.png)
To tell the truth, traditional clinical trials are expensive, slow, and frequently followed by problems that were visible in the data before anyone acted on them. A participant whose wearable data shows declining activity and disrupted sleep for two weeks before withdrawing was signaling something. A data gap that started as two missed device syncs before becoming a protocol deviation was visible the day it started. A safety signal that showed up in heart rate trends three days before the scheduled site visit could have been caught sooner.
The gap between when a signal appears and when it gets acted on is where trial integrity breaks, timelines slip, and costs compound. For research organizations, AI-powered patient monitoring changes when those signals get noticed, and by whom.
Traditional site monitoring catches problems after the fact. A clinical research associate reviews data during a site visit and identifies a protocol deviation that happened three weeks ago. An adverse event gets flagged when a participant reports it at a scheduled visit rather than when the physiological indicator first appeared. A dropout is recorded when a participant stops responding, not when their engagement pattern started to suggest they were heading in that direction.
Risk-based monitoring (RBM) improved the targeting of monitoring resources toward higher-risk sites and participants, but it still depends largely on aggregated data reviewed on a lag. The fundamental constraint has not changed: monitoring looks backward at what already happened.
Decentralized and hybrid trial designs have made this gap more visible, not less. When participants are not coming to a clinical site every two weeks, the window between observations widens. Without a mechanism to monitor what is happening between scheduled touchpoints, the first sign of a problem in a decentralized trial is often the dropout itself.
Trials that include wearable or remote monitoring components have access to a data stream that traditional monitoring models were not designed around: continuous physiological and behavioral signal collected daily, outside the clinic, for every participant in the study.
This changes the monitoring question from "what happened at the last visit" to "what has been happening every day." The practical implications run across three areas.
A participant whose resting heart rate has been trending upward for five days, whose sleep has become fragmented, and whose HRV has dropped below their individual baseline is showing a physiological pattern worth investigating. These signals often precede a clinical event or adverse report by days. When monitoring systems can see them in continuous data, the investigation can begin before the participant reaches the severity threshold that would prompt a report at the next visit.
Activity-based endpoints, sleep-based endpoints, or any protocol requirement tied to behavioral patterns can be tracked in near real time rather than verified retrospectively through participant self-report at a visit. The gap between what participants do and what they report doing is a known limitation of self-reported outcomes in trials, which shrinks when objective continuous data is available.
Find more details on integrating wearable data in clinical trials!
Declining engagement, reduced device wear time, and increasing variability in daily patterns are behavioral signals that often precede dropout. In trials with demanding protocols, the data can reveal which participants are struggling with compliance before they formally withdraw, creating an intervention opportunity.
The volume of continuous wearable data generated in a trial with hundreds or thousands of participants is not reviewable by a monitoring team looking at dashboards manually. AI is what converts that volume of signal into a prioritized, actionable set of alerts.
AI applies anomaly detection models trained on each participant's own longitudinal baseline. A resting heart rate of 78 is not inherently concerning. For a participant whose baseline over the prior three weeks is 58, it warrants attention. Personalized baseline models identify deviations that fixed thresholds would not catch, and they reduce false positives by calibrating alerts to the individual rather than the population.
Combined with multi-signal pattern recognition, AI monitoring surfaces combinations of signals that have historically preceded adverse events rather than reacting to a single out-of-range value. Relevant signal combinations include:
AI uses behavioral and physiological patterns to identify participants at elevated risk of withdrawing before they do. Patterns associated with pre-dropout states include:
An AI model trained on historical trial data can assign a dropout risk score to each active participant on a rolling basis, enabling proactive site outreach to participants who are showing early disengagement signals before they become withdrawals.
AI distinguishes between different types of missing data, each of which has different implications for protocol integrity. A gap in wearable data can mean:
These are not equivalent from a data quality or protocol deviation standpoint. AI classification of gap type, using device-specific sync patterns, battery data, and user behavior signals, allows monitoring systems to route each gap to the appropriate response: a participant nudge, a site alert, a device replacement workflow, or a pipeline investigation. Catching gaps on day two rather than day fourteen is the difference between a recoverable data quality issue and a protocol deviation that affects endpoint analysis.
Generating AI signals is only half of the monitoring problem. The other half is integrating those signals into the operational workflow of a running trial in a way that CRAs, site coordinators, and data managers can act on them.
The most effective implementations treat AI monitoring output as a risk-tiered alert layer sitting above the continuous data stream. Alerts are classified by severity and type:
Each alert carries the underlying data that triggered it, so the reviewer has context rather than just a notification.
Integration with existing trial management infrastructure matters here. Alerts that require manual re-entry into a separate system, or that are surfaced in a tool the monitoring team does not use in their daily workflow, get ignored. Effective AI monitoring connects to the systems where trial operations already live.
Escalation thresholds and alert configurations need to be defined before the trial starts, not tuned reactively during the study. What constitutes a safety-relevant deviation for this population and this protocol is a clinical and regulatory question. The AI layer executes against those definitions. Getting the configuration right upfront is what separates a monitoring system that generates actionable signal from one that generates noise.
AI monitoring at trial scale only works if the underlying data infrastructure is built to support it. Several requirements are non-negotiable.
Most trials with wearable components allow participants to use their own devices or select from a provided list. That means heart rate, HRV, sleep, and activity data arriving from multiple manufacturers with different sensor technologies, sampling rates, and metric definitions. AI models trained on normalized data produce reliable signal. Models trained on inconsistently normalized data produce systematic biases that may not be apparent until endpoint analysis.
Every data point that informs a monitoring alert needs a traceable provenance:
Regulators reviewing a trial dataset need to be able to follow that chain. This is not a post-hoc documentation exercise, it needs to be built into the data architecture from the start.
Continuous wearable data collected in a clinical trial is special category health data under GDPR Article 9. Participants need to give explicit, purpose-specific consent for the data to be used in monitoring workflows, and that consent needs to be recorded in a way that supports regulatory audit. Pseudonymization of participant identifiers needs to be maintained across the full data lifecycle, including within the monitoring alert system.
AI-generated signals in a clinical trial are not, in themselves, clinical decisions. They are inputs to human review. Documenting that distinction clearly, in the protocol, the statistical analysis plan, and the monitoring plan, is important for regulatory submissions under EMA, FDA, or other frameworks. How AI monitoring outputs are characterized, logged, and handled in the event of a triggered alert needs to be specified before the trial starts.
Thryve's Multimodal Clinical Trial Platform is built for research teams who need continuous wearable data collection at scale, with the data infrastructure required to support AI-powered monitoring alongside it.
The platform connects 500+ wearables and health data sources, normalizes 250+ health metrics across devices, and delivers a unified longitudinal dataset for every participant, structured for both endpoint analysis and real-time monitoring use cases. Key capabilities relevant to AI patient monitoring include:
The infrastructure problem in AI-powered trial monitoring is not the algorithm. It is the data layer underneath it: normalized, traceable, pseudonymized, and governed to the standard that regulated research requires. That is where Thryve operates.
If you’re curious about wearable data in clinical trials, download our “Between Two Tests” White paper on how wearables can help fix engagement and retention problems in the subscription diagnostic business.
If you are designing a clinical trial that includes wearable data collection and want to understand what the monitoring infrastructure looks like in practice, book a demo with Thryve.
Friedrich Lämmel is CEO of Thryve, the plug & play API to access and understand 24/7 health data from wearables and medical trackers. Prior to Thryve, he built eCommerce platforms with billions of turnover and worked and lived in several countries in Europe and beyond.
.png)
.png)