost at all. Converted at the Bank of Canada 2018 average of 0.7718 USD/CAD and moved to 2026 by BLS CPI-U medical care (x1.2223), that is US$4,525. It is a floor rather than a middle: those are Canadian public-system prices, and Wenzl 2018 finds US cardiac implant prices two to six times German ones. The base and high apply 1.5x and 3x for US commercial pricing — deliberately below the 2-6x that paper found for devices, because it measured devices and this is services, and read across as a direction rather than a multiplier. That uplift is an expert assumption and a scoped supplier quote should replace it. The previous figure here was an unsourced 600/1100/2200, which the sourced floor alone exceeds fourfold. Sources: `BLS-CPI-MEDICAL`, `BOC-FXUSDCAD-2018`, `CTG-DEVICE-FEASIBILITY`, `PAPER-AKPINAR-2019`, `PAPER-WENZL-2018`
Algorithm development and training	$171K	9 person-month	our planning assumption	28m
Quantity — our planning assumption. A planning placeholder. Rate — our planning assumption. A placeholder. Public wage tables give base salary by occupation and not loaded startup cost, and no wage source was read in this pass.
Design controls and quality system	$66K	3 person-month	our planning assumption	38m
Quantity — our planning assumption. A planning placeholder. Rate — our planning assumption. A placeholder. Consulting rates are quoted, not published.
Submission preparation	$66K	3 person-month	our planning assumption	0m
Quantity — our planning assumption. A planning placeholder. The user fee is published; the work of assembling the submission is not. Rate — our planning assumption. A placeholder. Consulting rates are quoted, not published.
Software verification, validation and cybersecurity documentation	$76K	4 person-month	our planning assumption	28m
Quantity — our planning assumption. A planning placeholder. Rate — our planning assumption. A placeholder. Public wage tables give base salary by occupation and not loaded startup cost, and no wage source was read in this pass.
Human factors and usability evaluation	$44K	2 person-month	our planning assumption	28m
Quantity — our planning assumption. A planning placeholder. Rate — our planning assumption. A placeholder. Consulting rates are quoted, not published.
Annotated pediatric heart-sound dataset access	unpriced	needs a quote — you can settle this
EU notified-body conformity assessment	deferred	An EU route, named so it is visibly considered and deferred so it neither costs nor delays the US submission it is not part of.

Quantity — our planning assumption. Neither the hours nor the rate is published; both are planning placeholders.

Rate — our planning assumption. A placeholder.

Analytical and bench performance evaluation $57K 3 person-month our planning assumption 29m

Quantity — our planning assumption. A planning placeholder. K252284 reports a non-inferiority design against its predicate but does not state the sample size, so the effort behind it is not observable.

Rate — our planning assumption. A placeholder. Public wage tables give base salary by occupation and not loaded startup cost, and no wage source was read in this pass.

Pivotal clinical performance study $1.83M 269 participant borrowed from a comparable case 0m

Quantity — borrowed from a comparable case. A transfer from other people's studies, never an observation of this device. 120 / 269 / 916 participants: the 10th, 50th and 90th percentiles of 12 registered multi-site diagnostic studies of FDA-regulated devices, with actual counts and actual dates. Correction, twice: the second one was an outlier problem. The figure before this one read 53 / 265 / 1,534 and came from 37 studies — every multi-site diagnostic study, rather than every multi-site diagnostic study of a device. 25 of the 37 were not device studies, and the biggest enrolled 17,294 people into a health-systems access program. Filtering on the registry's own FDA-regulated-device flag left the median almost unchanged (265 to 269) and pulled in both tails at once: the top fell 40% and the bottom doubled. The figure before that one was 51 / 88 / 254, from the auscultation family, which is dominated by single-site observational work — feasibility-shaped, not pivotal. Twelve studies is thin, and it is thinnest exactly where it is quoted highest: with n=12 the 90th percentile is simply the second-largest study. Read 916 as 'about as big as the larger ones', not as a percentile. A scoped protocol from a specialist should replace all three numbers.

