Assumption-Agnostic Deep Learning Framework for Holistic Clinical Trial Monitoring.

Area of Science:

• Clinical data science
• Machine learning in healthcare
• Regulatory science

Background:

• Current machine learning (ML) methods for quality tolerance limit (QTL) surveillance in clinical trials have limitations.
• These include reliance on parametric assumptions, data type restrictions (Poisson/Bernoulli), and treating visits as independent, hindering performance in complex trials.

Purpose of the Study:

• To propose a novel, assumption-agnostic framework for anomaly detection in clinical trials.
• To enable continuous, centralized detection of anomalies, including QTL deviations, across diverse data types and hierarchical levels.

Main Methods:

• A hierarchical, non-parametric, multi-dimensional deviation scoring scheme.
• A long short-term memory autoencoder to learn joint temporal distributions of numeric variables.
• Ingestion of streaming data from various sources, inferring a shared latent manifold without predefined mappings.

Main Results:

• The framework demonstrated substantial improvements in anomaly signal discrimination.
• It significantly reduced unnecessary follow-up and showed strong computational scalability.
• Evaluated through simulations based on real-world trial data and anomaly patterns.

Conclusions:

• The proposed framework offers a practical tool for earlier hazard detection and enhanced participant safety.
• It aligns with the risk-based monitoring paradigm (ICH E6(R3)).
• Enables streamlined clinical trial operations through advanced anomaly detection.