Company Spotlight: Tyto Care Tyto Insights - The FDA's Blueprint for Updating AI

Maya: Imagine an AI tool that listens to your lungs at home and flags a wheeze with a specificity of 0.99, meaning almost zero false alarms, but a sensitivity of just 0.55.

Sam: Today we are looking at exactly that tool, by diving into the public FDA 510(k) summary for K252844, published June 2, 2026, by the US Food and Drug Administration. The device is Tyto Insights for Wheeze Detection, and this submission introduces something called a PCCP.

Maya: Before we dive in, we must mention that our voices are AI-generated. This episode is a summary of a publicly available document from the FDA. It is not promotional material, and we have no ties to the company.

Sam: Right. This is an independent breakdown for our listeners who are building or buying health AI. We are looking at a product from Tyto Care Ltd, based in Netanya, Israel. Maya, let us start with the basics. What exactly is this product, and who is it designed for?

Maya: According to the indications for use, this is an over-the-counter artificial intelligence enabled decision support software system. It is used to evaluate lung sounds in adults and pediatrics, specifically patients 2 years and older.

Sam: Over-the-counter? So, anyone can just use this without a prescription? That sounds like a massive shift for respiratory diagnostics.

Maya: Yes, the intended users include professional users in a clinical environment, but also lay users between 18-65 years old in a non-clinical home environment. However, the document is very clear on its limits. It explicitly states it is not intended to detect other abnormal or normal lung sounds, only a specific abnormal lung sound suggestive of a Wheeze.

Sam: Just wheezes. And since it is over-the-counter, does it just give the patient a diagnosis right there on their phone? Do they just look at the screen and say, 'Okay, I have asthma'?

Maya: No, and that is a critical point for clinicians and health-system leaders listening. The document specifically states a licensed health care professional's advice is required to understand the meaning of the result. Healthcare providers are instructed to consider the device result in conjunction with the recording and other relevant patient data.

Sam: That makes total sense for a decision support tool. It is not replacing the doctor; it is giving them another data point. Let us talk about the architecture. The document breaks down the system into a few different subsystems. How does the audio actually get from the patient's home to the AI?

Maya: It requires a specific piece of hardware. The software processes recordings from the FDA-cleared compatible Tyto Stethoscope, cleared under 510(k) K181612. The recordings are sent by a third-party point of care app to the clinician app through a telehealth server.

Sam: So it is not just any microphone, it has to be that specific stethoscope.

Maya: Exactly. That telehealth server then sends the audio to the Tyto Insights Web Server using an API. The actual brain of the operation is the Algorithm Server. It receives the audio file, runs the machine learning neural network, and returns a result.

Sam: And what are the possible results it can spit back out to the user?

Maya: It provides these outputs: Positive, meaning wheeze suspected. Negative, meaning wheeze not suspected. Or, a message saying the system was not able to analyze the recording.

Sam: Why wouldn't it be able to analyze it? Does the summary explain that?

Maya: It does. It mentions constraints on the signal length. Recordings are processed in segments of up to 12 seconds, and the document explicitly states that signals shorter than 6 seconds will not be processed.

Sam: Got it, so you need at least 6 seconds of continuous, clear audio. Now, let us get to the numbers you teased at the very beginning of the episode. How well does this neural network actually perform when put to the test?

Maya: The performance evaluation section gives us the real numbers. The Area Under the Curve, or AUC, is 0.96. The specificity is 0.99. But the sensitivity is 0.55.

Sam: Okay, let us unpack that for a second because that 0.55 sensitivity stands out. If I am interpreting that correctly, if you have a dataset of recordings that truly contain a wheeze, the AI will only flag them at a rate of 0.55. It is missing a pretty significant chunk of actual wheezes.

Maya: That is exactly what a sensitivity of 0.55 means. However, the flip side is that specificity of 0.99. Specificity reflects the true negative rate among cases that truly do not have a wheeze. So, if we look only at the recordings where a patient's lungs are genuinely clear, the algorithm correctly identifies them as negative at a rate of 0.99. False positives are very rare in that population.