Rate — borrowed from a comparable case. The low end is sourced and the rest is judgment, and those are different things. Akpinar 2019 gives CA$4,797 per enrollee for a device trial once the device itself is removed — clinical services 2504, management 1901, ethics 392 — and the device is 74% of their total, which for a phone app and a clip-on stethoscope is not a cost at all. Converted at the Bank of Canada 2018 average of 0.7718 USD/CAD and moved to 2026 by BLS CPI-U medical care (x1.2223), that is US$4,525. It is a floor rather than a middle: those are Canadian public-system prices, and Wenzl 2018 finds US cardiac implant prices two to six times German ones. The base and high apply 1.5x and 3x for US commercial pricing — deliberately below the 2-6x that paper found for devices, because it measured devices and this is services, and read across as a direction rather than a multiplier. That uplift is an expert assumption and a scoped supplier quote should replace it. The previous figure here was an unsourced 600/1100/2200, which the sourced floor alone exceeds fourfold.

Sources: BLS-CPI-MEDICAL, BOC-FXUSDCAD-2018, CTG-AI-MURMUR-SET, CTG-DIAGNOSTIC-PIVOTAL, PAPER-AKPINAR-2019, PAPER-WENZL-2018

FDA 510(k) user fee $6,517 1 submission measured, from a named source 0m

Quantity — measured, from a named source. One submission per milestone, at the published small-business rate.

Rate — measured, from a named source. Published MDUFA fee for 1 Oct 2025 to 30 Sep 2026. Exact, not a range — FDA charges what it charges.

Sources: FDA-MDUFA-FY2026

Device feasibility study $75K 11 participant borrowed from a comparable case 0m

Quantity — borrowed from a comparable case. 4 / 11 / 40 participants: the 10th, 50th and 90th percentiles of 253 registered DEVICE_FEASIBILITY studies with actual counts and dates. This is the early study — does it work, is it safe enough to study properly — and 208 of the 253 ran at one site. An order of magnitude smaller than the pivotal that follows it, which is the whole reason a submission program needs both lines rather than one.

Sources: BLS-CPI-MEDICAL, BOC-FXUSDCAD-2018, CTG-DEVICE-FEASIBILITY, PAPER-AKPINAR-2019, PAPER-WENZL-2018

Algorithm development and training $171K 9 person-month our planning assumption 28m

Quantity — our planning assumption. A planning placeholder.

Rate — our planning assumption. A placeholder. Public wage tables give base salary by occupation and not loaded startup cost, and no wage source was read in this pass.

Design controls and quality system $66K 3 person-month our planning assumption 38m

Quantity — our planning assumption. A planning placeholder.

Rate — our planning assumption. A placeholder. Consulting rates are quoted, not published.

Submission preparation $66K 3 person-month our planning assumption 0m

Quantity — our planning assumption. A planning placeholder. The user fee is published; the work of assembling the submission is not.

Rate — our planning assumption. A placeholder. Consulting rates are quoted, not published.

Software verification, validation and cybersecurity documentation $76K 4 person-month our planning assumption 28m

Quantity — our planning assumption. A planning placeholder.

Rate — our planning assumption. A placeholder. Public wage tables give base salary by occupation and not loaded startup cost, and no wage source was read in this pass.

Human factors and usability evaluation $44K 2 person-month our planning assumption 28m

Quantity — our planning assumption. A planning placeholder.

Rate — our planning assumption. A placeholder. Consulting rates are quoted, not published.

Annotated pediatric heart-sound dataset access unpriced needs a quote — you can settle this

EU notified-body conformity assessment deferred An EU route, named so it is visibly considered and deferred so it neither costs nor delays the US submission it is not part of.

Rows are rounded for reading and the total is not, so the rows may not add to it exactly. The unrounded figures travel in the export.

Where these numbers come from

Pack ai-auscultation-2026-09-15. Two of six rates are published facts; the rest are planning assumptions and each row says which it is.

Record Supports Does not support

product code DQD Six AI/murmur-analysis devices have been cleared under product code DQD: K252284, K251494, K213794, K220766, K181988, K131044. The join is the product code, which is an FDA-assigned identifier. Whether any of these is an acceptable predicate for a particular new claim. Substantial equivalence is a regulatory judgment about intended use and technology, not a database lookup, and this record is a list of what exists.

