What AI Can't Fix In Decentralized Clinical Trial Data
By John Oncea, Chief Editor, Clinical Tech Leader
AI is the part of the decentralized trial technology stack getting all the attention right now. Yasir Shafiullah, a Ph.D. researcher in RF and analog IC design at the University of Oulu’s CWC-RF department, thinks that attention is arriving at the wrong layer.
“I do think AI can help,” Shafiullah said, describing its potential role in validating measurement data, flagging when a reading changed abruptly, or when a sensor appears to have moved outside expected conditions, tasks currently managed by a human reviewing data after the fact. He sees real upside there: faster identification of sensor errors, faster flagging of anomalies, and less reliance on catching problems only after a full dataset has already been collected.
Where The Limits Are
But Shafiullah was direct about where AI’s usefulness stops. It can’t distinguish a hardware problem from genuine physiological variation if the underlying data was never trustworthy to begin with. The analog front-end limitations, sensor placement errors, timestamp drift, and packet loss discussed elsewhere in this stack all happen before AI ever touches the data. No model, however sophisticated, can recover information a sensor never captured accurately in the first place.
That’s the throughline across this entire technology stack: data quality is decided at the hardware and network layers, long before it becomes a training input or an analysis output. AI can help interpret and validate what arrives; it can’t retroactively fix what was wrong when it left the sensor.
Availability Isn’t The Same As Integrity
Asked directly whether a complete dataset can still be an unreliable one, Shafiullah didn’t hesitate: yes. A patient moving during a measurement, poor skin contact, or a sensor removed overnight can all produce data that’s fully present and still not trustworthy. His answer for closing that gap wasn’t a single tool; it was a mindset: check data over time for whether it behaves consistently and makes physiological sense, and build in a feedback loop rather than assuming a complete dataset is automatically a valid one. “We cannot automatically guarantee that the data we receive is valid,” he said. “Problems can come from the network, the device, or human behavior.”
Realistic, Not Dismissive, About The Hype
Shafiullah’s view on AI’s broader trajectory is measured rather than skeptical. He uses AI himself for automation in his own IC design work, generating code, automating measurement workflows, and pointed to early research suggesting AI may eventually assist with analog front-end design itself, with engineers supervising rather than executing every step manually. His caution is specifically about sequencing: “I think AI is useful, but we should not blindly jump into it... maybe we are overestimating it in some ways, but it still has value.”
For decentralized trials specifically, that translates into treating AI as the layer that interprets and validates data other engineering decisions have already made trustworthy, not a substitute for getting those decisions right.
Questions Worth Asking Instead Of “Is It AI-Enabled?”
For sponsors and CROs evaluating remote monitoring platforms, Shafiullah’s perspective suggests a different vetting checklist than the one most vendor conversations default to. Rather than leading with AI capabilities, it’s worth asking how sensor drift and placement errors are detected, what happens to data during a connectivity interruption, how timestamp synchronization is managed across device types, and what validation methods exist before AI or analytics ever touch the dataset. He also pointed out two practical, non-technical factors sponsors underweight: confirming realistic connectivity and latency at the deployment site, and making sure staff are properly trained on sensor use and data transmission, problems no algorithm can compensate for.
The common misconception worth retiring, in Shafiullah’s view, isn’t really about AI at all. It’s the assumption that wearable technology is inherently dependable simply because consumer brands have made them feel familiar and low risk. Consumer adoption and clinical-grade trustworthiness are different questions and confusing them is where a lot of the industry’s confidence in decentralized data outpaces the engineering reality.
Together with “Why Trustworthy Clinical Trial Data Starts as Analog, Not Digital,” “The Synchronization Problem Clinical Research Isn’t Watching,” and “Wireless Connectivity Is a Data Integrity Issue, Not an IT Issue,” this piece completes the full stack Shafiullah walked through: physics, timing, transmission, and finally, interpretation. The order matters, and so far, most of the industry’s attention has been aimed at the last step instead of the first three.