Digital Measures in Trials: Unpacking the FDA's New Framework

Sam: Imagine running a clinical trial where you do not have to wait for a patient to show up at the clinic to know if your new treatment is working.

Maya: Instead, you capture continuous data from an AI-enabled wearable while they go about their life. But regulators need proof that data is valid, which is exactly why we are looking at a document published in August 2026 by the U.S. Food and Drug Administration, titled Key Considerations for the Development and Use of Digitally Derived Measures for Clinical Investigations.

Sam: Just a quick reminder before we dive in. Our voices are AI-generated, and this podcast is a summary of a public document. But this framework is something every pharma, medtech, and digital health leader needs to understand if they want to use modern digital endpoints.

Maya: Absolutely. This document is a massive signal to the industry. It was jointly published by multiple branches of the FDA, including the Center for Biologics Evaluation and Research, the Center for Drug Evaluation and Research, the Center for Devices and Radiological Health, and the Oncology Center of Excellence.

Sam: So it is a completely unified front across drugs, devices, biologics, and cancer treatments.

Maya: Exactly. They are looking to support responsible innovations using a risk-based approach. They want to tailor the evidence requirements based on how a digital measure is actually going to be used.

Sam: Let us start with some basic vocabulary, because the FDA uses very specific terminology here. They talk about DHTs and DDMs. What is the difference?

Maya: A DHT is a digital health technology. That is the system itself. The computing platforms, the connectivity, the software, the sensors. This could range from general wellness applications to actual medical device functions under section 201(h) of the Federal Food, Drug and Cosmetic Act.

Sam: So the DHT is the physical tool or the AI software capturing the data.

Maya: Right. And then the DDM is the digitally derived measure. That is the actual metric or measure that you derive from the data collected by the technology. The FDA notes these measures have the potential to streamline clinical investigations because you might detect treatment effects or safety signals earlier than you would with traditional clinic visits.

Sam: Because you are tracking the person continuously, out in the real world. That is the dream for decentralized trials.

Maya: It is. But the FDA makes it very clear that you cannot just measure something because you have a fancy sensor. You have to start by measuring what is meaningful to patients.

Sam: I noticed that. The technical guidance starts with a very human focus. They use the phrase meaningful aspect of health. How do they define that?

Maya: They describe a meaningful aspect of health as any specific aspect of feeling or functioning in daily life that is important to patients. And you need to identify this early, and keep reevaluating it.

Sam: So you cannot just have clinicians sitting in a boardroom deciding what matters.

Maya: Exactly. The document specifically says determining this is not exclusive to clinicians and regulators. You need qualitative data from patients and caregivers. Their input is essential to make sure you are identifying the right aspects of the disease.

Sam: It makes sense. If you are designing an AI wearable for a specific disease, you need to understand who is actually going to wear it. The document mentions defining the target population.

Maya: Yes, they call it defining the context of use. You have to look at age, sex, and relevant co-morbidities. Those factors will influence how the technology performs, how usable it is, and what actually constitutes a meaningful change for that specific group.

Sam: To help figure that out, the FDA suggests developing a disease model. Basically mapping out how the condition affects bodily structures and processes, and how that translates to the patient's daily life.

Maya: Right. Which brings us to the next big term in the document: the concept of interest. The concept of interest is the specific aspect of an individual's clinical, biological, physical, or functional state that your assessment is trying to capture.

Sam: Wait, so how is a concept of interest different from a meaningful aspect of health?

Maya: The meaningful aspect of health is the broad focus, like being able to walk around your house. The concept of interest might be synonymous with it, but it often gets more specific about what might be affected by the medical product. To make this concrete, the FDA provides a recurring example throughout the paper.

Sam: The walking bout example for cardiovascular disease. Let us unpack that, because it really makes the whole framework click.

Maya: Okay, so if the meaningful aspect of health for a patient with cardiovascular disease is ambulatory function, you need conceptual clarity on what exactly you are measuring. The document says it might be important to define which aspects of walking are relevant. Is it bout frequency? Bout definition? Bout length? Or gait speed?

Sam: So instead of just counting total daily steps, which could just be tiny shuffling steps around the kitchen, you might define a concept of interest as the ability to walk continuously for a certain length of time. A walking bout.

Maya: Exactly. You need to build a clear justification for why bout length is a better measure of cardiovascular function than total step count alone. And once you have that concept of interest locked in, you have to prove your technology can actually measure it. That is where we get into the heavy lifting: verification and validation.

