I start with the flow of all FDA-authorized radiology AI devices, because the final count cannot exceed that pipeline. FDA updates summarized by The Imaging Wire imply 115 radiology AI authorizations from January-May 2025, 83 from June-September 2025, and 55 in Q4 2025; over an 86-day window, those rates correspond to about 65.50, 58.51, and 51.41 radiology authorizations respectively. Separately, Innolitics counted 19 radiology AI clearances in March 2026 and 15 in April 2026, which is a softer recent pace equivalent to about 48.03 over 86 days. Blending the stronger 2025 pace with the softer March-April 2026 pace, I center total radiology AI authorizations in the resolution window of May 18-August 11, 2026 at roughly 50. (theimagingwire.com)

The main reduction from total radiology AI authorizations to qualifying devices is the question’s strict transparency filter. A 2025 npj Digital Medicine review of 1,016 FDA-authorized AI/ML devices found that 51.6% reported no performance metric at all in their public summaries; only 23.9% reported sensitivity, 21.7% specificity, 10.9% AUROC, 6.4% accuracy, 6.5% PPV, and 5.3% NPV. A separate 2025 JAMA analysis of 903 FDA AI devices found clinical performance studies for 55.9% of devices, and among those studies sensitivity, specificity, and AUC were reported for only 36.2%, 34.9%, and 16.2% respectively. Since this question is narrower than either paper—it wants radiology only, diagnostic use only, and explicitly patient/exam/case/study-level numerical metrics—I think a countable share much above the low-20%s of all radiology authorizations would be hard to justify. (nature.com)

Recent 2026 device mix also points to a minority outcome rather than a majority outcome. March 2026 had 19 radiology AI clearances and April 2026 had 15, but many recent radiology clearances are clearly non-qualifying hardware/image-processing or workflow tools. Examples include Aquilion ServeSP, whose AI evidence is image-reconstruction/phantom quality rather than disease-level diagnosis; Synapse 3D Base Tools, whose reported metrics are mainly Dice-style segmentation metrics; and JLK-NCCT, which explicitly says it is not intended as a primary diagnostic device and is for triage/prioritization only. QIH, the huge generic image-processing product code, had 287 cleared devices overall and still produced multiple April 2026 entries of that general type, reinforcing that a lot of radiology AI volume is outside the resolution criteria. (fda.innolitics.com)

On the other hand, there is a steady stream of clearly countable-style devices. RevealAI-Lung is CADx software for characterizing incidentally detected lung nodules and reports patient/case-level performance, including a validation set of 675 patients with AUC greater than 0.8 and an MRMC reader study where mean AUC improved from 0.538 to 0.719. Product-code mix supports more such devices ahead: the QDQ diagnostic code had four clearances from February-April 2026, and the POK CADx code had two from January-March 2026. Those codes explicitly require standalone diagnostic-accuracy testing using measures such as sensitivity, specificity, predictive values, and ROC/AUC, so devices cleared under them are much more likely than generic QIH/JAK/LNH hardware or workflow products to satisfy this market’s metric-reporting requirement. That said, not every diagnostic-looking code will qualify, because some devices may still report lesion-, tooth-, or image-level metrics rather than patient/exam/case-level metrics. (fda.innolitics.com)

Putting those pieces together, my central model is: about 50 total radiology AI authorizations in the window, multiplied by a qualifying share around 21%-22%, yielding an expected value near 10.8. I then widen the distribution because both components are uncertain: the overall FDA pace could land closer to the recent March-April 2026 slowdown or rebound toward the stronger 2025 run rate, and the fraction of summaries with truly qualifying patient-level metrics can swing materially depending on whether the mix leans toward CADx/mammo/prostate/lung tools or toward reconstruction, segmentation, dental finding-level tools, and triage software. That gives me a mode around 10-11, with most of the mass in the high single digits to mid teens. My rough 80% interval is about 6-15, and my rough 90% interval is about 4-18. (theimagingwire.com)

I treated this as total radiology AI authorizations in the 86-day window times a very strict eligibility-and-reporting filter. Radiology remains the dominant FDA AI specialty: one JAMA review of devices listed through August 31, 2024 found 692 of 903 AI-enabled devices (76.6%) were in radiology, and a radiology-specific JAMA review found 723 of 950 FDA-authorized AI/ML devices (76%) were radiology devices. (jamanetwork.com)

For the authorization-volume base rate, I manually counted the dated entries on FDA’s official AI-enabled medical devices list in the same May 18-August 11 window and got about 27 radiology authorizations in 2023, 32 in 2024, and 61 in 2025. That count is my inference from the line-by-line FDA list rather than a published FDA subtotal, but it shows clear acceleration into 2025. (fda.gov)

I do not translate 61 one-for-one into qualifying devices, because the fine print excludes many common authorization types. A Nature taxonomy of FDA-authorized AI devices found that among 736 unique devices, 427 were quantification/feature-localization, 84 were triage, and 106 used AI for image generation, while only 47 were diagnosis, 45 detection, and 40 detection/diagnosis. The same paper found 117 of 736 devices were intervention rather than assessment. That implies only a minority of authorizations are in the part of the market most likely to satisfy this question’s diagnostic-use rule. (nature.com)

Transparency is the second hard filter. In a JAMA Network Open study of 903 AI-enabled devices, clinical performance studies were reported for 505 devices (55.9%), but discriminatory performance metrics were reported in only 200 summaries, including sensitivity for 183 devices, specificity for 176, and AUC for 82. A separate radiology-specific JAMA review found clinical testing in only 208 of 717 radiology device summaries with documentation. So even before applying the question’s extra requirement that the metric be patient/exam/case/study level, the public-summary reporting rate is far from universal. (jamanetwork.com)

A recent concrete example shows why I am cautious. The 2025 Saige-Dx clearance is clearly radiology and diagnostic, and its summary describes case-level output plus standalone testing on 2,002 screening mammograms from unique women, but the publicly available summary text shown here does not provide a numerical sensitivity, specificity, accuracy, PPV/NPV, or AUC value. Under this market’s resolution logic, devices like that are at real risk of not counting. (accessdata.fda.gov)

There is also a timing risk at the end of the window. FDA’s AI-enabled-device page says the list is updated periodically and that AI-enabled devices whose decision summaries are not published within the data-collection period will be incorporated only in a subsequent update. Since the question requires positive evidence from public FDA summary materials by August 11, 2026, late-window decisions are somewhat less likely to qualify. (fda.gov)

My working decomposition is roughly 60-75 radiology AI authorizations in the window, about 20-30% surviving the diagnostic-use exclusions, and then about 30-50% of that surviving set actually providing a qualifying patient-level metric in the public summary or via an explicit incorporated reference. That produces a plausible range from the mid-single digits to the mid-teens, with the center of mass around 9. I therefore center the forecast a little above 9, with most probability on about 6-13 and a thinner right tail into the high teens. (fda.gov)

My forecast is dominated by one issue: the question resolves only for devices that appear on FDA’s AI-Enabled Medical Devices List by August 11, 2026, and FDA says that list is updated only periodically; it also says devices whose decision summaries are not yet published for a given data-collection period are rolled into a later update. CDRH New shows the AI-enabled list was updated on March 4, 2026. Public reporting shows earlier list updates on about May 2024, August 12, 2024, and July 10, 2025, so the cadence has been irregular and often separated by many months rather than continuously refreshed. That means a mechanically resolving zero is very plausible even if underlying clearances continue at a healthy pace. (fda.gov)