K252284 A cleared AI murmur-analysis device carries an explicit pediatric indication. The 510(k) summary states verbatim: 'eMurmur Heart AI is intended for use on pediatric and adult patients.' It cleared by the 510(k) route citing predicate K220766, on a non-inferiority design reporting sensitivity 90.0% (95% CI 78.8-95.9), specificity 90.0% (78.8-95.9) and accuracy 90.0% (82.8-94.5). The validation sample size is not stated in the summary, so the study cost and effort behind it are not observable here. It also does not establish that some other pediatric claim — referral rather than decision support — would clear the same way.

K220766 The predicate chain is real and joinable: K252284 names K220766 in its own summary. Its indication text was not read, so its own pediatric scope is unverified here.

K213794 The regulatory comparison named in the OpenStetho reference-case spec exists and is correctly identified: product code DQD, applicant Eko Devices Inc. Its indications text was not read in this pass, so nothing here says which populations it covers.

NCT06155643, NCT05176899, NCT04400513, NCT06070298, NCT00564122, NCT04445012, NCT05987670, NCT07376785, NCT04933890, NCT05459545 Registered enrollment for the ten AI murmur studies that query returns (totalCount 10, so the set is complete for it, not a first page): 68, 103, 200, 220, 300, 300, 371, 577, 1145, 1150. Eight are OBSERVATIONAL and two INTERVENTIONAL. Three include a CHILD age band (NCT00564122 at 300, NCT04445012 at 1150, NCT07376785 at 220, all observational); seven are adult or older-adult only. The median is 300 across all ten and also across the three pediatric ones. Cost, effort, duration or site count. Registration records design and enrollment and almost never records a budget, so nothing here prices a study. It does not size one for an unreleased device either: eight of the ten are observational and none is a submission-grade clinical performance study for the device being planned, so carrying 300 across is a transfer from other people's studies and is labeled as one.

MDUFA FY2026 fee schedule Exact published fees for 1 Oct 2025 to 30 Sep 2026, in USD: 510(k) standard 26067, 510(k) small business 6517, De Novo standard 173782, De Novo small business 43446, annual establishment registration 11423. Small business means gross receipts and sales under 100 million for the most recent tax year, certified through the Small Business Determination program. Anything about the work of preparing a submission. A user fee is what FDA charges to review it; the effort to assemble it is a separate line and this record says nothing about it. Small-business eligibility is also a determination the applicant must obtain, not a rate anyone may simply choose.

auscultation family, interventional, includes children; n=21 with an ACTUAL enrollment Of 220 auscultation studies in the bulk export, 91 carry an ACTUAL enrollment; 51 of those are interventional and 21 of those include a CHILD age band. Their enrollment is 51 at the 10th percentile, 88 at the median and 254 at the 90th, and their median duration is 10 months. Restricted to enrollmentInfo.type = ACTUAL, which the bulk CSV does not expose: 19% of registered enrollments are ESTIMATED, and an estimate is a recruitment target rather than a count of who took part. Cost or effort, which the registry never records. And 21 studies is a thin base: the 90th percentile rests on two or three of them. It also does not establish that a submission-grade performance study for an unreleased device resembles this family, which is a judgment about analogy and not something a percentile can settle.

PMC6861404 A per-enrollee cost decomposition for medical device trials, in CA$ 2018. Experimental arm: device 13446, clinical services 2504, management 1901, ethics 392, total 18243. Control arm: device 4332, clinical services 7202, management 1901, ethics 392, total 13827. Both columns sum exactly. The structural finding: the device itself is 74% of the experimental per-patient cost, so for a device with no meaningful per-patient hardware — a phone app and a clip-on stethoscope — the transferable part is the other 26%, which is 4797 CA$ 2018. Mean trial duration 44.8 months (SD 18.2, range 6-66). A US figure. These are Canadian public-system prices from one centre, and the paper leaves out the manufacturer's own monitoring, data management and regulatory costs, so it prices a site rather than a trial. 18 trials, 399 enrollees, 56% cardiology, 50% implantable, and the authors state it is about 8% of the Canadian national sample. The 44.8-month mean is dominated by implantables with long follow-up and does not transfer to a short auscultation study.