Sam: It is tuned to be very conservative. It does not want to cry wolf. But for a pediatrician relying on this for remote monitoring, they need to know that a negative result does not completely rule out a wheeze. Does the 510(k) summary tell us anything about the dataset they used to get those numbers? Like, how many patients were in the trial, or where they were located?

Maya: No, it does not. The document does not provide the size of the test dataset, the patient demographics, or the study design for the original clearance. It only gives us those final statistical estimates and their 95% confidence intervals.

Sam: That is a great reminder for our listeners about what a 510(k) summary actually is. This is not a full clinical trial report or an academic paper. It is a regulatory document proving substantial equivalence to an existing device on the market.

Maya: Exactly. And in this case, the predicate device is an earlier version of itself. It was compared to the Tyto Insights for Wheeze Detection, cleared under 510(k) number K232237 on December 13, 2023. The summary states that the intended use and technological characteristics are identical to that predicate.

Sam: If it is identical to the December 13, 2023 version, why did they file a new 510(k) submission in 2026? What actually changed?

Maya: The submission highlights specific changes. First, there was a minor software update where third-party libraries were updated to address known vulnerabilities. This aligns with their software development lifecycle and cybersecurity procedures. But the summary notes this minor update did not impact device performance.

Sam: So they patched some cybersecurity vulnerabilities. That is standard. But no new clinical validation was needed for that?

Maya: Right. The document states the clinical validation and human factors validation previously conducted for the predicate device remain applicable. No additional validation was required for the software patch.

Sam: Okay, so a routine cybersecurity patch. But that brings us to the next change, which is the really exciting part for the medtech developers listening. The new device name includes the phrase in parentheses: with PCCP. What exactly is a PCCP?

Maya: PCCP stands for Predetermined Change Control Plan. The FDA's approval letter explains that under section 515C(b)(1) of the Act, a company does not need to submit a new premarket notification for a change to a cleared device, as long as that change is consistent with an established PCCP.

Sam: That is huge. Normally, if a company wants to retrain their machine learning algorithm with new data to make it better, they have to go through the whole FDA clearance process again, which takes time and money. This plan is essentially a pre-negotiated contract with the FDA saying, 'if we follow these exact testing rules, we can push an update without asking for permission first.'

Maya: Exactly. But the FDA letter is careful to warn that if deviations from the established PCCP result in a major modification to the intended use, or could significantly affect safety or effectiveness, then a new submission is required under 21 CFR 807.81(a)(3).

Sam: And if they fail to submit when they are supposed to?

Maya: The letter is very strict on that. Failure to submit such a premarket submission would constitute adulteration and misbranding.

Sam: So the stakes are high. You have to follow the plan exactly. The summary actually lists the exact modifications they are allowed to make under this plan. Let us walk through them because this is a masterclass in how to write a PCCP. What is the first allowed change?

Maya: The first is modifications related to performance specifications enhancement. This means re-training the ML model with additional data to improve the clinical performance, measured by sensitivity and specificity, while maintaining the same type and range of input signal.

Sam: And how are they required to test that new, retrained model to prove it is still safe?

Maya: They have to run a retrospective study on a dataset of compatible stethoscope recordings that are representative of the intended patient population. The data has to be acquired by real-world use of the stethoscope. They must evaluate software verification and validation, and then look at the stand-alone accuracy.

Sam: What are the hard success criteria? They cannot just say it feels better clinically. What numbers do they have to hit?

Maya: The document lays out a very strict rule. The modification will be considered successful only if the lower bound of the 95% confidence interval for both sensitivity and specificity is not lower than the corresponding lower confidence bounds of the most recently validated and released version.

Sam: So they can never step backward. If the current version's lower bound for sensitivity is 0.43, any future model's lower bound has to be 0.43 or higher. They also have secondary endpoints listed for this retraining, right?

Maya: Yes, they have to descriptively evaluate the Positive Predictive Value and Negative Predictive Value, comparing the newly trained algorithm to the prior one. They also have to evaluate the performance across specific clinically relevant subgroups.

Sam: What subgroups are they required to track?

Maya: The document lists age group, specifically 2-18 years vs. >18 years, gender, user type, meaning lay user versus healthcare professional, and stethoscope model.