If the list is updated in time, the underlying radiology-AI clearance flow is clearly large enough to support a nontrivial count. FDA’s official list, by my manual count of entries in the analogous May 18-August 11 windows, shows roughly 34 radiology AI authorizations in 2024 and roughly 63 in 2025. Separately, AuntMinnie reported that 2025 alone saw at least 255 radiology AI devices cleared through the 510(k) pathway, and the March 2026 FDA-list coverage article put radiology at 1,104 cumulative AI-enabled authorizations through the end of 2025. So the pipeline size is not the bottleneck; publication and filtering are. (fda.gov)

The strict inclusion filters remove most of those raw authorizations. Many radiology AI entries are scanners or reconstruction/image-processing products rather than standalone diagnostic software, and many others are triage or workflow tools that the resolution explicitly excludes. The 2024 npj Digital Medicine paper on FDA-cleared image-interpretation AI found that triage devices were the largest category, at 59% of products, and specifically noted that product code QAS corresponds to triage-style devices; the paper also excluded LLZ and QIH from its CAD-focused set because those codes often cover image-processing or management functions rather than explicit CAD assistance. That is directly relevant because recent FDA-list windows contain many IYN/LNH/JAK/LLZ/QIH/QAS entries. (nature.com)

Even among genuinely diagnostic products, the patient-level-metric requirement is a hard screen. The 2025 transparency study covering 1,012 accessible FDA AI/ML summaries through December 2024 found that 51.6% reported no performance metric at all, and only 53.1% reported a clinical study. A recent diagnostic example illustrates the problem: the August 2025 Saige-Dx summary is clearly diagnostic, describes case-level output, and describes standalone testing on 2,002 screening mammograms from unique women, but the visible public summary does not state a numerical sensitivity, specificity, AUROC, PPV, NPV, or accuracy value. Under this market’s fine print, near-misses like that do not count unless the summary explicitly provides a qualifying patient-level number or clearly incorporates one by reference. (nature.com)

Putting those pieces together, I used a three-scenario model. Scenario 1 is no relevant FDA-list update before resolution, which forces a 0; I assign that 56%. Scenario 2 is an update that captures only part of the target window and yields a low single-digit resolved count; I model that with a Poisson mean of 3 and weight 25%. Scenario 3 is a broader/timelier update that captures more of the window and yields a mid-to-high single-digit count; I model that with a Poisson mean of 8 and weight 19%. This gives a forecast mean of about 2.27, a median of 0, and a mode of 0. In words: I think the most likely resolution is zero because of update timing, but conditional on a timely update I think something like 3-8 qualifying devices is more plausible than either none or a double-digit surge. (fda.gov)

The key difficulty is that the resolution is much stricter than simply counting radiology AI clearances. A device only counts if, by August 11, 2026, it is on the FDA AI-enabled list, has Radiology as the lead panel, has a genuinely diagnostic intended use, and its public FDA summary positively shows at least one numerical patient/exam/case/study-level metric such as sensitivity, specificity, accuracy, AUC, PPV, or NPV. The FDA also notes that its AI-enabled list includes devices only after public decision summaries are available, which matters because missing or thin public summaries will fail this question’s positive-evidence test. (fda.gov)

For volume, radiology remains the dominant FDA AI specialty. The Imaging Wire’s March 12, 2026 summary of the FDA update through December 2025 reports 1,451 total AI-enabled medical devices, 1,104 radiology devices (76%), and 255 radiology authorizations in calendar year 2025 alone. The same article also notes that the FDA list includes not just standalone software but also imaging hardware with embedded AI, which is important because many such entries are outside this question’s inclusion rules. (theimagingwire.com)

Using the official FDA AI-enabled list itself as an analog window, my manual count of entries dated May 20, 2025 through August 11, 2025 yields 64 radiology-panel AI authorizations in that comparable summer span. Separately, Innolitics’ 2026 tracking found 19 radiology AI/ML SaMD clearances in March 2026 and 15 radiology AI authorizations in April 2026, which is broadly consistent with a raw 86-day radiology-AI flow in the neighborhood of roughly 60-65 devices rather than something dramatically lower or higher. (fda.gov)

But only a minority of those raw radiology AI authorizations are likely to survive the question’s screening. Recent FDA-market examples show that radiology clearances include RT contouring, surgical guidance, workflow/post-processing tools, image-reconstruction hardware, and other products that are not straightforwardly diagnostic under the resolution rules. Innolitics also highlights how broad product code QIH is, spanning things like dental X-rays, enteric tube placement, follicle scanning, general radiology suites, and 3D imaging base tools; that breadth implies many radiology-panel AI authorizations are not countable diagnostic devices here. My inference is that only about one-fifth to one-quarter of raw radiology AI authorizations in this period will clear the question’s diagnostic-use filter. (innolitics.com)

Performance-metric reporting is the second major bottleneck. In a review of 1,012 accessible FDA AI/ML summaries through December 2024, 51.6% reported no performance metric at all; only 23.9% reported sensitivity and 21.7% specificity. That paper found only modest improvement in transparency after the 2021 guidance, with the average ACTR transparency score still just 3.7 in 2024. A JAMA Health Forum study of 691 FDA-cleared AI/ML devices through 2023 was similarly pessimistic, finding sensitivity reported in 24.0% and specificity in 22.0% of summaries. FDA’s transparency principles, current as of June 13, 2024, encourage sharing device performance, training/testing data characteristics, and clinical study summaries, but they remain guidance rather than a mandatory standardized reporting template. (nature.com)

There is, however, a reason not to forecast too low. A 2026 medRxiv study focused on recent radiology computer-assisted triage/notification software from 2024-2025 found 38 systems and reported that 91% of pathology entries disclosed sensitivity and specificity, while 63% also reported ROC-AUC. I treat that result as an upper bound on reporting propensity for the eligible subset rather than a direct analog, because many triage-only products would be excluded by this question even if they report metrics. Still, it suggests that once you narrow to certain higher-risk diagnostic radiology software categories, explicit performance reporting can be common. (medrxiv.org)

Two concrete FDA examples show why the resolution count should sit well below the raw authorization count. First, the De Novo summary for Allix5 gives detailed patient-level performance, including a 5-year AUC of 0.70 with 95% CI 0.69-0.72, but the summary explicitly states that Allix5 is not intended to diagnose or detect breast cancer; under this market’s rules, that makes it a non-counter despite having strong patient-level metrics. Second, Saige-Dx is clearly diagnostic and case-level in nature, and its public 510(k) summary describes retrospective standalone testing on 2,002 screening mammograms from 12 U.S. sites, but the publicly available summary text shown here gives noninferiority language and dataset details without explicit numerical sensitivity/specificity/AUC in the public summary itself. Under the question’s strict positive-evidence standard, summaries of that sort are at real risk of not counting unless the incorporation-by-reference language is explicit. (accessdata.fda.gov)

Putting this together, my base model is: about 62 raw radiology AI authorizations over the 86-day window, multiplied by an overall qualifying fraction of about 14% after applying both the diagnostic-use screen and the patient-level-metric screen. That gives a mean near 8.7. I then widen uncertainty because the biggest error bars are not authorization volume but classification ambiguity: whether a device is truly diagnostic under the wording used, whether the public summary publishes a numerical patient-level metric, and whether sponsors use sparse special-510(k) summaries versus fuller direct/De Novo decision summaries. My distribution therefore centers on 8-9, with most probability mass in the mid-single-digits through low teens and only a thin tail above the high teens. In short: the raw radiology AI pipeline is large, but this question’s documentation requirements are strict enough that only a modest fraction should resolve positively. (fda.gov)

My median forecast is 8 devices, with a mean of about 8.8. The key reason I am below the raw authorization pace is that this market resolves on a very narrow subset: the device must be on the FDA AI-enabled list, be Radiology-led, have a genuinely diagnostic intended use, and publicly report at least one qualifying patient/exam/study-level metric in the FDA summary by August 11, 2026. The FDA itself says its AI-enabled list is not comprehensive and that some authorized devices are only added in a later update once decision summaries are published; the list page currently shows content current as of March 4, 2026. That creates a real downward pressure from publication/listing lag, especially for decisions near the end of the window. (fda.gov)