Bank of Canada series FXUSDCAD, 2018 The 2018 daily average over 250 observations is 1.2957 CAD per USD, so 1 CAD = 0.7718 USD. Purchasing power. An exchange rate says what the currency is worth, not what a service costs where it is bought.

BLS series CUUR0000SAM Annual averages 484.71 in 2018 and 592.44 across the first eight months of 2026: a factor of 1.2223. Medical care rather than all items, which ran 1.3208 over the same period — the cost being moved is lab tests, imaging and procedures, so the medical index is the apt one and the choice is recorded because the two disagree by eight points. Anything about Canada, and anything about a specific site's pricing.

10.1377/hlthaff.2017.1367 US prices for cardiac implants run two to six times those in Germany, where they are cheapest: dual-chamber pacemakers about $4,200 in the US against $1,400 in Germany in 2014. Cited in the Alberta paper's own reference list. It is the reason a Canadian figure converted at an exchange rate is treated here as a floor for a US site rather than an estimate of one. Anything about clinical service prices, which is what the transferable part of the Alberta figure is. Canada is not in the comparison either. It establishes that a single-payer system's prices sit well below US commercial ones for devices, and it is being read across to services as a direction, not as a multiplier.

primaryPurpose = DEVICE_FEASIBILITY; n=253 with actual enrollment and actual dates The registry has an explicit DEVICE_FEASIBILITY primary purpose. Across 253 such studies with an actual rather than estimated enrollment and actual start and completion dates: enrollment 4 / 11 / 40 at the 10th, 50th and 90th percentiles, and duration 2 / 13 / 46 months. 208 of the 253 run at a single site. This is the early, safety-and-workability study, and it is an order of magnitude smaller than a pivotal. Cost or effort. And a feasibility study does not support a marketing claim: it establishes that the thing works well enough to study properly, which is why a submission program needs the pivotal as well.

primaryPurpose = DIAGNOSTIC, 6 or more sites, oversightModule.isFdaRegulatedDevice = true; n=12 Enrollment 120 / 269 / 916 and duration 6 / 25 / 58 months across 12 multi-site diagnostic studies that the sponsor declared FDA-regulated devices, with actual counts and actual dates. The site-count ladder is the clearest signal in the registry for how big a device study is: 1 site gives a median of 40 participants and 14 months, 2-5 gives 51 and 18, 6-20 gives 120 and 25, and 20+ gives 265 and 41. A pivotal is multi-site by construction. Cost. And the top of the range is thin in a way no wording fixes: Twelve studies means the 90th percentile is just the second-largest one. Read 916 as 'about as big as the larger studies we found', not as a percentile. Correction: this cohort was filtered after the fact, and the figure before it was wrong. It read 53 / 265 / 1,534 over 37 studies, because the cohort was every multi-site diagnostic study rather than every multi-site diagnostic study of a device. 25 of those 37 are not declared device studies at all. The largest, NCT04081051, enrolled 17,294 people in 'Advancing Access to Diagnostic Innovation Essential for UHC and AMR Prevention' — a health-systems access program, not a device pivotal. The median barely moved when the non-device studies came out (265 to 269), which is the reassuring part; both tails moved a long way, which is the point. The 90th percentile fell 40%, from 1,534 to 916, and the 10th rose from 53 to 120. The contaminating studies were stretching the range at both ends. It also does not establish that the device in front of you needs a study of that shape — how large a pivotal has to be follows from a sensitivity target nobody has set, which is a question for a specialist and not for a percentile. The cohort is rebuilt by dev/trials/pivotal_cohort.py, which prints its members by name so this check is repeatable rather than a claim.

