The false positive paradox: Examining real-world clinical predictive performance of FDA-authorized AI devices for radiology using clinical prevalence

This study analyzes FDA-authorized radiology AI devices to demonstrate how low disease prevalence creates a false positive paradox that undermines positive predictive value, arguing for the mandatory disclosure of false discovery and omission rates to guide clinically and ethically appropriate AI selection.

The Gist

The Big Idea: The "False Alarm" Trap

Imagine you buy a super-smart security system for your house. The company tells you it is 99% accurate. They say, "If a burglar is there, it will almost certainly catch them! And if no one is there, it will almost certainly stay quiet!"

You feel safe. But then, you move into a neighborhood where burglaries are incredibly rare—maybe only 1 in 1,000 houses gets broken into in a whole year.

Suddenly, your alarm starts going off every single day.

You rush outside, but there is no burglar. It was just a cat, a falling branch, or a gust of wind. You check your phone, and the app says, "Intruder Detected!" 365 times a year. Even though the system is "99% accurate," 99% of those alarms are wrong.

This is the False Positive Paradox described in the paper. It happens when a test is very good at finding a disease, but the disease is so rare that the test ends up screaming "DANGER!" at healthy people far more often than it actually finds sick people.

What the Paper Actually Did

The authors looked at 38 different AI tools used by doctors to read X-rays, CT scans, and MRIs. These tools are approved by the FDA (the US government agency that checks medical devices).

The companies selling these tools brag about their Sensitivity (how good they are at finding the bad stuff) and Specificity (how good they are at ignoring the good stuff). They say things like, "We are 95% accurate!"

The Problem:
The paper found that these companies often hide the most important number: How common is the disease in the real world?

• In the lab (The "Enriched" Test): The companies tested their AI on a pile of scans where they forced a lot of sick patients to be there (like putting 50 burglars in a neighborhood of 100 houses just to test the alarm). In this fake world, the AI looks amazing.
• In the hospital (The Real World): The AI is now used on regular patients. Most people are healthy. The disease is rare.

Because the disease is rare, the AI's "high accuracy" numbers turn into a nightmare of False Positives.

The "Cry Wolf" Effect in Radiology

The paper explains that when an AI flags a healthy patient as "sick," it causes a chain reaction:

1. The Doctor's Dilemma: The doctor sees the AI flag. Even if the doctor thinks, "That looks normal," they are terrified of missing a real case. If they ignore the AI and the patient does have the disease, the doctor could get sued.
2. Defensive Medicine: To be safe, the doctor orders more tests, more scans, or sends the patient for a biopsy.
3. The Cost:
   • For the Patient: Unnecessary anxiety, radiation exposure, and invasive procedures for something that wasn't there.
   • For the System: Wasted money and time.
   • For the Doctor: They start to lose trust in the AI because it's "crying wolf" too much.

The "Magic Math" Solution

The paper doesn't just complain; it offers a solution. It says doctors and hospitals don't need to wait for the AI companies to fix their reports. They can do the math themselves using a simple formula (Bayes' Theorem).

The Analogy:
Think of the AI's accuracy as a recipe.

• The AI company gives you the ingredients (Sensitivity and Specificity).
• But they forget to tell you the serving size (Prevalence).

If you bake a cake for 100 people using a recipe meant for 10, it's going to be a mess. The paper says: "Don't just look at the ingredients; look at the serving size!"

If a hospital knows that only 1% of their patients have a specific type of brain bleed, they can plug that number into the formula. They will instantly see that even with a "95% accurate" AI, 7 out of every 10 alarms will be false.

The Authors' Recommendations

The authors are asking the FDA and the AI companies to stop hiding the ball. They want:

1. Honest Reporting: Companies must report how common the disease is in the test data. If they tested on a "super-sick" group, they need to say so.
2. Real-World Math: Reports should show what the "False Alarm Rate" (False Discovery Rate) would be in a normal hospital, not just in a lab.
3. Choice: Doctors should be able to choose an AI setting that fits their needs. Sometimes you want to catch every possible disease (even if it means more false alarms). Other times, you want to avoid scaring healthy people. The data should let them choose.

The Bottom Line

Just because a medical AI says it is "99% accurate" doesn't mean it will work well in your local hospital. If the disease is rare, the AI will likely be wrong more often than it is right.

The takeaway: Don't trust the marketing slogan. Look at the math. If the disease is rare, expect a lot of false alarms, and make sure your hospital is ready to handle them without panicking.

Technical Summary

1. Problem Statement

The paper addresses a critical disconnect between the reported diagnostic accuracy of FDA-authorized Artificial Intelligence (AI) devices in radiology and their actual clinical utility.

• The Metric Gap: FDA submissions (510(k)) and vendor marketing predominantly rely on sensitivity and specificity as primary performance indicators. While these metrics describe a device's intrinsic ability to detect disease, they do not account for disease prevalence (the base rate).
• The False Positive Paradox (FPP): In clinical settings where disease prevalence is low, even AI systems with high sensitivity and specificity can generate a disproportionately high number of false positives. This leads to a low Positive Predictive Value (PPV) and a high False Discovery Rate (FDR).
• Clinical Consequences:
  • Base Rate Neglect: Clinicians may overestimate the reliability of an AI flag due to high sensitivity/specificity, ignoring the low prevalence of the target condition.
  • Defensive Medicine: High FDRs incentivize clinicians to order unnecessary follow-up imaging or biopsies to avoid legal liability associated with missing a case the AI "flagged."
  • Resource Strain: Unnecessary procedures increase patient anxiety, radiation exposure, and healthcare costs.
  • Data Enrichment Bias: Many FDA summaries use enriched datasets (artificially high prevalence) to stabilize accuracy estimates, resulting in inflated PPV and deflated FDR values that do not reflect real-world practice.