For the denominator, I used the official FDA AI-enabled list as the closest base rate source. Manually counting FDA radiology-panel entries in the comparable May 18 to August 11 window gives roughly 29 in 2023, 34 in 2024, and 64 in 2025. Separately, AuntMinnie reported that at least 255 radiology AI devices were cleared in 2025 alone. That combination suggests the raw number of AI-enabled radiology authorizations in this 2026 window is likely still high, and I center it around about 70, with a plausible rough range of 60 to 80. (fda.gov)

But only a minority of raw radiology AI entries appear likely to satisfy the question’s diagnostic-use filter. In the 2025 window, many FDA-listed radiology AI entries are obviously ultrasound/MR/CT system upgrades, reconstruction or image-processing tools, PACS/viewing tools, or radiation-treatment/planning products rather than standalone diagnostic software. More generally, a curated npj Digital Medicine study of AI devices for medical image interpretation found 140 FDA clearances from January 2016 to October 2023 for 104 unique AI-enabled CAD products, a much narrower set than the full radiology AI universe. I therefore use a diagnostic-share prior in roughly the 18% to 25% range rather than anything close to half of all radiology AI authorizations. That is partly an inference from the FDA product mix, not a directly reported FDA statistic. (fda.gov)

The toughest filter is public reporting of a qualifying patient-level metric. Broad evidence says this is far from universal. A JAMA Network Open analysis of 903 FDA AI-enabled devices through August 31, 2024 found clinical performance studies reported for 505 devices, but discriminatory metrics were reported in only 200 summaries. An npj Digital Medicine analysis of 1,012 FDA AI/ML summaries through December 2024 found that 51.6% reported no performance metric at all; sensitivity appeared in 23.9% and specificity in 21.7%. Those are all-device figures, so diagnostic radiology should do better than the aggregate, but they still imply that public metric reporting is a substantial bottleneck. (jamanetwork.com)

Recent device-level evidence supports not assuming near-100% reporting even for clearly diagnostic radiology tools. Saige-Dx is a Radiology QDQ device with decision date August 11, 2025. Its official FDA database page classifies it as radiological computer assisted detection/diagnosis software, and its PDF summary describes standalone multicenter testing on 2,002 DBT screening mammograms from unique women. But the public summary states only that the test met prespecified criteria and does not provide numerical sensitivity, specificity, accuracy, PPV, NPV, or AUROC in searchable text. Under this market’s strict resolution rules, that kind of summary would likely fail the qualifying patient-level-metric requirement despite being diagnostically oriented. (accessdata.fda.gov)

Putting those pieces together, my base calculation is: raw radiology AI authorizations about 70 multiplied by diagnostic-share about 0.22 multiplied by public numerical patient-level-metric reporting about 0.60 multiplied by a small accessibility/lag haircut of about 0.95, which gives about 8.8. I also checked a pessimistic case of 60 x 0.15 x 0.45 x 0.90 = 3.6 and an optimistic case of 75 x 0.25 x 0.70 x 0.95 = 12.5. Blending those scenarios leaves me centered in the high single digits rather than the teens. So my final forecast has median 8, mean 8.8, and most of the probability mass from roughly 4 to 14 devices. (fda.gov)

How many FDA-authorized AI radiology diagnostic devices will report at least one patient-level performance metric from May 18 to August 11, 2026?

• My most likely prediction for the number of FDA-authorized AI radiology diagnostic devices reporting at least one patient-level performance metric is 19.50, with a 50% chance of falling between 14.50 and 24.50.
• Radiology remains the most active sector for FDA AI authorizations, with roughly 255 devices cleared in 2025, providing a robust baseline of approximately 60 clearances for the 86-day window.
• Stringent resolution criteria will significantly filter the total number of authorized devices; only those with diagnostic intent (excluding triage/workflow) and explicit patient-level metrics in public summaries will count.
• Historical data suggests that over 50% of AI device summaries lack specific performance metrics, and many report lesion-level rather than the required patient-level data.
• Regulatory documentation lags and potential FDA staffing reductions are significant factors that may lower the final count of publicly accessible summaries before the August 11, 2026 deadline.
• Industry financial health remains strong, with major players like GE HealthCare and Siemens Healthineers maintaining high R&D spending to support a steady pipeline of new submissions.

Key figures

Historical context

The FDA began tracking AI-enabled medical devices in 1995, with radiology consistently dominating the field. By December 2025, 1,104 radiology AI devices had been authorized, representing approximately 76% of all AI medical devices. The volume has increased significantly over time: from 155 radiology AI clearances in 2023 to at least 255 in 2025. Despite this volume, transparency has historically been a challenge. A 2024 study published in npj Digital Medicine found that 51.6% of authorized AI/ML devices did not report specific performance metrics in their summaries. When metrics were reported, sensitivity (23.9%) and specificity (21.7%) were the most common. The FDA has attempted to address this through various initiatives, such as the 2021 AI/ML SaMD Action Plan and the January 2025 draft guidance on lifecycle management, both of which encourage more robust and transparent clinical validation.

Tailwinds

• Increased Regulatory Pressure: New FDA guidance (January 2025) and international standards (TRIPOD+AI) are pushing manufacturers toward more transparent reporting of clinical metrics.
• Maturing AI Market: As the market for radiology AI matures, healthcare providers are increasingly demanding ‘robust peer-reviewed evidence,’ incentivizing manufacturers to include patient-level metrics in their submissions.
• De Novo and PMA Pathways: While most clearances are 510(k), an increase in De Novo or PMA applications would naturally lead to more detailed clinical summaries and metrics.
• Rejection of Exemptions: The FDA’s recent denial of petitions to exempt certain radiology AI from 510(k) requirements ensures that these devices must continue to submit detailed performance data.

Headwinds

• FDA Personnel Reductions: Reports of a 19% reduction in the FDA workforce in early 2025 could lead to longer processing times or delays in publishing decision summaries.
• Strict Metric Definitions: The requirement for ‘patient-level’ metrics is a high bar; many manufacturers only report ‘finding-level’ (lesion-level) data, which would disqualify those devices.
• Reporting Lags: If the FDA database or the ‘AI-Enabled Medical Devices List’ is not updated promptly, devices cleared in July or early August 2026 may not have public summaries available by the August 11 deadline.
• Shift to Workflow Tools: A market trend toward ‘workflow optimization’ (which is excluded) could reduce the proportion of authorized devices that meet the ‘diagnostic’ requirement.

Detailed reasoning

My prediction is based on a multi-stage filtering process applied to the historical and projected volume of FDA radiology AI authorizations.

1. Authorization Volume Baseline: Historical data shows that FDA authorization of AI-enabled radiology devices has accelerated, with at least 255 devices cleared in 2025. This averages approximately 21 devices per month. The resolution window from May 18 to August 11, 2026, spans 86 days (approximately 2.8 months). If the 2025 rate holds, we would expect roughly 60 total radiology AI authorizations during this window. Given the steady R&D investment reported by industry leaders like GE HealthCare and Siemens Healthineers (averaging $345M and €487M per quarter respectively), a decline in submission volume is unlikely.

2. Diagnostic Intended Use Filter: The question strictly requires “Diagnostic Intended Use” and excludes triage-only, workflow, or measurement-only tools. While radiology accounts for ~76% of all AI authorizations, a significant portion (estimated at 25-35%) of these are for triage (CADt) or workflow optimization. Applying this filter reduces the pool of potential devices from ~60 down to approximately 39-45.

