Will AI Redefine Burn Triage? Unpacking the FDA's DeepView AI De Novo Order

Maya: Welcome to AI in Medicine - Smart Summaries. I am Maya, and along with my co-host Sam, we are AI-generated voices here to unpack a single public document for you today.

Sam: Today we are diving into a really interesting milestone from the U.S. Food and Drug Administration. The document is a De Novo classification order dated May 21, 2026, for a device called the DeepView AI System, submitted by Spectral MD, Inc.

Maya: This is a crucial read for both clinicians looking at the future of burn care and health-tech developers navigating regulatory pathways. The single most interesting finding here is that the FDA has officially established an entirely new regulatory category for software that uses artificial intelligence to predict burn wound healing potential without direct contact.

Sam: Let us start right at the top of the document. This is addressed to Trudy Estridge, the Senior Director of Regulatory Affairs at Spectral MD in Dallas, Texas. The FDA is granting their request for De Novo classification, which essentially means there was no legally marketed device on the market to base substantial equivalence on.

Maya: Exactly. The letter clearly outlines the regulatory mechanism. It states that under section 513(f)(2) of the Food, Drug and Cosmetic Act, any person who determines there is no legally marketed device to use as a comparison can request the FDA to make a risk-based classification under section 513(a)(1) of the Act, without having to first submit a 510(k).

Sam: The document gives some really helpful historical context on that process, too. It notes that the law was amended by the Food and Drug Administration Safety and Innovation Act on July 9, 2012, which created two options for De Novo classification. And then later, on December 13, 2016, the 21st Century Cures Act removed a requirement that a De Novo request had to be submitted within 30 days of receiving a not substantially equivalent determination.

Maya: That legislative history explains why Spectral MD could just go straight to the FDA. The document shows the FDA received this De Novo request on March 6, 2026, and by law, the FDA shall classify the device within 120 days of receiving such a request. They met that deadline with this letter dated May 21, 2026.

Sam: So they asked for a risk-based classification, and the FDA decided to place it into Class II. What exactly does this DeepView AI System do that makes it a Class II device?

Maya: According to the exact indications for use in the text, the DeepView AI-Burn software analyzes images obtained using the system. It uses artificial intelligence to detect deep burn tissue, which they specify as deep partial thickness and full thickness burns. It then provides segmentation of specific areas in a burn wound by highlighting areas that are unlikely to heal within 21 days with conservative treatment alone.

Sam: Wait, really? It actually draws a map on the image and highlights areas based on a specific 21 days healing threshold? That sounds like it could dramatically alter how a surgeon plans a treatment.

Maya: It does provide that map, but the FDA placed incredibly strict guardrails on how that map can be used. The document states that the absence of a highlighted region indicates the absence of deep burn tissue. It also calculates wound area measurement and percent total body surface area burn. But here is the critical limitation for the ward: the system is an aid to physicians in the management of thermal burn wounds and considering various treatment plans. It should be used in conjunction with clinical assessment and not as a stand-alone diagnostic device.

Sam: And the FDA gets even more specific about surgical planning. The letter explicitly states that the device's segmentation contour does not provide definitive delineation of the deep burn margin for precise excision sizing. It also says it is not for delineating the precise borders of non-healing tissue.

Maya: That phrasing is vital for any surgeon using this. You cannot just look at the AI-generated contour and trace your scalpel along it for precise excision sizing. The FDA is making it very clear that this is adjunctive information. It guides your clinical judgment, but it does not replace the physical boundaries you identify during the actual surgery.

Sam: Because they granted this De Novo request, the FDA also had to define an entirely new generic type of device. The regulation name they created is software-aided adjunctive diagnostic device for use by healthcare providers in skin wound assessment.

Maya: And the definition they wrote for this generic category gives us a clear look into the boardroom, specifically at how the FDA views the future of this technology. They define it as a device that uses a software algorithm to analyze optical or other physical properties of a skin wound without direct wound contact. It returns characterizing information that may include classification, severity, or healing potential of the skin wound.

Sam: That is a broad definition! Optical or other physical properties without direct wound contact. It sounds like they are leaving room for technologies beyond just standard digital photography, as long as it is non-contact. But because it is a Class II classification, that means there are special controls attached, right?

Maya: Absolutely. The document states that to classify it into Class I or Class II, the proposed class must have sufficient regulatory controls to provide reasonable assurance of safety and effectiveness. The FDA believes that Class II special controls, combined with general controls, provide that reasonable assurance. They outline a detailed table of identified risks to health and the mitigation measures required.

Sam: Let us walk through those risks, because if you are a developer building wound assessment AI, this table is essentially your mandatory to-do list. The first risk listed is false negative results, leading to delayed treatment or failure to treat appropriately. And right next to it, false positive results, leading to unnecessary referrals, medical therapy like medication, and unnecessary medical procedures like surgery.