Costs of Selected Clinical Trials, 103 costed trials, 87 with NCT numbers 103 trials with an actual total cost and 87 with an NCT number, sourced from CIRM awards (57), Knowledge Ecology International FOIA responses to four universities (24), SEC filings (16) and FOIA'd NIH grants (6). Per-patient cost runs $3,961 to $2,190,911 with a median of $97,233. It bounds a device rate from above — a stethoscope app trial costs less than a cell therapy — and the cheapest twelve trials cluster at $4,000 to $11,000, which is the band this pack's own floor sits in. The phase gradient runs in two directions at once. Normalizing the phase column first — the sheet spells the same rung 'Phase 3' and a bare '3', and reading it raw splits every rung — median enrollment climbs from 15 at Phase 1 (n=29) to 182 at Phase 3 (n=16), while median per-patient cost falls from $258,024 to $33,528. Median total cost barely moves: $6,192,579 to $7,695,272. Twelve times the patients at an eighth of the price each is roughly the same trial bill, because Phase 1's cost is fixed cost — batch manufacture, inpatient monitoring — divided over very few people. A device rate. There is no device trial in it: the two rows tagged 'device' are ViaCyte NCT02239354 and NCT03163511, which ClinicalTrials.gov classifies as COMBINATION_PRODUCT — an encapsulated cell therapy delivered in a device, at $127,302 and $359,136 per patient. Everything else is cell therapy, gene therapy, biologics or small molecules, which are the most expensive trials in medicine. Using any of it to price a software device would be the transfer this pack exists to refuse.

docket FDA-2011-D-0469 FDA's Center for Devices and Radiological Health recommends human factors and usability engineering processes so that a device is safe and effective for the users it is intended for, the uses it is put to and the places it is used, to minimize use error and the harm that follows. Final guidance, docket FDA-2011-D-0469, current as of 2026-08-03. That is the authority for a submission plan carrying human factors work at all when the device is operated by a clinician whose mistake could reach a patient. How much of it, or what it costs. The guidance is risk-based: how far human factors work has to go follows from a use-related risk analysis of the specific device, which nobody has done for this one. It sets no study size, no participant count and no price, and a recommendation to apply a process is not a requirement to run a summative validation study. The quantity and rate on WP-USABILITY remain expert assumptions and say so.