3. Performance Metric Filter: This is the most restrictive criterion. Analysis of historical FDA summaries indicates that roughly 51.6% of devices do not report specific performance metrics at all. Furthermore, the resolution requires these metrics to be “patient-level” (e.g., sensitivity or AUC calculated per patient or per exam). Many radiology devices, particularly for lung or breast imaging, report “finding-level” metrics (performance on individual lesions), which would not count unless patient-level data is also provided. I estimate that approximately 40-50% of the diagnostic devices will meet this specific reporting standard.

4. Documentation and Publication Lag: For a device to be counted, its summary must be publicly accessible by August 11, 2026. As of May 19, the FDA’s AI-enabled list only showed entries through December 2025, indicating a significant lag in list updates. While individual 510(k) summaries often appear faster than the consolidated list, the requirement for positive evidence in a publicly accessible summary will likely “shave off” devices authorized in the final weeks of the window (late July to August).

Synthesizing these factors (60 base authorizations × 70% diagnostic × 45% reporting metrics) leads to a most likely value of 19.5 devices. The 50% interval of 14.5 to 24.5 accounts for fluctuations in FDA throughput and the variable quality of summary documentation.

Key uncertainties

• Database Update Cadence: The most critical uncertainty is how frequently the FDA updates the public 510(k) database and AI-enabled list during the final weeks of the window.
• Interpretation of ‘Patient-Level’: There is a risk that many summaries will report metrics that are ambiguous regarding whether they apply to the patient or the individual finding, leading to conservative resolution (not counting them).
• Workforce and Policy Shifts: Potential changes in FDA leadership or policy regarding AI transparency could abruptly change the level of detail required in published summaries.
• Seasonal Authorization Fluctuations: While year-end (Q4) typically sees a surge in clearances, the May-August period may experience different seasonal volume patterns.

Conclusion

• Expect a count near 19-20 devices, representing about 30-40% of the projected gross radiology AI authorizations for the period.
• The “Patient-Level” metric requirement is the most significant filter; monitor FDA decision summaries for “case-level” or “study-level” sensitivity/specificity rather than just “lesion-level” data.
• Watch for FDA database update lags in July and August, as the resolution requires the summaries to be publicly accessible by August 11, 2026.

The question asks how many FDA-authorized AI-enabled radiology diagnostic devices will have Decision Dates within May 18–Aug 11, 2026 and whose FDA decision summaries report at least one standalone patient/exam/case-level performance metric. Forecasters used FDA AI-device list growth to estimate ~33 devices/month (A ≈ 94 in the 86-day window) and agreed radiology makes up ~74% of additions (R ≈ 0.74), but then apply stricter filters for diagnostic intended use (central D ≈ 0.25, biggest source of uncertainty) and for whether decision summaries include qualifying patient-level metrics (central M ≈ 0.60). Combining those inputs yields a central forecast of about 10–11 devices, with a median ≈10, a 50% interval ~7–13, an 80% interval ~4–18 and a 90% tail into the low 20s; the forecasters consider an upper cap near 50 implausible under current trends. Teams broadly agree on the base-rate approach and that uncertainty is driven most by how strictly “diagnostic” is interpreted and by summary-level reporting practices; they recommend auditing a past 86-day window and monitoring FDA list growth and summary availability to refine the estimate. Practical caveats include lumpy decision-date timing, possible delays in posting decision summaries, and sensitivity to any FDA policy changes that would raise or lower reporting of performance metrics.

Forecast and Reasoning

Forecast: Probability distribution:

• 5.00% chance of value below 1.0
• 10.00% chance of value below 2.0
• 20.00% chance of value below 4.0
• 30.00% chance of value below 6.0
• 40.00% chance of value below 8.0
• 50.00% chance of value below 10.0
• 60.00% chance of value below 12.0
• 70.00% chance of value below 15.0
• 80.00% chance of value below 18.0
• 90.00% chance of value below 24.0
• 95.00% chance of value below 32.0

Forecast Reasoning

We are forecasting the number of FDA-authorized AI-enabled radiology (lead panel 90) diagnostic devices with decision dates from May 18 to Aug 11, 2026 that have publicly accessible FDA summaries (or valid incorporated-by-reference summaries) showing at least one qualifying patient/exam/case/study-level numerical metric (sensitivity/specificity/accuracy/AUC/PPV/NPV), with silence/ambiguity counting as exclusion. The main uncertainty is the compound filter: (1) total radiology AI authorizations in a ~2.8-month window, (2) fraction that are truly diagnostic under the fine print (excluding workflow, triage-only, measurement-only, reconstruction, etc.), (3) fraction of those with explicit patient-level (not lesion-level) standalone AI metrics in the FDA-accessible summary, and (4) whether the needed summary is publicly available by Aug 11. On volume, the background implies ~255 radiology AI clearances in 2025 (~21/month), suggesting roughly 55–65 radiology AI decisions in this 86-day window, with moderate variance. On the diagnostic filter, Forecaster 2’s 65% seems high given how many radiology AI clearances are enhancement/quantification/triage/workflow, while Forecaster 3’s 25–35% may be a bit low; I center around ~30–45% diagnostic. On metrics, requiring patient-level (not lesion-level) and standalone metrics is materially stricter than “any clinical evaluation”; I center around ~45–65% of diagnostic devices meeting the metric rule, with a meaningful downside risk. On availability, I apply a modest haircut (some summaries posted late, statement-instead-of-summary cases, or list update lag), but not as extreme as Forecaster 4’s implied large cutoff effect; I center around ~85–95% availability. Multiplying midpoints (e.g., 60 total 0.38 diagnostic 0.55 metric * 0.9 availability) yields ~11–12, but given downside risks from the patient-level requirement and strict default-to-exclusion, I place the median at 10 and keep a wide upper tail to reflect Forecaster 2’s plausible “more diagnostic + more reporting” scenario.

An amateur forecast is likely to (a) extrapolate total radiology AI authorizations for ~3 months and (b) apply a generic “some fraction report metrics” assumption, while missing key exclusionary details that sharply reduce the count: lead-panel ordering, diagnostic-vs-workflow/triage/measurement distinctions, patient-level (not lesion-level) metric specificity, standalone-AI-only requirements, and the rule that missing/unclear information means exclusion plus the requirement that summaries be publicly accessible by Aug 11. My forecast explicitly models these multiplicative filters and keeps a broader interval to account for short-window volatility and documentation/publication uncertainty, so I’m more likely to be well-calibrated than a single-point or lightly-adjusted amateur estimate.

Forecast rationale (numeric):

— Iteration 1 — The forecasts converge on a simple pattern: many FDA radiology AI authorizations may occur in the window, but only a small subset will qualify as “FDA-authorized AI radiology diagnostic devices” that publicly report at least one standalone patient-level performance metric.

Main factors driving the estimate

• Underlying authorization volume is expected to be fairly high.
  Recent FDA radiology AI clearance pace suggests dozens of authorizations could appear over the period, so the raw pool is not the main constraint.

• Heavy filtering comes from the definition of what counts.
  Many radiology AI products are excluded because they are:

  • triage or workflow tools,
  • measurement-only systems,
  • reconstruction or image-processing products,
  • treatment-planning tools,
  • or hardware-embedded rather than standalone diagnostic devices.

• Public performance reporting is the biggest bottleneck.
  Even among qualifying diagnostic devices, many summaries do not clearly report an explicit, patient-level metric. Some may report lesion-level or other non-equivalent measures, while others may lack accessible summaries or have delayed publication.

• Administrative and publication lags matter.
  FDA list updates and public summaries may not appear quickly enough to capture every authorization in the window, which can push the observed count downward.

Areas of agreement

• The final number should be much smaller than the total number of radiology AI authorizations.
• The estimate is driven more by classification and reporting rules than by raw clearance volume.
• The most important uncertainty is how many devices are truly diagnostic and how many provide clean patient-level metrics.

