The Rise of the Digital Health Technology–AI Coupled System in the AI Era of Drug Development

Abstract

The modern clinical investigation is undergoing a profound structural evolution. Previously, a comprehensive landscape analysis of regulatory submissions to the US Food and Drug Administration (FDA) during 2016 to 2021 demonstrated a rapid, exponential increase in the application of artificial intelligence and machine learning (AI/ML) across the drug development lifecycle.^{1, 2} At the same time, an expansion of digital health technologies (DHTs) in drug development is also observed, such as smart watches, phone applications, and other mobile technologies, enabling continuous capture of vital physiological and behavioral data real-time and remotely outside of traditional trial site.^3-5 However, these technologies are not advancing in isolation; many DHT applications have the potential to be intrinsically coupled with AI.

For the purposes of this commentary, we explore these coupled technologies as an AI–DHT coupled system. The AI–DHT coupled system describes the synergistic relationship between DHTs and AI in drug development. Within this emerging AI–DHT coupled system, DHT collects large amounts of data from patients remotely outside of traditional clinical trial sites, and AI can provide the necessary computational engine for signal processing, feature selection, and clinical translation. Together, they are driving a steady paradigm shift in drug development, gradually augmenting the traditional, episodic, site-centric model with a continuous, patient-centric model, as reflected in FDA's guidance on conducting clinical trials with decentralized element⁶ (Figure 1).

To fully appreciate the scope of this evolving relationship, it is essential to trace how the AI–DHT coupled system is embedded into the operational and clinical continuum of drug development. Rather than serving as isolated point solutions, these coupled technologies create a continuous thread of data generation and algorithmic translation in all stages of clinical trials. In reviewing recent regulatory submissions, the use of this AI–DHT coupled system can be mapped across four distinct chronological phases: optimizing initial trial enrollment, engagement, and protocol adherence; establishing baseline digital phenotypes for trial enrichment; conducting active safety surveillance; and developing novel digitally derived endpoints and biomarkers.

The initial impact of the AI–DHT coupled system is observed directly in promoting trial enrollment and expanding patient access to clinical trials. AI–DHT coupled systems can help patient enrollment in trials through automation of enrollment procedures or enabling timely enrollment of patients with acute and transient conditions, for example, patients with sepsis in emergency department, where the onset could take from a few hours to a few days. By leveraging the remote data collection capabilities established by DHTs, investigators can also enroll historically underrepresented or difficult-to-reach populations, such as individuals with limited mobility, patients with rare diseases, and those who live in rural settings, and potentially any individual who faces logistical hurdles in traditional clinical site visits thereby resulting in a high dropout rate.

AI algorithms can enhance traditional measurements by processing those data streams continuously, including daily activities, heart rates, and glucose fluctuations to monitor a participant in real-time and may offer future potential for tracking medication ingestion and metabolism. By identifying subtle behavioral patterns captured by DHT, AI models have shown the potential to identify a lapse in patient adherence to trial protocols or suggest early drop out, with the potential to enable proactive, remote clinician monitoring. This allows trial sponsors and investigators to deploy targeted retention and interventions strategies, such as in-person visits or dose adjustment, preserving consistent high data quality, and sustained patient engagement without fundamentally increasing the burden on the participant.

Beyond basic operational engagement, the AI–DHT coupled system is used increasingly for patient selection and identification. Recent submissions to FDA indicate a growing trend of leveraging baseline DHT data for patient inclusion/exclusion and enrichment. While digital phenotyping can rely on the continuous, passive tracking of daily activities via DHTs, baseline phenotyping and predictive/prognostic information can also be captured through single-time active tasks. Examples include targeted walking, speaking, or tapping assessments recorded via sensor-based mobile applications.

By applying AI-driven feature selection to these datasets, researchers can improve patient inclusion/exclusion and enrichment strategies for future studies. One primary application involves calculating digital scores to identify specific sub-populations whose digital phenotypes suggest a significant risk of adverse events or predict a high likelihood of responding to the targeted therapy, which can be used as exclusion/enrichment criteria. Another strategy leverages the high sensitivity of algorithmic processing to flag anomalous patients whose baseline digital readouts subtly deviate from classical pathological patterns. By excluding these atypical cases and refining both inclusion and exclusion criteria based on objective physiological data, the AI–DHT coupled system may reduce variability (noise) and enhance the statistical power and efficiency of the ensuing investigation.