Sam: That is incredibly comprehensive. It ensures the algorithm does not suddenly get worse for children while getting better for adults. Okay, that is model retraining. What is the next type of allowed modification in the PCCP?

Maya: The second is modifications related to technical performance specifications. This covers tweaking data preprocessing, data augmentation, and hyper-parameters to improve running time, memory consumption, and CPU utilization.

Sam: So making the code run faster or cheaper in the cloud, basically. But without changing the core AI.

Maya: Correct. These changes must be implemented without modifications to the core model architecture, intended use, or output of the device. And the success criteria for testing that are the exact same co-primary endpoints. Both sensitivity and specificity lower bounds cannot drop.

Sam: And there is one final allowed change in the PCCP table. What is that one?

Maya: Modifications related to device input expansion. This allows them to expand the algorithm to include new sources of the same signal type. Specifically, different models of FDA compatible stethoscopes with equivalent audio acquisition specifications.

Sam: Ah, so if Tyto Care releases a newer model of their physical stethoscope hardware, they do not need a whole new FDA clearance for the AI software to analyze its audio.

Maya: Right. As long as the new hardware is FDA cleared for over-the-counter use, they just run the retrospective dataset test and prove the sensitivity and specificity hold up. The subgroup analysis must evaluate performance for each compatible stethoscope model.

Sam: This whole document is basically a template for how AI companies can navigate adaptive regulation. But Maya, one thing I noticed in the FDA clearance letter is that the FDA reminds the company about the Quality Management System Regulation, or QMSR. What does that mean for these updates? Do they just update the code and ship it?

Maya: Not at all, and that is a crucial point for companies to understand. The FDA letter states that regardless of whether a change requires premarket review, the QMSR requires device manufacturers to review and approve changes to device design and production.

Sam: So they still have to follow all the internal compliance steps.

Maya: Exactly. It specifically cites ISO 13485 clause 7.3 and ISO 13485 clause 7.5. You must document changes and approvals in the Medical Device File. You are still fully accountable for nonconforming product procedures, and corrective and preventative actions.

Sam: The PCCP cuts red tape at the FDA submission desk, but it does not eliminate the rigorous engineering and documentation work inside the company. Before we wrap up, I want to touch on the Reference Device mentioned in the document. We talked about the predicate device, which was the older wheeze detector. But they also list a reference device called Tyto Insights for Rhonchi Detection. Why include that?

Maya: The document notes that the predicate wheeze device from 2023 did not include a PCCP. However, the Rhonchi detection device, cleared on April 07, 2025 under 510(k) K243567, did include an FDA authorized PCCP.

Sam: So they used the Rhonchi device as a precedent.

Maya: Yes. Tyto Care cited the Rhonchi device to demonstrate that the types of software modifications, data management practices, re-training practices, and update procedures they are proposing for the Wheeze device are consistent with a framework the FDA has already authorized for them.

Sam: They basically said, 'Hey FDA, you already agreed to this exact testing plan when we built our Rhonchi algorithm in 2025, we just want to apply the same rules to our Wheeze algorithm.' That is very smart regulatory strategy.

Maya: It certainly is. And the FDA agreed, concluding that the proposed modifications under the PCCP do not raise new questions of safety and effectiveness, and determined the device is substantially equivalent.

Sam: Alright, let us bring it all together. As a crisp recap, today we looked at the 510(k) summary for Tyto Insights for Wheeze Detection with PCCP. It is an over-the-counter AI software that analyzes 6 to 12 seconds of lung audio to detect wheezing.

Maya: And while it boasts a specificity of 0.99, its sensitivity is 0.55. This makes it a highly specific decision-support tool, not a standalone diagnostic that replaces clinical judgment.

Sam: And more importantly for the industry, the document outlines a Predetermined Change Control Plan that allows the company to retrain their machine learning model, optimize performance, and support new stethoscope hardware without filing a new 510(k), as long as they prove through rigorous retrospective testing that the statistical performance does not drop.

Maya: It provides a clear roadmap for post-market algorithm updates.

Sam: A quick reminder that this episode is not medical advice, and we encourage you to read the primary source yourself. You can find a link to the original FDA source document in our show notes. Thanks for listening to AI in Medicine - Smart Summaries, we will catch you next time.