Main disagreement

• The forecasts differ mainly on how strict the filter is and how well performance metrics are reported:
  • more conservative views land in the mid-single digits,
  • others place the center in the low-to-mid teens,
  • with a less likely upside if clearance pace and transparency both exceed expectations.

Overall, the collective reasoning points to a small qualifying count relative to total FDA radiology AI activity, with the outcome most sensitive to device type exclusions, reporting transparency, and FDA/publication timing.

— Iteration 2 — Overall, the forecasts use the same basic framework: start with a recent-rate estimate of FDA radiology AI clearances over the 86-day window, then narrow that pool by applying a series of eligibility filters.

Main reasoning patterns

• Base-rate approach: All estimates begin by projecting how many radiology AI devices will be authorized during the period, using recent quarterly clearance pace as the anchor.
• Sequential filtering: The projected total is then reduced by excluding devices that are:
  • triage-only,
  • workflow or measurement tools,
  • treatment-planning or other non-diagnostic systems,
  • or not clearly in radiology.
• Metric-specific requirement: A further reduction is applied for devices that do not explicitly report a patient-level performance metric (for example, study-, exam-, or case-level sensitivity/specificity or similar standalone performance results).
• Timing and listing lag: A major uncertainty is whether newly authorized devices will appear on the FDA AI-enabled device list and whether their summaries will be detailed enough before the deadline.
• Strict interpretation of ambiguity: Some reasoning treats unclear or incomplete summary language as exclusionary, which pushes the count lower.

Areas of agreement

• The answer depends less on authorization volume alone than on how many devices both qualify as diagnostic radiology tools and publicly report patient-level metrics.
• Reporting lag and summary wording are important sources of uncertainty.
• The final count is highly sensitive to how strictly “patient-level performance metric” is interpreted.

Main differences

• The biggest spread comes from how aggressively the initial pool is filtered:
  • a more permissive view yields a count in the low-to-mid 20s,
  • a moderate view centers around the low teens,
  • a stricter view drives the estimate into the single digits.
• Most reasoning suggests the 50-device cap is unlikely to bind, though a faster-than-expected clearance surge could make it relevant in a looser scenario.

Bottom line

The collective logic points to a moderate-to-low count, driven primarily by the combination of clearance pace, exclusion of non-diagnostic tools, and incomplete patient-level metric reporting. The main uncertainty is not how many devices will be authorized, but how many will satisfy the stricter reporting and eligibility criteria by the deadline.

— Iteration 3 — The forecasts converge on a funnel-shaped reasoning process:

1. Baseline volume is fairly high
   They expect roughly 50–60 FDA radiology AI authorizations in the May 18–Aug. 11, 2026 window, extrapolating from the strong late-2025 pace.

2. Most devices are screened out by strict eligibility rules
   The count is then reduced by several filters:

   • must be a radiology diagnostic device rather than triage, workflow, or measurement software
   • the FDA device summary must be publicly available by Aug. 11
   • the summary must explicitly report a patient/exam/case/study-level performance metric
   • devices with only lesion/finding-level metrics, or no usable public metric, are excluded

3. Transparency is the main bottleneck
   A recurring assumption is that many radiology AI summaries are incomplete or non-transparent, so only a minority of otherwise relevant devices will qualify.

Areas of consensus

• The authorization count itself is likely substantial.
• The number that actually report a patient-level metric is much smaller.
• The key uncertainty is not total approvals, but how many are diagnostic, publicly posted in time, and sufficiently specific in their performance reporting.

Main differences in reasoning

• Some estimates are more permissive about what counts as qualifying, producing a higher central estimate.
• Others apply stricter interpretations of “diagnostic” and “patient-level metric,” leading to a much lower result.
• The biggest swing factor is the assumed reporting/transparency rate among authorized devices.

Overall synthesis

Taken together, the reasoning supports a modest final count relative to the approval volume, likely in the single digits to low teens for a strict interpretation, with a plausible higher outcome if reporting is more complete and more devices meet the diagnostic criteria. The distribution is wide because both FDA throughput and, especially, public metric disclosure are uncertain.

Summary of Rationales

The forecasting rationales universally approach this question by treating the resolution criteria as a strict “cascading funnel.” Starting from a baseline expectation of total AI radiology clearances during the target window, each subsequent regulatory and administrative filter significantly reduces the pool of eligible devices.

• Administrative and Publication Lags: Although the question specifies a window ending on August 11, the rationales agree that the effective timeframe is much shorter. The FDA typically takes weeks to publish public 510(k) decision summaries, meaning devices cleared late in the window (late July or August) are highly unlikely to have accessible documents by the deadline. Furthermore, the FDA’s AI-Enabled Medical Devices List is updated in periodic batches. Because of this tracking lag, recent clearances may not appear on the official list by the cutoff date, creating a structural risk that devices cleared within the timeframe will fail to qualify.
• Narrow Diagnostic Criteria: The resolution criteria strictly limit eligible devices to those with a “Diagnostic Intended Use” (e.g., CADe and CADx). Forecasters note that the vast majority of the current AI radiology market consists of triage (CADt), workflow optimization, and measurement-only tools, all of which are explicitly excluded. This filter eliminates most of the initially cleared devices.
• Strict Metric Reporting Standards: Among the surviving diagnostic devices, public summaries notoriously lack transparency. Many developers rely on older predicate devices or omit standalone performance data entirely to protect proprietary information. Additionally, when performance metrics are provided, detection tools frequently report at the lesion, finding, or image level (such as FROC) rather than the required patient, exam, case, or study level. Because the verification logic strictly penalizes ambiguity, any failure to explicitly identify the required denominator will disqualify the device.

Conclusion Across all samples, forecasters agree that the compounding effects of administrative publication delays, the exclusion of non-diagnostic tools, and the poor public reporting of explicit patient-level metrics drastically constrain the final estimate. These bottlenecks reduce an initial pool of dozens of expected clearances to a very small handful, with the list-update lags creating a distinct possibility that zero devices will fulfill all technical requirements by the deadline.

SUMMARY

Question: How many FDA-authorized AI radiology diagnostic devices will report at least one patient-level performance metric from May 18 to August 11, 2026? Final Prediction: Probability distribution:

• 10.00% chance of value below 1.976667
• 20.00% chance of value below 2.876667
• 40.00% chance of value below 4.576667
• 60.00% chance of value below 6.543333
• 80.00% chance of value below 9.96
• 90.00% chance of value below 13.36

Total Cost: extra_metadata_in_explanation is disabled Time Spent: extra_metadata_in_explanation is disabled LLMs: extra_metadata_in_explanation is disabled Bot Name: extra_metadata_in_explanation is disabled

Report 1 Summary

Forecasts

Forecaster 1: Probability distribution:

• 10.00% chance of value below 0.49
• 20.00% chance of value below 1.19
• 40.00% chance of value below 2.19
• 60.00% chance of value below 3.49
• 80.00% chance of value below 6.49
• 90.00% chance of value below 9.49

Forecaster 2: Probability distribution:

• 10.00% chance of value below 4.49
• 20.00% chance of value below 5.49
• 40.00% chance of value below 7.49
• 60.00% chance of value below 9.49
• 80.00% chance of value below 12.49
• 90.00% chance of value below 15.49

Forecaster 3: Probability distribution:

• 10.00% chance of value below 0.95
• 20.00% chance of value below 1.95
• 40.00% chance of value below 4.05
• 60.00% chance of value below 6.65
• 80.00% chance of value below 10.9
• 90.00% chance of value below 15.1

Research Summary

The research summarizes recent FDA authorization trends and reporting practices for AI/ML medical devices in radiology through December 2025 and examines how those patterns affect the ability to identify devices that explicitly report patient-level performance metrics during the target window (May 18–Aug 11, 2026). Key quantitative findings: by end-2025 there were roughly 1,400–1,450 total FDA-authorized AI/ML medical devices with radiology accounting for about 75–76% (~1,100–1,104) of them; in 2025 roughly 295 AI devices were cleared overall and ~211 were radiology, giving an average ~17–18 radiology AI clearances per month (Q4‑2025: 55 radiology devices over three months). Using 2025 rates, the 85-day window corresponds to roughly ~49–51 expected radiology device clearances. Reporting and transparency findings: studies and FDA-summary reviews found substantial variation and gaps—many summaries omit study design, fewer than 2% cite randomized trials, only small proportions explicitly report PPV/NPV (~10%) or prevalence (~19%), and many devices report lesion- or finding-level metrics rather than patient/exam/case-level metrics. Regulatory context noted includes finalized PCCP guidance and other FDA AI lifecycle/transparency efforts from 2024–2026, continuing reliance on 510(k) pathways for most devices, and no identified prediction-market contracts specific to this question.

