Comparing generative and extractive approaches to information extraction from abstracts describing randomized...

Systematic reviews of Randomized Controlled Trials (RCTs) are an important part of the evidence-based medicine paradigm. However, the creation of such systematic reviews by clinical experts is costly as well as time-consuming, and results can get quickly outdated after publication. Most RCTs are structured based on the Patient, Intervention, Comparison, Outcomes (PICO) framework and there exist many approaches which aim to extract PICO elements automatically. The automatic extraction of PICO information from RCTs has the potential to significantly speed up the creation process of systematic reviews and this way also benefit the field of evidence-based medicine.

Previous work has addressed the extraction of PICO elements as the task of identifying relevant text spans or sentences, but without populating a structured representation of a trial. In contrast, in this work, we treat PICO elements as structured templates with slots to do justice to the complex nature of the information they represent. We present two different approaches to extract this structured information from the abstracts of RCTs. The first approach is an extractive approach based on our previous work that is extended to capture full document representations as well as by a clustering step to infer the number of instances of each template type. The second approach is a generative approach based on a seq2seq model that encodes the abstract describing the RCT and uses a decoder to infer a structured representation of a trial including its arms, treatments, endpoints and outcomes. Both approaches are evaluated with different base models on a manually annotated dataset consisting of RCT abstracts on an existing dataset comprising 211 annotated clinical trial abstracts for Type 2 Diabetes and Glaucoma. For both diseases, the extractive approach (with flan-t5-base) reached the best $F_1$ score, i.e. 0.547 ( $\pm 0.006$ ) for type 2 diabetes and 0.636 ( $\pm 0.006$ ) for glaucoma. Generally, the $F_1$ scores were higher for glaucoma than for type 2 diabetes and the standard deviation was higher for the generative approach.

In our experiments, both approaches show promising performance extracting structured PICO information from RCTs, especially considering that most related work focuses on the far easier task of predicting less structured objects. In our experimental results, the extractive approach performs best in both cases, although the lead is greater for glaucoma than for type 2 diabetes. For future work, it remains to be investigated how the base model size affects the performance of both approaches in comparison. Although the extractive approach currently leaves more room for direct improvements, the generative approach might benefit from larger models.