On Graph Construction for Classification of Clinical Trials Protocols Using Graph Neural Networks

Abstract

A recent trend in health-related machine learning proposes the use of Graph Neural Networks (GNN's) to model biomedical data. This is justified due to the complexity of healthcare data and the modelling power of graph abstractions. Thus, GNN's emerge as the natural choice to learn from increasing amounts of healthcare data. While formulating the problem, however, there are usually multiple design choices and decisions that can affect the final performance. In this work, we focus on Clinical Trial (CT) protocols consisting of hierarchical documents, containing free text as well as medical codes and terms, and design a classifier to predict each CT protocol termination risk as "low" or "high". We show that while using GNN's to solve this classification task is very successful, the way the graph is constructed is also of importance and one can benefit from making a priori useful information more explicit. While a natural choice is to consider each CT protocol as an independent graph and pose the problem as a graph classification, consistent performance improvements can be achieved by considering them as super-nodes in one unified graph and connecting them according to some metadata, like similar medical condition or intervention, and finally approaching the problem as a node classification task rather than graph classification. We validate this hypothesis experimentally on a large-scale manually labeled CT database. This provides useful insights on the flexibility of graphbased modeling for machine learning in the healthcare domain.

Deep learning-based risk prediction for interventional clinical trials based on protocol design: A retrospective study

“…Figure 10 sketches the general schema of the predictive risk assignment model. Given a CT protocol, a phase, and a condition (group), after a basic pre-processing step, these features are fed to two machine learning models based on the transformer 30 , 31 ( Figure 10 B) and GNN 27 , 28 , 32 ( Figure 10 C) architectures. During the training phase, these models learn the CT protocol representation conditioned to the phase and clinical condition, and at the inference phase they predict the CT risk.…”

“… 43 and Jumper et al., 44 as well as our recent work addressing a simplified scenario of CT classification. 27 , 28 …”

“…As an alternative, deep learning, 24 that is, artificial neural network architectures with several learning layers, has emerged as a paradigm based on automatic extraction of useful representations from data through multiple stages of processing, and with successful examples across digital medicine. 25 , 26 As for the use of deep learning for CT risk assessment, recent studies 27 , 28 , 29 have explored the use geometric deep learning, a branch of deep learning used for graph-based data, to predict whether a CT would successfully complete a particular phase. By exploiting the hierarchical nature of the CT protocol document, graph-based models provide significant performance improvements compared with models encoding the protocol using non-hierarchical representations.…”

Deep Learning-Based Risk Prediction for Interventional Clinical Trials Based on Protocol Design: A Retrospective Study

“…CT risk prediction models Figure 10 sketches the general schema of the predictive risk assignment model. Given a CT protocol, a phase, and a condition (group), after a basic pre-processing step, these features are fed to two machine learning models based on the transformer 30,31 (Figure 10B) and GNN 27,28,32 (Figure 10C) architectures. During the training phase, these models learn the CT protocol representation conditioned to the phase and clinical condition, and at the inference phase they predict the CT risk.…”

“…This choice is furthermore driven by many successful applications of GNNs in various fields, e.g., the works of Stokes et al 43 and Jumper et al, 44 as well as our recent work addressing a simplified scenario of CT classification. 27,28 The GNN-based CT risk prediction algorithm, as shown in Figure 10C, works by extracting raw vectorial features from individual textual pieces of the CT protocol and initializes the graph structure defined by a standard CT template with the vectorial features on their corresponding nodes. 27 The nodes of the graph will then propagate information between each other using L layers (L = 4 in our experiments) of the message passing algorithm.…”