Sources used (from the research searches and reviews):

• U.S. Food and Drug Administration (FDA) — AI/ML and medical devices resources and decision summaries (FDA 510(k) and related pages): https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-resources and https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/pmn.cfm
• Peer-reviewed analyses and systematic reviews of FDA-cleared radiology AI devices (2024–2026) — bibliographic/search access: https://pubmed.ncbi.nlm.nih.gov
• Radiological Society of North America (RSNA) materials and policy commentary: https://www.rsna.org
• American College of Radiology (ACR) materials and policy commentary: https://www.acr.org
• News and industry coverage of FDA AI device clearances and policy (examples of outlets searched): https://www.statnews.com, https://www.fiercebiotech.com
• Prediction-market platforms checked for relevant contracts: https://manifold.markets, https://polymarket.com, https://metaculus.com

(If you want, I can now enumerate or link the specific FDA decision summaries for radiology-authorized devices from mid‑May 2026 onward for direct verification.)

RESEARCH

Report 1 Research

I’ll help you research this question about FDA-authorized AI radiology diagnostic devices. Let me conduct a systematic investigation of the relevant trends, base rates, and prediction markets.

Agent Forecast (discrete, confidence: low)

• 9: 8.5%
• 8: 8.5%
• 10: 8.0%
• 7: 8.0%
• 11: 7.2%

Base Rate: Q4 2025: 55 radiology AI devices cleared (18.3/month); March 2026: 19 radiology SaMD clearances. Historical performance metric reporting: ~24% of all AI devices reported sensitivity through 2023 (JAMA study of 691 devices). Recent (2024–2025) diagnostic radiology AI devices show somewhat higher reporting per medRxiv 2026 analysis, but no direct estimate exists for patient-level vs. finding-level subdivision.

Initial Reasoning:

• Expected ~51–57 total radiology AI devices authorized in the 85-day window (May 18–Aug 11, 2026) based on a confirmed rate of ~18–22 radiology AI devices/month through Q4 2025 and March 2026 data.
• Applying a three-stage filter: (1) ~60% qualify as ‘diagnostic’ after excluding hardware-embedded devices, RT planning tools, surgical guidance, and workflow-only devices; (2) ~36% of qualifying diagnostic devices are estimated to report patient-level (vs. finding/lesion-level) performance metrics, based on historical base rates (~20–24% sensitivity/specificity reporting across all AI devices through 2023) adjusted upward for the trend seen in 2024–2025 FDA submissions where metrics are ‘commonly disclosed.’ This yields a central estimate of approximately 12 qualifying devices.
• The 90% CI is roughly 3–21, reflecting uncertainty in the authorization rate, the strict definitional filters, and the patient-level vs. finding-level metric distinction.

Key Uncertainties:

• Patient-level vs. finding/lesion-level metric reporting rate has no direct empirical anchor; estimate of ~25-28% is interpolated between historical 24% sensitivity reporting and qualitative observations of higher recent transparency
• Fraction of radiology AI clearances that qualify as ‘diagnostic’ under the strict resolution criteria (excluding hardware-embedded, RT planning, surgical guidance, workflow tools) — estimated ~55% but not directly measured
• FDA authorization pace may shift in mid-2026 due to potential staffing changes or regulatory shifts; March 2026 data suggests pace is holding
• Strict ‘positive evidence required’ resolution rule means ambiguous summaries default to ‘not counted,’ which may further reduce counts beyond filter estimates

Evidence Sources (8):

• https://theimagingwire.com/2026/03/11/numbers-from-the-fda-show-radiology-is-maintaining-its-lead/
• https://innolitics.com/articles/march-2026-ai-samd-clearances/
• https://innolitics.com/articles/year-in-review-ai-ml-medical-device-k-clearances/
• https://www.medrxiv.org/content/10.64898/2026.05.08.26352766v1
• https://www.nature.com/articles/s41746-025-02052-9
• https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2833324
• https://www.medrxiv.org/content/10.64898/2026.03.25.26349197v1
• https://pubmed.ncbi.nlm.nih.gov/41004181/

Grounding Facts (passed to reviewers):

1. Today’s date is May 19, 2026.

Draft: 12: 7.2%, 13: 7.0%, 11: 7.0%, 10: 6.8%, 14: 6.8%

Reviewer Panel:

Reviewer (reasoning)

Reviewer (calibration)

{ “key_objections”: [

"The upward adjustment from a 20-24% historical rate of sensitivity/specificity reporting to a 36% rate for *patient-level*

Reviewer (fact-checker)

2 factual error(s). 2 unsupported claim(s). Verdict: significant_errors

Unsupported claims:

• ~36% of qualifying diagnostic devices are estimated to report patient-level (vs. finding/lesion-level) performance metrics
• patient-level vs. finding-level metric distinction

Objections:

• [minor] confirmed rate of ~18–22 radiology AI devices/month — Corpus says: Corpus shows 18.3/month in Q4 2025 and 19 in March 2026; no support for upper bound of 22
• [moderate] ~60% qualify as ‘diagnostic’ after excluding hardware-embedded devices, RT planning tools, surgical guidance, and workflow-only devices — Corpus says: Corpus states 63% of all cleared devices were diagnostic-purpose; no breakdown or exclusion list specific to radiology AI devices

Adjustment Rationale:

Adjusted central estimate downward from ~12 to ~9 in response to fact-checker and calibration reviewer concerns. The fact-checker correctly flagged that the 36% patient-level metric reporting rate was an unsupported upward extrapolation from the ~24% historical sensitivity-reporting base rate. Revised the patient-level metric filter to ~25-28%, closer to the historical base rate with only a modest uplift for the 2024-2025 trend. The strict resolution criteria — requiring positive evidence for every criterion and defaulting silent/ambiguous summaries to ‘not counted’ — further supports a more conservative estimate. Kept the ~55-60% diagnostic filter despite the fact-checker’s note that radiology-specific breakdowns aren’t directly available; the exclusion categories (hardware-embedded, RT planning, surgical guidance) are well-documented qualitatively even if exact percentages are estimates. Tightened authorization rate to 18-19 radiology AI devices/month (corpus-confirmed range). Calculation: ~53 total radiology AI × 0.55 diagnostic × 0.27 patient-level = ~7.9 → centered around 8-9 with wide uncertainty. Distribution remains right-skewed normal-like with mean ~9, std ~5.5, reflecting substantial compound uncertainty across three sequential filters.

forecast_sdk | 482.5s | research $0.8533 + reviewers $0.0094 + reconcile $0.1603 = $1.0230