Record	Supports	Does not support
`product code DQD`	Six AI/murmur-analysis devices have been cleared under product code DQD: K252284, K251494, K213794, K220766, K181988, K131044. The join is the product code, which is an FDA-assigned identifier.	Whether any of these is an acceptable predicate for a particular new claim. Substantial equivalence is a regulatory judgment about intended use and technology, not a database lookup, and this record is a list of what exists.
`K252284`	A cleared AI murmur-analysis device carries an explicit pediatric indication. The 510(k) summary states verbatim: 'eMurmur Heart AI is intended for use on pediatric and adult patients.' It cleared by the 510(k) route citing predicate K220766, on a non-inferiority design reporting sensitivity 90.0% (95% CI 78.8-95.9), specificity 90.0% (78.8-95.9) and accuracy 90.0% (82.8-94.5).	The validation sample size is not stated in the summary, so the study cost and effort behind it are not observable here. It also does not establish that some other pediatric claim — referral rather than decision support — would clear the same way.
`K220766`	The predicate chain is real and joinable: K252284 names K220766 in its own summary.	Its indication text was not read, so its own pediatric scope is unverified here.
`K213794`	The regulatory comparison named in the OpenStetho reference-case spec exists and is correctly identified: product code DQD, applicant Eko Devices Inc.	Its indications text was not read in this pass, so nothing here says which populations it covers.
`NCT06155643, NCT05176899, NCT04400513, NCT06070298, NCT00564122, NCT04445012, NCT05987670, NCT07376785, NCT04933890, NCT05459545`	Registered enrollment for the ten AI murmur studies that query returns (totalCount 10, so the set is complete for it, not a first page): 68, 103, 200, 220, 300, 300, 371, 577, 1145, 1150. Eight are OBSERVATIONAL and two INTERVENTIONAL. Three include a CHILD age band (NCT00564122 at 300, NCT04445012 at 1150, NCT07376785 at 220, all observational); seven are adult or older-adult only. The median is 300 across all ten and also across the three pediatric ones.	Cost, effort, duration or site count. Registration records design and enrollment and almost never records a budget, so nothing here prices a study. It does not size one for an unreleased device either: eight of the ten are observational and none is a submission-grade clinical performance study for the device being planned, so carrying 300 across is a transfer from other people's studies and is labeled as one.
`MDUFA FY2026 fee schedule`	Exact published fees for 1 Oct 2025 to 30 Sep 2026, in USD: 510(k) standard 26067, 510(k) small business 6517, De Novo standard 173782, De Novo small business 43446, annual establishment registration 11423. Small business means gross receipts and sales under 100 million for the most recent tax year, certified through the Small Business Determination program.	Anything about the work of preparing a submission. A user fee is what FDA charges to review it; the effort to assemble it is a separate line and this record says nothing about it. Small-business eligibility is also a determination the applicant must obtain, not a rate anyone may simply choose.
`auscultation family, interventional, includes children; n=21 with an ACTUAL enrollment`	Of 220 auscultation studies in the bulk export, 91 carry an ACTUAL enrollment; 51 of those are interventional and 21 of those include a CHILD age band. Their enrollment is 51 at the 10th percentile, 88 at the median and 254 at the 90th, and their median duration is 10 months. Restricted to enrollmentInfo.type = ACTUAL, which the bulk CSV does not expose: 19% of registered enrollments are ESTIMATED, and an estimate is a recruitment target rather than a count of who took part.	Cost or effort, which the registry never records. And 21 studies is a thin base: the 90th percentile rests on two or three of them. It also does not establish that a submission-grade performance study for an unreleased device resembles this family, which is a judgment about analogy and not something a percentile can settle.
`PMC6861404`	A per-enrollee cost decomposition for medical device trials, in CA$ 2018. Experimental arm: device 13446, clinical services 2504, management 1901, ethics 392, total 18243. Control arm: device 4332, clinical services 7202, management 1901, ethics 392, total 13827. Both columns sum exactly. The structural finding: the device itself is 74% of the experimental per-patient cost, so for a device with no meaningful per-patient hardware — a phone app and a clip-on stethoscope — the transferable part is the other 26%, which is 4797 CA$ 2018. Mean trial duration 44.8 months (SD 18.2, range 6-66).	A US figure. These are Canadian public-system prices from one centre, and the paper leaves out the manufacturer's own monitoring, data management and regulatory costs, so it prices a site rather than a trial. 18 trials, 399 enrollees, 56% cardiology, 50% implantable, and the authors state it is about 8% of the Canadian national sample. The 44.8-month mean is dominated by implantables with long follow-up and does not transfer to a short auscultation study.
`Bank of Canada series FXUSDCAD, 2018`	The 2018 daily average over 250 observations is 1.2957 CAD per USD, so 1 CAD = 0.7718 USD.	Purchasing power. An exchange rate says what the currency is worth, not what a service costs where it is bought.
`BLS series CUUR0000SAM`	Annual averages 484.71 in 2018 and 592.44 across the first eight months of 2026: a factor of 1.2223. Medical care rather than all items, which ran 1.3208 over the same period — the cost being moved is lab tests, imaging and procedures, so the medical index is the apt one and the choice is recorded because the two disagree by eight points.	Anything about Canada, and anything about a specific site's pricing.
