Classification of hierarchical text using geometric deep learning: the case of clinical trials corpus

Abstract

We consider the hierarchical representation of documents as graphs and use geometric deep learning to classify them into different categories. While graph neural networks can efficiently handle the variable structure of hierarchical documents using the permutation invariant message passing operations, we show that we can gain extra performance improvements using our proposed selective graph pooling operation that arises from the fact that some parts of the hierarchy are invariable across different documents. We applied our model to classify clinical trial (CT) protocols into completed and terminated categories. We use bag-of-words based, as well as pre-trained transformer-based embeddings to featurize the graph nodes, achieving f1-scores 0.85 on a publicly available large scale CT registry of around 360K protocols. We further demonstrate how the selective pooling can add insights into the CT termination status prediction. We make the source code and dataset splits accessible.

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

“…To benefit from the power of end-to-end deep learning and to avoid using handcrafted approaches, our recent work [7] targeted the CT classification problem using GNN's. This was motivated by the highly hierarchical structure of CT's as shown in Fig.…”

“…While an important difficulty to do machine learning on such representations is the high dimensionality imposed by the number of the tokens of the collection, one can benefit from their high sparsity to project them to much lower dimensions, as e.g. in [7], where a very practical setup has been implemented suitable for the classification task and with low latencies.…”

“…x 7 x 8 All CT's as one graph: The above setup considers each protocol independently, so a graph-classification task can be performed to classify them, as in our prior work [7].…”

“…In the second abstraction, we further require them to have the same trial phase. To benchmark this idea against the case where CT's are considered entirely independently (e.g., as in [7]), we consider these super-nodes as an empty graph, i.e., without any edges between them, while of course considering all the nodes and edges within the super-nodes.…”

“…While the use of GNN's has been shown to be very effective, in this work, we show that the way the graph is constructed and the problem is formulated is also of prime importance. In particular, we build up on our prior work [7], where we encode Clinical Trials (CT) protocols in a hierarchical graph to predict their termination risk, by further linking them with edges according to connections between the CT phases, conditions and interventions. Therefore, rather than considering multiple disconnected graphs and posing the problem as graph classification, we consider a single but very large graph and target a node classification problem to classify CT protocols.…”