Extracted Percentile Values: {0.1: 2, 1.0: 3, 5.0: 4, 10.0: 5, 15.0: 6, 20.0: 7, 25.0: 8, 30.0: 9, 35.0: 10, 40.0: 11, 45.0: 12, 46.0: 12, 47.0: 13, 48.0: 13, 49.0: 13, 50.0: 14, 51.0: 14, 52.0: 14, 53.0: 15, 54.0: 15, 55.0: 16, 60.0: 17, 65.0: 19, 70.0: 21, 75.0: 23, 80.0: 26, 85.0: 29, 90.0: 33, 95.0: 39, 99.0: 46, 99.9: 48}

Model’s Answer:

Analysis: FDA-Authorized AI Radiology Diagnostic Devices with Patient-Level Metrics (May 18 - August 11, 2026)

Understanding the Resolution Criteria

This question requires counting devices that meet ALL of the following:

1. Listed on FDA’s AI-Enabled Medical Devices List
2. Radiology (Panel Code 90) as lead review panel
3. Diagnostic intended use (detect, diagnose, classify, screen, identify)
4. Decision date between May 18, 2026 and August 11, 2026
5. Report at least one patient-level performance metric (sensitivity, specificity, AUC, accuracy, PPV, NPV) explicitly at patient/exam/case/study level

Estimating Base Volume

Authorization Rate Analysis:

• 2025 saw approximately 211-255 radiology AI devices cleared (sources vary slightly)
• This equates to roughly 17-21 radiology AI clearances per month
• The forecast window spans approximately 85 days (~2.8 months)
• Expected radiology AI authorizations: approximately 48-60 devices

Filtering Steps:

Step one: Filtering for diagnostic use. Not all radiology AI devices are diagnostic—many are workflow tools, triage-only systems, or measurement applications. Based on historical patterns, approximately 55-65% of radiology AI devices serve diagnostic purposes. This yields an estimated 27-39 diagnostic radiology AI devices.

Step two: Filtering for patient-level metrics. This is the most restrictive criterion. The JAMA Network Open analysis found only 55.9% of AI devices report any clinical performance studies, with sensitivity reported in ~36% and AUC in only ~16%. Critically, many devices report lesion-level rather than patient-level metrics. I estimate only 30-40% of diagnostic devices clearly report patient-level metrics meeting the strict criteria.

Scenario Analysis

Low Outcome (Why this could happen):

• FDA review slowdowns or fewer submissions in pipeline
• Higher proportion of non-diagnostic tools (workflow optimization, triage-only)
• Many devices using lesion-level rather than patient-level metrics
• Ambiguous summaries excluded per verification rules
• Estimated: 4-7 devices

Central Estimate:

• ~55 radiology AI authorizations × 60% diagnostic × 35% with patient-level metrics ≈ 12-14 devices

High Outcome (Why this could happen):

• Regulatory pressure (RAPID pathway, enhanced scrutiny) increases metric transparency
• Acceleration in radiology AI submissions
• Higher proportion of diagnostic applications
• Industry response to academic criticism improves reporting
• Estimated: 25-35 devices

Critical Assumptions and Failure Probabilities

Assumption 1: Authorization volume maintains 2025 pace

• Probability of significant deviation: ~25%
• Impact: Could shift estimate by ±5-8 devices

Assumption 2: Patient-level metric reporting rate is 30-40%

• Probability this is incorrect: ~35%
• If higher (50%+), could add 3-6 devices; if lower (<25%), could reduce by 3-5 devices

Assumption 3: Diagnostic vs. non-diagnostic split at 55-65%

• Probability of significant variance: ~30%
• Pipeline composition could shift this substantially

Final Distribution Rationale

Given substantial uncertainty across multiple filtering steps and the strict verification criteria (“ambiguous = doesn’t count”), I’m centering my distribution around 13-14 devices with wide uncertainty bands. The distribution is slightly right-skewed to account for the possibility that regulatory pressure has improved transparency in recent months. The explicit bounds of 1 and 49 require all predictions to be strictly within this range.

Percentile .1: 2 Percentile 1: 3 Percentile 5: 4 Percentile 10: 5 Percentile 15: 6 Percentile 20: 7 Percentile 25: 8 Percentile 30: 9 Percentile 35: 10 Percentile 40: 11 Percentile 45: 12 Percentile 46: 12 Percentile 47: 13 Percentile 48: 13 Percentile 49: 13 Percentile 50: 14 Percentile 51: 14 Percentile 52: 14 Percentile 53: 15 Percentile 54: 15 Percentile 55: 16 Percentile 60: 17 Percentile 65: 19 Percentile 70: 21 Percentile 75: 23 Percentile 80: 26 Percentile 85: 29 Percentile 90: 33 Percentile 95: 39 Percentile 99: 46 Percentile 99.9: 48

TL;DR

I forecast 8 qualifying devices, with a median of 8 and a 90% interval of 2 to 16.

Context

This is a narrow count, not a count of all radiology AI authorizations. A device must be on FDA’s AI-enabled list, have Radiology as lead panel, have a diagnostic intended use, and disclose at least one standalone patient-, exam-, case-, or study-level metric in an official public FDA summary or an explicitly incorporated predicate summary.

The FDA list is the right gatekeeper, but it is an imperfect source for this question. FDA says the list is based mainly on public authorization documents and classification, that linked summaries are not all-inclusive, and that devices with unpublished decision summaries can be added in later updates (FDA AI-enabled medical devices list). That matters because this resolves on public positive evidence by August 11, 2026.

Evidence

The historical backbone is the same calendar window in prior FDA AI-list vintages. The current public FDA page visible at forecast time is reverse chronological and runs through December 30, 2025; I counted lead-Radiology AI rows by final decision date in the May 18-August 11 window (FDA AI-enabled medical devices list). These are denominator counts before the diagnostic and patient-level metric filters.

The 2025 jump should get most of the weight. FDA-list reporting through December 2025 showed 1,451 AI-enabled medical devices since tracking began and 1,104 radiology devices, or 76% of the total; Q4 2025 alone had 55 radiology AI authorizations (The Imaging Wire, March 12, 2026). A separate April 2026 report said FDA cleared at least 255 radiology AI devices through 510(k) in 2025, which implies about 60 radiology devices in an 86-day window under straight-line timing (AuntMinnie, April 13, 2026). I set the 2026 window denominator at mean 66 and standard deviation 16. That is close to the 2025 same-window count, with room for update lag and year-to-year batching.

The qualifying fraction is much lower. A 2025 npj Digital Medicine review covered FDA AI/ML devices through December 2024: 1,016 devices were addressed by FDA, 1,012 had accessible summaries, 769 were radiology, and 522 of the 1,012 summaries, or 51.6%, reported no performance metric at all (npj Digital Medicine, 2025). The same study found sensitivity in 242 summaries, specificity in 220, AUROC in 110, PPV in 66, accuracy in 65, and NPV in 54 (npj Digital Medicine, 2025). This is an upper bound for this question because it counts many metrics that are not patient-level and does not exclude triage-only, measurement-only, workflow, reconstruction, hardware, or treatment-planning devices.

The product mix explains the discount. The 2025 window contains many Radiology AI rows with product codes tied to scanners, ultrasound systems, image reconstruction, treatment planning, measurement, dental finding analysis, segmentation, or triage. AuntMinnie’s summary of FDA’s CAD categories lists QAS and QFM as radiological computer-aided triage and notification or prioritization software, while QDQ/QBS are closer to computer-assisted detection or diagnosis software (AuntMinnie, April 13, 2026). That distinction is decisive here.