Sam: This seems to be the absolute core of the FDA's guidance. They say the technology and the measure must be considered fit-for-purpose. And getting to fit-for-purpose requires a very specific sequence of evidence.

Maya: Yes, and they break it down into distinct steps. Let us start with verification. Verification is basically confirming with objective evidence that the physical parameter is measured accurately and precisely.

Sam: So in our walking example, the physical parameter is just acceleration.

Maya: Right. Before you talk about steps or walking bouts, you just have a sensor. Verification means you run tests to confirm the accelerometer accurately captures raw acceleration values. You ensure the hardware functions properly.

Sam: Do trial sponsors have to build all that evidence from scratch?

Maya: Not necessarily. The FDA notes you can often leverage verification data made available by the device manufacturers, sometimes publicly or by right of reference. You can also demonstrate conformance to consensus performance standards if they exist.

Sam: Okay, so verification is proving the hardware measures the physics correctly. Then we move to validation, which the FDA splits into major parts. What is the first part?

Maya: The first part is analytical validation. This is proving the device appropriately assesses the clinical event or characteristic in your target population. So taking those raw verified acceleration values, and using an AI algorithm to accurately convert them into a step count.

Sam: And the document gives another really specific example here regarding stride length, right?

Maya: Yes. They mention a scenario where you want to measure the 50th percentile of stride length derived from steps taken over a specified duration, like a week. Analytical validation means providing evidence that the technology accurately measures both the strides and the stride length in that specific patient population.

Sam: And they point out that a disease might actually change how a patient takes a stride compared to a healthy person. You cannot just test your AI algorithm on healthy volunteers in a lab and assume it works for someone with advanced cardiovascular disease.

Maya: Exactly. You have to understand how the disease impacts the stride, and how your algorithm detects that from the raw sensor signal. Usually, this means comparing your digital measure against a reference measure.

Sam: But what if there is no reference measure? What if you are using AI to measure a totally novel digital biomarker that no one has ever captured before?

Maya: The FDA acknowledges that directly. If there is no appropriate reference measure for direct comparison to a novel concept, they suggest discussing how to best support the rationale and validity with the relevant FDA review division. Do not just guess. Talk to them early.

Sam: That is the recurring theme. Talk to the FDA review division prior to putting it in your trial protocol. Now, the document also talks about how much evidence you need for this analytical validation.

Maya: Right, the scope of the validation studies depends on the intended use. If your digital measure is going to be the primary endpoint in a pivotal study, the FDA expects prospective validation with pre-specified performance thresholds in your target population.

Sam: So you have to design a study just to validate the measure before you even use it for the drug trial.

Maya: Exactly. But if you are just using it as an exploratory endpoint, they say you might be able to rely on retrospective or bridging evidence. It is a sliding scale of risk.

Sam: The document also brings up a very practical headache for clinical trials: missing data. Wearables run out of battery, or patients take them off. How does the FDA want sponsors to handle that?

Maya: They want you to assess how much data is actually needed to provide reliable estimates during the analytical validation phase. For instance, how many minutes of data, or time spent walking, do you actually need to reliably estimate that 50th percentile of stride length?

Sam: If you need a specific amount of data to get a precise estimate, and the device only captures a fraction of that because it is uncomfortable, your measure fails.

Maya: Which leads perfectly into the next part of validation. Once you have analytically validated that the algorithm counts steps correctly, you move to clinical validation.

Sam: Okay, so what is clinical validation?

Maya: Clinical validation confirms that your measure actually tracks what it purports to measure conceptually, and that it reflects a meaningful aspect of health and tracks changes in the patient's clinical status.

Sam: Going back to our cardiovascular example, we verified the accelerometer, we analytically validated that it calculates walking bouts correctly. Clinical validation is proving that walking bout length actually tells us something real about the patient's heart disease.

Maya: Precisely. The FDA outlines specific ways to demonstrate clinical validity. One is whether the measure reflects a clinician's or patient's impression of disease severity. Another is if it reliably measures disease status when no changes are reported.

Sam: So if you have mild cardiovascular disease versus severe, the walking bout length should look distinctly different between those groups of known disease status.

Maya: Yes. And another way is to ask if the measure captures changes in disease status when a patient or clinician reports a change. If the patient says they feel much better this week, does the walking bout data show that improvement?

Sam: If it does, then the values generated correspond to the specific health experiences of the patient. That makes a lot of sense. But even if the math and the clinical theory are perfect, what if the patient just cannot use the device properly?

Maya: That is a huge section of this document. The FDA emphasizes the role of usability studies. They want evidence that users actually understand and can follow the instructions for use.

