Company Spotlight: Imvaria Fibresolve - AI for Lung Fibrosis with a Built-In Update Plan

Maya: Imagine being able to classify a severe lung disease like idiopathic pulmonary fibrosis using a routine CT scan as an adjunct prior to invasive testing, all while the manufacturer is legally pre-cleared to update and re-train the AI model over time.

Sam: That is exactly what we are unpacking today in our Company Spotlight on Imvaria's Fibresolve, based on their FDA 510(k) summary cleared on November 7, 2025. And as always, a quick reminder that our voices are AI-generated, and this is a summary of a publicly available FDA document, not company marketing material.

Maya: Right, we are looking strictly at the facts in the FDA clearance for Fibresolve with PCCP, which stands for Predetermined Change Control Plan, and seeing what it actually means for the ward and the boardroom.

Sam: Before we get into the really fascinating model update part, let us set the regulatory stage. This document comes from the FDA's Center for Devices and Radiological Health. It clears the device under the 510(k) number K252041.

Maya: Yes, it is cleared as a Class II device under the regulation number 21 CFR 892.2085, with the product code QWO. And like all these clearance letters, it includes standard boilerplate language stating the device is substantially equivalent to legally marketed predicate devices marketed in interstate commerce prior to May 28, 1976, or to reclassified devices.

Sam: The letter also notes it is subject to general controls, meaning things like annual registration, good manufacturing practice, and device labeling that must be truthful and not misleading. It even mentions compliance with the Unique Device Identification System rule, so it has to have a UDI on its label and package.

Maya: Exactly. So let us look at what this software actually does. According to the Indications for Use, Fibresolve is a software-only device that receives and analyzes lung computed tomography, or CT, imaging data. The goal is to provide a diagnostic subtype classification in suspected cases of interstitial lung disease, which is commonly abbreviated as ILD.

Sam: And the document is very specific about how this fits into the clinical workflow. It supplements the standard-of-care by providing this qualitative, diagnostic classification output based on machine learning pattern recognition. It says this is to provide adjunctive information as part of a referral pathway to an appropriate Multidisciplinary Discussion, or as part of an MDD itself.

Maya: Right, the MDD is a core part of lung disease management. But the most critical clinical detail here is what it is an adjunct for. The summary states specifically that the tool is used to serve as an adjunct in the diagnosis of idiopathic pulmonary fibrosis, or IPF, prior to invasive testing.

Sam: That phrase, prior to invasive testing, is huge. It frames exactly why a hospital might want this. But it is not a standalone doctor. The results are intended to be used only by clinicians qualified in the care of lung disease, specifically in caring for patients with ILD, in conjunction with clinical history, symptoms, other tests, and professional judgment.

Maya: Definitely. It does not replace the specialist. And there are strict eligibility criteria for which patient scans you can actually run through Fibresolve. First, the patient must be an age strictly greater than 22 years old.

Sam: That specific cutoff, greater than 22, tells us the pediatric and young adult populations are excluded from the intended use. And the second criterion is that the patient must have pulmonary symptoms suggestive of possible ILD including IPF.

Maya: You cannot just run this on every random chest CT. So, if a patient meets those criteria, the input must be a DICOM-compliant lung CT scan. The document then breaks the device down into 3 components to explain how it processes that scan.

Sam: Let us walk through the tech stack. The first of the 3 components is the Image Receiver API, which handles image acquisition in the cloud. The document says a hospital can either access the API directly via secure software integration, like from their PACS, or the images can be transmitted manually, for example by mail, to the manufacturer.

Maya: It is interesting they explicitly include mailing the images as an option. From there, the API passes the images to the second component, the Ingestion Pipeline and Analysis System. The pipeline identifies the applicable CT series, verifies it is valid, completes quality checks, and confirms it is adequate for analysis.

Sam: For the software developers listening, the summary actually gives us some great technical detail. It says the system is based on a machine learning model component and a Docker-based HTTP/1.1 Representational State Transfer software application programming interfaces. Basically, standard REST APIs for image transfer and analysis.

Maya: Once it passes those quality checks, it hits the core component, which they call the Fibresolve Model Inference Graph. The document describes this as an ensemble 3D deep learning model that was developed and trained using images from multiple facilities.

Sam: An ensemble model usually means multiple models working together to produce the final output. The algorithm development included pre-training, training to the disease target, architecture optimization, threshold determination, and validation. But there is a massive limitation here regarding what the output actually looks like.

Maya: Yes, this is vital for clinicians to understand. The document explicitly states that no segmentation is performed as part of the Analysis System. And in the comparison table, it reiterates that the device does not mark, annotate, or direct users' attention to a specific location in the original image.

Sam: So it is not drawing a box around the fibrosis. It also says there is no alteration of the original image. It just provides a binary result stating whether the data are determined to be suggestive for the target disease state. That data is then sent out by the third component, the Output API.