Maya: To mitigate those false positive and false negative risks, the FDA requires clinical performance testing, non-clinical performance testing, and labeling. The next risk is false or inaccurate results, or failure to generate a result due to use error or improper device use.

Sam: For use error, the mitigation measures expand to include precision testing and human factors testing, along with labeling. Then there is a risk of device failure or malfunction causing inaccurate results. That requires non-clinical performance testing, precision testing, software verification, validation, and hazard analysis.

Maya: There is also a very physical risk mentioned: electrical, thermal, or light exposure-related injury. The examples they give are a burn or eye damage. That means this system likely uses a specific light source that could harm the patient if misused. The mitigation for that is electrical and thermal safety testing, software verification, validation, hazard analysis, and labeling.

Sam: Finally, the risk table lists interference with other devices, which must be mitigated by electromagnetic compatibility testing. After the risk table, the document dives deep into the exact special controls. There are seven numbered special controls, and they are incredibly detailed.

Maya: Special control number one focuses on clinical performance validation testing. The very first requirement under this control is huge for any AI developer. Data must demonstrate improvement of device-aided users' diagnostic characterization of the indicated wound compared to the accuracy of unaided users in the intended patient population and under anticipated conditions of use.

Sam: That means you cannot just prove the AI algorithm is highly accurate in a vacuum. You have to run a study showing that a clinician actually performs better when they have the AI output than when they are unaided. That is a massive usability and clinical workflow test.

Maya: Exactly. Furthermore, this clinical evaluation must include patients across risk factors representative of the intended use population. And it requires standalone device performance testing that demonstrates the accuracy of the device output relative to ground truth.

Sam: The phrase ground truth always brings up a lot of questions in AI. How does the FDA want companies to handle ground truth for burns?

Maya: The document specifies that justification must be provided for the determination of ground truth. They do not dictate what the ground truth must be, but you have to clinically justify whatever standard you use. The standalone testing must also include sensitivity and specificity of the device output, again with clinical justification of the reported results.

Sam: And for devices that provide mapping of the wound area, like showing wound borders or the extent of the wound, the testing must demonstrate segmentation accuracy. Wounds must be selected by representative users and include a justified quantity and range of wound types.

Maya: There is also a strict requirement for subgroup analysis. The analysis of standalone performance must include subgroup analysis by relevant risk factors. They want to ensure the algorithm does not just work well on average, but works consistently across different patient profiles within the intended use population.

Sam: Moving to special control number two, this covers non-clinical performance testing. It must demonstrate that the device performs as intended under anticipated conditions of use. Specifically, the device software must be tested for compatibility with specific signal or image acquisition hardware.

Maya: That compatibility testing must include a description of compatible hardware and processes, pre-specified compatibility testing protocols, and datasets. Also under non-clinical testing, performance testing must demonstrate the photobiological safety of any lamp or lamp systems. That ties directly back to the eye damage risk we discussed earlier.

Sam: Special control three is all about precision. Performance testing must demonstrate device precision, which includes repeatability and reproducibility of device performance across operators and challenging use conditions. You need to prove the device gives you the same answer if a different operator uses it or if the lighting in the room is not perfect.

Maya: Special control four covers electromagnetic compatibility and electrical and thermal safety of any electrical components. Special control five requires software verification, validation, and hazard analysis. And special control six is a human factors assessment to demonstrate that the device can be safely and correctly used by intended users based solely on the directions for use.

Sam: Based solely on the directions for use. That is a high bar. You cannot have a company representative standing over the doctor's shoulder explaining how to interpret the map. The instructions alone have to be enough to prevent use error.

Maya: That leads perfectly into special control seven, which is labeling. The labeling requirements are extensive. It must include a summary of both the standalone and clinical performance testing conducted with the device. That summary has to describe performance measures including sensitivity, specificity, and statistical confidence intervals.

Sam: And going back to those subgroups, the label must show the performance of the device for all clinically relevant subgroups within the intended use population. It also needs to provide information related to the limitations of device performance or subpopulations for which the device may not perform as expected.

Maya: The labeling must provide information needed to facilitate interpretation of all device outputs. It also requires a prominent statement that the device is not intended for use as a standalone diagnostic and is not for use to confirm a clinical diagnosis.

Sam: Finally, the labeling must include warnings to avoid unsafe exposure to any energy-emitting components of the device. The document explicitly gives an example of excluding use of the device on wounds close to the eye. It also notes that this is a prescription device and must comply with 21 CFR 801.109.

Maya: There is an interesting paragraph right after the special controls regarding combination products. The FDA advises Spectral MD that although the letter refers to their product as a device, some granted products may instead be combination products. They provide an email address, CDRHProductJurisdiction@fda.hhs.gov, if the company has questions about whether their product falls into that category.