Sam: Human factors. A wearable is only as good as the person wearing it.

Maya: Right. And the rigor of these usability studies has to match the regulatory role of your measure. The user interfaces must accommodate the anticipated sensory, cognitive, language, and motor capabilities of the target population.

Sam: Which goes right back to the context of use. If your target population is older adults with motor impairments, you cannot have a device that requires complex interactions to activate.

Maya: Exactly. The FDA specifically notes populations like older adults, children, and patients with impairments. The technology has to be user-friendly, require minimal effort to complete tasks, have a low burden of data entry, and offer options for user control preferences.

Sam: They also mention evaluating calibration processes. Some devices require the patient or caregiver to calibrate them, like setting a baseline for individual stride length. That calibration process has to be validated too.

Maya: Yes, you have to ensure the calibration results in accurate measurements and figure out the appropriate frequency for it. They recommend cognitive interviews with patients regarding the device instructions, combined with pilot testing. You have to confirm they actually understand the training.

Sam: And if they are using this technology at home, the FDA says you should conduct a comparative evaluation between measurements obtained remotely and those obtained in a clinical setting using the same device.

Maya: Any differences between the home data and the clinic data must be justified. And for devices intended for prolonged use, sponsors have to evaluate adherence, user burden, sustained usability, and potential user fatigue throughout the whole study duration.

Sam: Because wearing a chest strap for a short time is very different from wearing it for the entire study duration. User fatigue is real. Let us talk about what else can go wrong. The document has a section on potential sources of error.

Maya: Yes, identifying potential sources of error is critical to maintaining the validity of the measure. The FDA says you need to start by deeply understanding how the sensors and algorithms work to determine what experimental conditions are needed to characterize these errors.

Sam: They list a few specific examples of errors that could ruin your trial data. Incorrect placement is a big one. Putting a wearable on the wrist instead of the hip.

Maya: Right, or environmental factors that change between the home and the clinic, like temperature or humidity. They also bring up interference that the AI might misinterpret. Their example is riding in a vehicle being misclassified as a walking bout.

Sam: If a patient is on a bumpy bus ride, and the algorithm thinks they are power walking, your cardiovascular outcome data is completely compromised.

Maya: Exactly. They also mention physical factors like user posture, motion, or the height difference between the sensor and the heart when measuring blood pressure.

Sam: What about the software itself? Tech companies update their algorithms constantly.

Maya: The FDA explicitly calls this out. If a technology, an AI algorithm, or a general computing platform is updated during a clinical investigation, the sponsors must ensure the device remains fit-for-purpose.

Sam: Even just an operating system update on a smartphone?

Maya: Even an operating system update. If the digital measure might be affected, it is helpful to validate the measurements again after the update is introduced. You can use previously collected data or a new prospective study, but you have to ensure the measurements did not change because of the software.

Sam: That is a massive consideration for tech companies used to pushing updates frequently. In a clinical trial setting, every update introduces regulatory risk. You have to lock it down or be prepared to re-validate.

Maya: And there is another source of error the FDA mentions: changes in the values over time in the absence of an intervention or change in disease activity. You have to consider baseline variability and trends. Otherwise, you might attribute a random change in the data to your drug working, when it is just natural fluctuation.

Sam: This whole document is really a roadmap. The FDA wants this technology to succeed. They say they believe a collaborative, patient-centered approach may help unlock the full potential of digital health technologies.

Maya: They do. But they are making it clear that a cool algorithm is not enough. For AI models specifically, developers should consider a credibility assessment commensurate with model risk specific to the proposed context of use.

Sam: Right, they reference another draft guidance on using AI to support regulatory decision-making. The overarching lesson for medtech and pharma leaders seems to be: map out your evidence generation strategy from physical sensor, to algorithm, to clinical meaning, before you enroll a single patient.

Maya: And document your justification every step of the way. Why this aspect of health? Why this digital measure over a traditional outcome? Why is it better? That justification can and should be updated over time.

Sam: Let us quickly recap what we have covered. The FDA's framework requires trial sponsors to define a patient-centered context of use. You then have to verify the physical sensor, analytically validate the algorithm's output, and clinically validate that the metric actually tracks the disease.

Maya: You also need to run rigorous usability studies tailored to your specific patient population, and relentlessly monitor for sources of error, from bumpy bus rides to smartphone software updates.

Sam: If you are building or buying digital endpoints, this document is required reading. We have linked the full source document in the show notes. As always, this podcast is for informational purposes and is not medical advice.