`10.1377/hlthaff.2017.1367`	US prices for cardiac implants run two to six times those in Germany, where they are cheapest: dual-chamber pacemakers about $4,200 in the US against $1,400 in Germany in 2014. Cited in the Alberta paper's own reference list. It is the reason a Canadian figure converted at an exchange rate is treated here as a floor for a US site rather than an estimate of one.	Anything about clinical service prices, which is what the transferable part of the Alberta figure is. Canada is not in the comparison either. It establishes that a single-payer system's prices sit well below US commercial ones for devices, and it is being read across to services as a direction, not as a multiplier.
`primaryPurpose = DEVICE_FEASIBILITY; n=253 with actual enrollment and actual dates`	The registry has an explicit DEVICE_FEASIBILITY primary purpose. Across 253 such studies with an actual rather than estimated enrollment and actual start and completion dates: enrollment 4 / 11 / 40 at the 10th, 50th and 90th percentiles, and duration 2 / 13 / 46 months. 208 of the 253 run at a single site. This is the early, safety-and-workability study, and it is an order of magnitude smaller than a pivotal.	Cost or effort. And a feasibility study does not support a marketing claim: it establishes that the thing works well enough to study properly, which is why a submission program needs the pivotal as well.
`primaryPurpose = DIAGNOSTIC, 6 or more sites, oversightModule.isFdaRegulatedDevice = true; n=12`	Enrollment 120 / 269 / 916 and duration 6 / 25 / 58 months across 12 multi-site diagnostic studies that the sponsor declared FDA-regulated devices, with actual counts and actual dates. The site-count ladder is the clearest signal in the registry for how big a device study is: 1 site gives a median of 40 participants and 14 months, 2-5 gives 51 and 18, 6-20 gives 120 and 25, and 20+ gives 265 and 41. A pivotal is multi-site by construction.	Cost. And the top of the range is thin in a way no wording fixes: Twelve studies means the 90th percentile is just the second-largest one. Read 916 as 'about as big as the larger studies we found', not as a percentile. Correction: this cohort was filtered after the fact, and the figure before it was wrong. It read 53 / 265 / 1,534 over 37 studies, because the cohort was every multi-site diagnostic study rather than every multi-site diagnostic study of a device. 25 of those 37 are not declared device studies at all. The largest, NCT04081051, enrolled 17,294 people in 'Advancing Access to Diagnostic Innovation Essential for UHC and AMR Prevention' — a health-systems access program, not a device pivotal. The median barely moved when the non-device studies came out (265 to 269), which is the reassuring part; both tails moved a long way, which is the point. The 90th percentile fell 40%, from 1,534 to 916, and the 10th rose from 53 to 120. The contaminating studies were stretching the range at both ends. It also does not establish that the device in front of you needs a study of that shape — how large a pivotal has to be follows from a sensitivity target nobody has set, which is a question for a specialist and not for a percentile. The cohort is rebuilt by dev/trials/pivotal_cohort.py, which prints its members by name so this check is repeatable rather than a claim.
`Costs of Selected Clinical Trials, 103 costed trials, 87 with NCT numbers`	103 trials with an actual total cost and 87 with an NCT number, sourced from CIRM awards (57), Knowledge Ecology International FOIA responses to four universities (24), SEC filings (16) and FOIA'd NIH grants (6). Per-patient cost runs $3,961 to $2,190,911 with a median of $97,233. It bounds a device rate from above — a stethoscope app trial costs less than a cell therapy — and the cheapest twelve trials cluster at $4,000 to $11,000, which is the band this pack's own floor sits in. The phase gradient runs in two directions at once. Normalizing the phase column first — the sheet spells the same rung 'Phase 3' and a bare '3', and reading it raw splits every rung — median enrollment climbs from 15 at Phase 1 (n=29) to 182 at Phase 3 (n=16), while median per-patient cost falls from $258,024 to $33,528. Median total cost barely moves: $6,192,579 to $7,695,272. Twelve times the patients at an eighth of the price each is roughly the same trial bill, because Phase 1's cost is fixed cost — batch manufacture, inpatient monitoring — divided over very few people.	A device rate. There is no device trial in it: the two rows tagged 'device' are ViaCyte NCT02239354 and NCT03163511, which ClinicalTrials.gov classifies as COMBINATION_PRODUCT — an encapsulated cell therapy delivered in a device, at $127,302 and $359,136 per patient. Everything else is cell therapy, gene therapy, biologics or small molecules, which are the most expensive trials in medicine. Using any of it to price a software device would be the transfer this pack exists to refuse.
`docket FDA-2011-D-0469`	FDA's Center for Devices and Radiological Health recommends human factors and usability engineering processes so that a device is safe and effective for the users it is intended for, the uses it is put to and the places it is used, to minimize use error and the harm that follows. Final guidance, docket FDA-2011-D-0469, current as of 2026-08-03. That is the authority for a submission plan carrying human factors work at all when the device is operated by a clinician whose mistake could reach a patient.	How much of it, or what it costs. The guidance is risk-based: how far human factors work has to go follows from a use-related risk analysis of the specific device, which nobody has done for this one. It sets no study size, no participant count and no price, and a recommendation to apply a process is not a requirement to run a summative validation study. The quantity and rate on WP-USABILITY remain expert assumptions and say so.