Sam: That makes sense. If they ever packaged this software with a specific drug or biologic wound dressing, the regulatory landscape would shift again. The document also addresses what happens for future devices of this generic type. Under section 510(m) of the FD&C Act, the FDA can exempt a Class II device from premarket notification requirements if they determine it is not necessary to provide reasonable assurance of safety and effectiveness.

Maya: However, the FDA explicitly states in this order that they have determined premarket notification is necessary. Therefore, this generic device type is not exempt from the premarket notification requirements of the FD&C Act. Anyone who intends to market a software-aided adjunctive diagnostic device for skin wound assessment in the future must submit a premarket notification containing information on the software prior to marketing it.

Sam: So future companies will have to submit a 510(k), and they can use the DeepView AI System as their predicate device to demonstrate substantial equivalence, as long as they follow all these special controls we just discussed.

Maya: Correct. The letter also includes standard language reminding the company that granting this De Novo request does not mean the FDA has determined the device complies with other requirements of the FD&C Act or any other Federal statutes and regulations. They must still comply with registration and listing under 21 CFR Part 807, and labeling under 21 CFR Part 801.

Sam: They also have to comply with medical device reporting, which is the reporting of medical device-related adverse events, under 21 CFR 803. The letter mentions that if it is a combination product, postmarketing safety reporting falls under 21 CFR 4, Subpart B. There are also good manufacturing practice requirements set forth in the Quality Management System Regulation under 21 CFR Part 820.

Maya: And for combination products, current good manufacturing practices fall under 21 CFR 4, Subpart A. The letter even mentions that if applicable, they must follow the electronic product radiation control provisions under Sections 531 through 542 of the FD&C Act, which corresponds to 21 CFR 1000 through 1050.

Sam: One more major compliance note is the final Unique Device Identification System Rule, or UDI Rule. All medical devices, including Class I, unclassified devices, and combination product device constituent parts, must comply. The UDI Rule requires that a device bear a unique device identifier on its label and package unless an exception or alternative applies under 21 CFR 801.20(b).

Maya: The dates on the device label also have to be formatted in accordance with 21 CFR 801.18. Furthermore, the UDI Rule requires that certain information be submitted to the Global Unique Device Identification Database, or GUDID, under 21 CFR Part 830 Subpart E. The FDA provides a link to their UDI System webpage for more information on these requirements.

Sam: To make all of this official, the letter states that a notice announcing this classification order will be published in the Federal Register within 30 days. A copy of the order and supporting documentation are kept on file in the Dockets Management Branch in Rockville, Maryland, available for inspection between 9 a.m. and 4 p.m., Monday through Friday.

Maya: As a result of this order, the letter concludes that Spectral MD may immediately market their device as described in the De Novo request, subject to the general control provisions of the FD&C Act and the special controls identified in this order. The letter is signed by Julie Morabito, Ph.D., Director of the Division of General Surgery Devices.

Sam: That division is part of the Office of Surgical and Infection Control Devices, within the Office of Product Evaluation and Quality, at the Center for Devices and Radiological Health. They also provide a contact name, Scott Kominsky, and his phone number for any questions concerning the contents of the letter.

Maya: If we look at the big picture for a moment, this entire document represents a massive leap forward in standardizing how AI is evaluated for wound assessment. By establishing that the software must highlight areas unlikely to heal within 21 days with conservative treatment, the FDA is locking in a very specific clinical endpoint that future algorithms will likely be measured against.

Sam: And by enforcing that special control where you have to prove an aided user is more accurate than an unaided user, they are making sure companies cannot just build an algorithm that works in a lab. You have to prove it actually helps a doctor make a better assessment in the real world, under anticipated conditions of use.

Maya: But the strongest takeaway for any clinician listening is the boundary the FDA drew around surgical excision. The device calculates wound area and percent total body surface area burn, but its segmentation contour does not provide definitive delineation of the deep burn margin for precise excision sizing. It is an adjunctive tool, a guide, but not a replacement for surgical judgment at the margins of non-healing tissue.

Sam: That is exactly right. To quickly recap: The FDA granted De Novo classification to Spectral MD's DeepView AI System, making it a Class II device. It uses AI to analyze images without direct contact, detecting deep burn tissue and highlighting areas unlikely to heal within 21 days. The FDA created a new generic device category with strict special controls covering clinical validation against unaided users, human factors, precision, and photobiological safety.

Maya: And remember, future devices of this type are not exempt from 510(k) requirements and will need to meet these same rigorous special controls. You can find a link to the full FDA classification order in our show notes. Finally, please remember that this podcast is for informational purposes only and is not medical advice.