In clinical trials, AI–DHT coupled systems elevate patient monitoring and safety surveillance. Continuous or high-frequency data capture via DHTs can offer a robust enhancement to traditional scheduled clinical assessments by enabling uninterrupted safety profiling between scheduled site visits. This can help ensure the safety of trial participants through early detection and intervention and provide more safety information about the product investigated.

More critically, the AI–DHT coupled system enables highly sensitive, real-time adverse event (AE) detection and mitigation. Deep learning models applied to continuous DHT data such as accelerometry, electroencephalogram (EEG), or continuous glucose monitoring (CGM) have been used to assess real-time safety risks, such as falls, seizures, or hypoglycemia, allowing the system with the potential to prompt patients to take precautions to prevent impending adverse events. In cardiovascular safety monitoring, AI-enabled electrocardiogram (ECG) is increasingly deployed in a continuous manner to swiftly detect and predict potentially dangerous QTc prolongation. Additionally, high-frequency reporting tools, such as electronic Patient-Reported Outcomes (ePROs) or eDiaries, can aid in systematically identifying fluctuations of patients’ habits such as diet and nutritional intake. A specific example of combining eDiaries with AI–DHT coupled system is use of event-driven eDiaries, which are digital tools used to capture patient-reported outcomes immediately following specific occurrences such as seizures or pain flares.⁷ Together, these continuous surveillance mechanisms can allow trial sponsors to proactively mitigate risks, helping to support patient well-being throughout the trial while providing safety insight.

The clinical realization of the AI–DHT coupled system lies in its potential to augment how therapeutic treatment efficacy is measured. Initially, clinical measures captured by DHTs were predominantly positioned within regulatory submissions as exploratory endpoints—useful for generating hypotheses but not intended for driving formal regulatory decisions. However, as the continuous data streams from DHTs are increasingly processed and interpreted by robust algorithmic frameworks, these novel digitally derived measures are being explored as secondary and primary clinical endpoints in trials.

A recent example for algorithmic validation in digital pathology is the PathAI developed AI-powered tool that received formal FDA qualification through the Drug Development Tool (DDT) Biomarker Qualification Program.⁸ While this platform operates on lab-based histological data rather than continuous remote sensors, it currently serves as the publicly facing example demonstrating how an AI-based evaluation within clinical investigations can be used to support drug development and approval.

While currently in early stages, other digitally derived measures are progressing. For instance, DHTs equipped with advanced physiological and biomechanical sensors are being utilized to quantify nocturnal scratching in dermatological trials. The data collected by these sensors uses AI to discern scratching measures from normal ambient sleep behavior. Translating this data into an endpoint in a clinical trial is a complex process that involves patient centricity, measurement science, and data science.

Similarly, advancements in psychiatric and neurological investigations are proceeding through FDA's Innovative Science and Technology Approaches for New Drugs program (ISTAND),⁹ which sits under the DDT Qualification Program. Rather than simply digitizing static questionnaires, trial sponsors are exploring the integration of sophisticated multimodal AI systems. Recent projects accepted into the ISTAND pipeline for neuroscience have focused on capturing and analyzing nuanced behavioral signals—such as vocal biomarkers, facial kinematics, and speech patterns—directly via DHTs as a potential substitute for what were previously clinician-reported outcomes.¹⁰ These AI-powered systems have the potential to address inconsistencies and subjectivity inherent to current existing assessments.

The integration of AI and DHTs is no longer an exploratory exercise; it represents a fundamental restructuring of clinical drug development. As demonstrated by the evolving landscape of regulatory submissions, the AI–DHT coupled system can bridge the evidentiary gaps between the episodic assessments performed in the controlled environment of the traditional trial sites. By promoting sustained trial engagement, rigorously refining patient inclusion and exclusion through baseline digital phenotyping, enabling proactive, real-time safety surveillance, and driving the derivation of highly sensitive novel digitally derived endpoints, AI–DHT coupled systems can significantly enhance both trial integrity and statistical power. To ensure long-term robustness as outlined in CDER's AI draft guidance, future efforts may be needed to address model drift by adapting continuous lifecycle management of AI–DHT systems in clinical investigations. As regulatory science continues to evolve to the influx of high-frequency, multidimensional data streams, the continued maturation of the AI–DHT coupled system to conduct more efficient clinically precise therapeutic investigations will be paramount.