Maya: The Output API can integrate into a hospital's electronic health record for electronic transmission, or the manufacturer can transmit the report in a human-readable format, like via fax. The clinician then incorporates that report into their decision-making.

Sam: And the summary makes a point to say the system does not include an image viewer or visual output for diagnostic use. The clinical team has to review the source images using their facility's standard diagnostic viewer prior to submission, and the system does not replace imaging interpretation generally.

Maya: Exactly. Now, let us talk about how this device proved it was safe and effective to get FDA clearance. Because this is a 510(k), it is based on proving substantial equivalence to a legally marketed predicate device.

Sam: The predicate device here is the original Fibresolve, which was cleared under a De Novo pathway, number DEN220040. So this new clearance is essentially an update to that exact same predicate.

Maya: But this is where we have to point out what is missing from this document. Under the Summary of Non-Clinical Testing section, the text only says that Software Verification and Validation per IEC 62304 were performed to demonstrate safety based on current industry standards, and that the results indicate the subject device is equivalent to the predicate.

Sam: Right. Listeners who are used to seeing heavy clinical trial data in our deep dives need to know that there are absolutely no performance numbers in this 510(k) summary. It does not provide the size of the testing dataset, the demographic breakdown, or any sensitivity and specificity numbers. It simply states the testing showed equivalence.

Maya: Which means if you are a health-system leader evaluating whether to buy this software, you cannot rely on this public 510(k) summary for your clinical diligence. You would need to request the actual clinical validation studies from the company to see those real numbers. As the FDA letter reminds us, clearance does not mean the FDA evaluates contract liability warranties.

Sam: Exactly. But while it lacks those clinical numbers, this document contains something that is incredibly important for the digital health industry at large. The entire reason for this new clearance is that Fibresolve now includes a Predetermined Change Control Plan, or PCCP.

Maya: The PCCP is a game-changer. Normally, if a company makes a major change to an AI model, they have to submit a brand new 510(k). But under section 515C of the Act, the FDA allows companies to pre-specify the modifications they want to make and how they will validate them. If they stick to the plan, no new clearance is needed.

Sam: The summary breaks this down into SaMD Pre-Specifications, which are the specific modifications allowed, and the Algorithm Change Protocol, which outlines the process for data management, model re-training, and performance evaluation. And they list three specific modifications they are allowed to make. Let us go through them.

Maya: The first allowed modification is Model architecture modification. The rationale given in the document is that with advancements in underlying model architecture through new computer science developments, new model training allows for improvements in performance in terms of overall metrics as well as consistency and speed.

Sam: The testing method for that is statistical assessments following the same standards used in the original device clearance, ensuring non-inferior absolute performance. The benefit is enhanced performance and generalizability, but the listed risk is a reduction in clinical performance. To mitigate that, they have to evaluate the device model on test dataset metrics and execute unit and integration tests.

Maya: The second allowed modification is the Introduction of new training data. This is what everyone in AI wants. The summary states that with new training data, new model training allows for improvements in generalizability, robustness, and reductions in biases, which provides greater clinical value.

Sam: Again, the risk is a reduction in performance, and the mitigation is rigorous evaluation on test dataset metrics. The impact assessment explicitly notes this will result in revised sensitivity and specificity metrics for the system. That means the model's actual diagnostic accuracy can change over time through manufacturer updates, ideally for the better, without FDA re-review.

Maya: And the third allowed modification is perhaps the most interesting clinically. It is the Incorporation of ancillary inputs. The plan allows the addition of non-imaging ancillary inputs into the model to improve overall performance, with model component contributions displayed with the final result.

Sam: The document hints at exactly what those inputs will be. In the comparison table, under Software Inputs, the subject device lists the CT scan, but also lists, potentially age, sex, and pulmonary function tests, or PFTs. So they have pre-cleared a pathway to update this from a pure image-analysis tool into a multimodal AI.

Maya: And the benefit listed for adding those ancillary inputs includes enhanced performance and generalizability, but also explainability. If a clinician can see how a patient's pulmonary function tests contributed to the diagnostic result, that might make a binary output much more trustworthy on the ward.

Sam: But the FDA letter puts boundaries on this. It says if deviations from the established PCCP result in a major change or modification in the intended use, or could significantly affect safety or effectiveness, a new premarket notification would be required. Failure to submit it would constitute adulteration and misbranding.

Maya: It is a brilliant regulatory strategy. My memorable takeaway for health-system leaders is that AI procurement is no longer just buying a static algorithm. With a PCCP, you are buying a product where the manufacturer is legally authorized to update its architecture, train on new data, and add new clinical inputs over time without going back to the FDA for a new clearance.

Sam: A perfect wrap-up. That is it for our dive into the FDA 510(k) summary for Imvaria's Fibresolve with PCCP. We have linked the full public document in the show notes so you can read the details yourself. Remember, this podcast is for informational purposes only and is not medical advice. Thanks for listening.