2. Methodology

The study employed a retrospective analysis of public FDA regulatory data combined with real-world clinical prevalence estimates.

• Data Collection:
  • Cohort: 38 FDA-authorized radiological AI devices (Product Codes QAS, QBS, QDQ, QFM) with final decision dates in 2024 and 2025.
  • Source: FDA 510(k) Premarket Notification Database and public decision summaries.
  • Extracted Metrics: Sensitivity, Specificity, ROC-AUC, reported PPV/NPV, and test set prevalence.
• Prevalence Data:
  • Clinical Base Rates: The authors utilized scan-level incidence data from a large private multisite practice (e.g., Pulmonary Embolism: 2.52%, Intracranial Hemorrhage: 6.66%, Aortic Dissection: 0.32%) and published literature to represent real-world prevalence.
• Statistical Analysis:
  • Re-calculation: Using Bayes' theorem, the authors recalculated PPV, NPV, FDR (1-PPV), and FOR (1-NPV) for each device.
  • Inputs: The recalculations used the vendor-reported sensitivity and specificity paired with the real-world clinical prevalence rates rather than the enriched test set prevalence.
  • Meta-analysis: Generalized linear mixed modeling (GLIMMIX) was used to estimate overall mean sensitivity, specificity, and ROC-AUC across the cohort.

3. Key Results

The analysis of 57 device entries revealed significant discrepancies between vendor-reported metrics and clinically anticipated performance.

• High Accuracy, Low Predictive Value:
  • The meta-analytic mean sensitivity was 92.6% and specificity was 90.8%.
  • However, when applied to low-prevalence conditions, the FDR often exceeded 50%.
  • Example (Large Vessel Occlusion - LVO): A device with 90.6% sensitivity and 88.8% specificity yielded a predicted FDR of 70.7% (PPV ~29%) in a real-world setting with 4.87% prevalence. This means ~71 out of 100 AI flags would be false alarms.
  • Example (Aortic Dissection): With a prevalence of only 0.32%, a device with 92.7% sensitivity and 92.8% specificity resulted in a FDR of 96.01% (PPV ~4%).
• Reporting Inconsistencies:
  • Only 10% of summaries explicitly reported PPV/NPV.
  • 43% provided enough data to derive PPV/NPV, but many relied on enriched datasets.
  • Some summaries reported performance of radiologists using the AI rather than the AI's standalone performance.
  • Few summaries reported sensitivity/specificity across multiple thresholds, limiting the ability of clinicians to adjust for their specific risk tolerance.
• Comparison of Vendor vs. Clinical Estimates:
  • In cases where vendors reported PPV based on enriched datasets (e.g., 40-50% prevalence), the values were drastically higher than the clinically anticipated PPV (often <5% for rare conditions).
  • Only a minority of summaries (e.g., K250248 for PE) used prevalence rates close to real-world data, resulting in PPV/NPV estimates that aligned with clinical expectations.

4. Key Contributions

• Quantification of the Paradox: The study provides empirical evidence that high sensitivity/specificity does not guarantee high PPV in low-prevalence radiology workflows, demonstrating the "False Positive Paradox" across a broad cohort of FDA-cleared devices.
• Methodological Framework: It establishes a reproducible method for clinicians to calculate local FDR and FOR using public 510(k) data (sensitivity/specificity) and local prevalence estimates, bypassing the need for proprietary internal validation.
• Identification of Reporting Gaps: The paper highlights the lack of standardization in FDA summaries, specifically the omission of test set prevalence and the failure to report performance at multiple decision thresholds.
• Actionable Recommendations:
  1. Mandatory Prevalence Reporting: Vendors must report the prevalence of the test set alongside sensitivity/specificity.
  2. Threshold Transparency: Sensitivity and specificity should be reported across multiple thresholds (or full ROC curves) to allow clinics to optimize the trade-off between FDR and FOR based on their specific clinical needs.
  3. Contextualized PPV/NPV: If PPV/NPV are reported, the base rate used must be explicitly stated and clinically defensible.
  4. Enrichment Warnings: Summaries using enriched datasets must explicitly warn that PPV/NPV cannot be calculated using the test set prevalence.

5. Significance

• Clinical Decision Support: The paper empowers radiologists and hospital administrators to make informed purchasing and implementation decisions by understanding the true "cost" of AI flags in their specific patient populations.
• Patient Safety: By anticipating high FDRs, practices can mitigate the risks of defensive medicine, unnecessary invasive procedures, and patient anxiety caused by false AI alarms.
• Regulatory Impact: The findings suggest a need for the FDA to update guidance on AI device submissions, requiring more transparent reporting of predictive values adjusted for real-world prevalence to prevent misleading marketing claims.
• Ethical & Financial Implications: The study underscores that without adjusting for prevalence, AI deployment could lead to inefficient resource allocation and potential legal liabilities for clinicians who rely on uncalibrated AI outputs.

In conclusion, the paper argues that diagnostic accuracy (sensitivity/specificity) is insufficient for clinical deployment decisions. True clinical utility requires a nuanced understanding of predictive performance (PPV/FDR) derived from the intersection of device accuracy and local disease prevalence.