Two 2025 examples show the split. MammoScreen 4, cleared July 3, 2025, is a radiological computer-assisted detection/diagnosis device and reports mammogram-level AUC of 0.894 with a 95% CI of 0.870 to 0.919, which is the kind of exam-level metric that qualifies (FDA K243679 summary). Rapid CTA 360, cleared July 16, 2025, reports case-level sensitivity of 0.921 and specificity of 0.890 on 403 CTA cases, but the same summary says it is for triage and notification and is not intended to be used as a diagnostic device, so I exclude it under the fine print (FDA K251151 summary).

My final model is:

$$ N \sim \mathrm{NegBin}(\mu=66,\sigma=16), \quad q \sim \mathrm{Beta}(8.42,58.94), \quad X \mid N,q \sim \mathrm{Binomial}(N,q). $$

Here, N is the number of FDA AI-list lead-Radiology devices in the window, q is the chance a lead-Radiology AI device also passes the diagnostic-use and patient-level-metric tests, and X is the resolution count before the cap. The beta prior has mean 12.5% and standard deviation 4.0%. I chose 12.5% because broad public metric reporting is common enough to make 5% too low, but the actual resolution criteria remove most rows that look relevant on a naive keyword search. The model mean is 8.25, the mode is 6, the median is 8, the 10th to 90th percentile interval is 3 to 14, and the 5th to 95th percentile interval is 2 to 16. The probability of hitting the 50 cap is effectively zero.

What’s non-obvious

The easy mistake is to start from radiology’s 75%-plus share of FDA AI authorizations and expect a high answer. That overcounts. The FDA list is full of radiology AI that improves image acquisition, reconstruction, measurements, planning, segmentation, or workflow. Those devices can be valuable and still contribute nothing to this resolution.

The second trap is triage. Triage devices often publish exactly the metrics named in the question, including sensitivity and specificity. But if their indication says they are not diagnostic and only flag cases for workflow prioritization, the fine print excludes them. That single rule removes a visible cluster of recent high-metric summaries.

Limitations

The largest uncertainty is classification, not FDA volume. A future 2026 summary may use diagnostic words but report only lesion-level, tooth-level, vessel-level, or reader-aided metrics. Another may describe a strong standalone study while omitting the numerical metric from the public summary. Under this resolution, silence does not count.

The second uncertainty is FDA publication timing. FDA says AI-enabled devices with decision summaries not published inside a data-collection period can be incorporated into later updates (FDA AI-enabled medical devices list). A real late-window authorization could fail the question if its public summary or AI-list entry is not available by August 11, 2026.

The third uncertainty is product mix. If May-August 2026 is heavy in mammography, lung, prostate, or other diagnostic CADe/CADx submissions with explicit exam-level AUC or sensitivity/specificity, the answer can land in the low teens. If it is mostly scanner AI, reconstruction, triage, dental finding tools, segmentation, and planning software, the answer can land below 5.

Forecast: 14.99 devices

• Product code matters significantly. Devices with QDQ, POK, and MYN codes (CADe/CADx systems) regularly report patient-level sensitivity, specificity, and AUC metrics and clearly qualify. QAS/QFM codes (triage devices) also report case-level diagnostic metrics but have borderline “diagnostic use” status under the resolution criteria. QIH codes (image processing/measurement) typically report geometric metrics like Dice coefficient rather than patient-level diagnostic metrics and mostly don’t qualify.

• QIH devices are the main non-qualifying category. Recent evidence (AV Cardiac CT clearance) confirms QIH devices explicitly state they don’t perform automated diagnosis and report only measurement/visualization metrics. This clarifies that ~25% of radiology AI clearances likely fail both the diagnostic intent and patient-level metric requirements.

• Qualifying device types represent 50-60% of clearances. QDQ/POK/MYN codes alone account for ~28% of radiology AI clearances and have high metric-reporting rates. Adding QAS/QFM and other diagnostic codes brings the qualifying share to roughly half of all clearances, suggesting a central estimate around 15-18 devices from an expected ~50 clearances in the 85-day window.

• Metric reporting is routine but not universal. Recent 510(k) summaries (eyonis® LCS 1.1, Genius AI Detection 2.0) explicitly report patient-level metrics, but the resolution criteria require “positive evidence” in publicly accessible summaries—silent or delayed summaries don’t count.

• 2026 clearance pipeline appears active. Multiple Q1 2026 clearances have been confirmed, with no evidence of significant slowdown, supporting the ~50-clearance baseline.

• Substantial uncertainty remains around triage device eligibility, whether all summaries will be publicly available by August 11, and whether some QDQ devices

Forecast (percentiles, aggregated):

• p10: 3.0
• p25: 5.0
• p50: 7.0
• p75: 11.0
• p90: 16.0

Per-model forecasts:

• 1: p10=3, p25=6, p50=10, p75=16, p90=23
• 2: p10=2, p25=4, p50=7, p75=11, p90=17
• 3: p10=3, p25=5, p50=8, p75=13, p90=20

Rationale: Summary The median estimate of 7 qualifying devices is driven by a compounding series of strict technical and temporal constraints that drastically narrow the overall volume of FDA AI authorizations. The 86-day observation window from May 18 to August 11, 2026, would normally yield approximately 65 to 70 total radiology AI clearances, assuming a continuation of the roughly 21 to 25 clearances per month observed throughout late 2025 and early 2026. However, three critical bottlenecks significantly reduce the final count. First, the FDA typically experiences a 3- to 4-week administrative delay in publishing decision summaries after a device is cleared. Because the criteria require the summary to be publicly accessible by August 11, devices cleared in the latter half of July or early August are highly unlikely to qualify. This publication lag effectively shrinks the eligible timeframe to about 1.8 to 2 months, yielding an initial pool of roughly 35 to 45 devices. Second, the requirement for a “diagnostic” intended use excludes the vast majority of tools, such as triage-only (CADt), workflow optimization, and measurement-only software. True diagnostic applications (like CADe and CADx) historically represent only 25% to 35% of all radiology AI clearances, which leaves approximately 10 to 16 diagnostic devices. Finally, widespread inconsistencies in transparency and reporting further filter the pool. Research indicates that about half of AI devices fail to publicly report performance metrics in their decision summaries. Moreover, the metrics provided must be explicit, patient-level, and based on standalone AI performance rather than just lesion-level or human-plus-AI metrics. Applying these final strictures reduces the expected number of qualifying devices to around 7.

Strongest Arguments for Higher Values

• Rapid year-over-year acceleration in FDA AI medical device authorizations could push the baseline volume of monthly clearances much higher than historical averages by the summer of 2026.
• Heightened industry scrutiny and recent regulatory pushes for transparency could mean a much larger share of diagnostic devices now explicitly report patient-level metrics in their summaries.
• The FDA might process and publish decision summaries faster than the typical 3- to 4-week lag, expanding the effective time window and bringing late-July and August clearances into the final count.

Strongest Arguments for Lower Values

• The publication delay for FDA decision summaries frequently stretches beyond 4 weeks, which could severely truncate the eligible time window to just a few weeks of actual posted clearances.
• AI developers often focus heavily on triage (CADt) and workflow applications, meaning the proportion of true “diagnostic” tools could fall well below the historical 25% to 35% range.
• AI diagnostic tools commonly prioritize lesion- or finding-level metrics to highlight detection capabilities, rather than the explicitly labeled patient-, case-, or exam-level metrics demanded by the criteria.

Key Uncertainties

• FDA Publication Timelines: The variance in how quickly the FDA makes decision summaries publicly available is the largest swing factor, as persistent delays could easily halve the eligible pool.
• Metric Categorization Nuance: The strictness of requiring explicitly labeled “patient-level” and “standalone AI” metrics means many devices reporting general accuracy data might be disqualified on a technicality.
• Market Shift Toward Non-Diagnostic Tools: A potential industry shift toward administrative, workflow, and measurement tools—where regulatory hurdles are generally lower—could heavily dilute the pipeline of eligible diagnostic AI submissions.