Assumption-Agnostic Deep Learning Framework for Holistic Clinical Trial Monitoring

Recently developed machine learning–enabled approaches to quality tolerance limit (QTL) surveillance offer efficiencies over labourintensive source data verification. However, their reliance on parametric assumptions, restriction to Poisson or Bernoulli data types, and treatment of subject visits as independent observations limit their performance in the heterogeneous, longitudinal settings typical of modern clinical trials. We propose an assumptionagnostic framework for anomaly detection that combines two key components: (1) a hierarchical, nonparametric, multi-dimensional deviation scoring scheme based on discrepancies between reconstructed and observed values, and (2) a long shortterm memory (LTSM) autoencoder that learns the joint temporal distribution of all numeric variables. Together, these components enable continuous, centralized detection of anomalies—including QTL deviations—across all numeric data types and hierarchical levels (program, study, site, and participant). The framework ingests streaming data from electronic data capture systems, vendors, and embedded numeric representations of text, inferring a shared latent manifold that captures complex crossdomain relations and intrasubject dynamics without relying on predefined mappings or historical priors. As new data accrue, deviation score thresholds are continuously updated to maintain robust performance. We evaluated the framework using simulations modeled on real-world trial structures and anomaly patterns, along with a supporting case study. Results demonstrate substantial improvements in anomaly signal discrimination, significant reductions in unnecessary follow-up, and strong computational scalability. Anchored in the riskbased monitoring paradigm articulated in ICH E6(R2), the proposed framework provides sponsors with a practical toolset for earlier hazard detection, enhanced participant safety, and streamlined trial operations.