Exploring Multi Feature Optimization for Summarizing Clinical Trial Descriptions

Documenting Clinical Trial Descriptions of patients can help doctors with diagnostics and treatment plans and can be used for future reference. However, with the rapid growth of population, manually checking all previous files of a patient is not feasible. We address this challenge by providing summaries of clinical trial descriptions. We present a framework for automatically summarizing Clinical Trial Descriptions, which takes advantage of different features in semantic and syntactic space. We propose a multi objective optimization technique which uses position and similarity of the sentences. The similarity is established based on TF-IDF and WMD. We evaluate the proposed method on clinical trial dataset, and compare the results against human gold standard summaries using ROUGE metrics. Our approach is unsupervised in nature which has an advantage over supervised models in the advent of new diseases where a large quantity of quality data is not available. We provide a detailed ablation study to show the contribution of each feature in our approach and release our code on GitHub.¹

I. Introduction

The recent outbreak of COVID-19 has catalyzed the development of novel coronavirus vaccines. However, the creation of any vaccine or drug takes much time because of the multiple steps involved, and one of the significant steps after the development of the vaccine is clinical trials. Clinical trials are a type of research that studies new tests and treatments and evaluates their effects on human health. Clinical Trial summarization is a task of creating salient, non-redundant and condensed variants of trial descriptions. Extractive summarization methods using sentence features are popular. The informational sentences are selected from long descriptions using sentence scoring features. Some of the sentence scoring features are sentence length, position, similarity and dissimilarity with title and Maximum Marginal Relevance(MMR)[1]. It has been shown in the literature survey that the chances of a sentence being in the generated summary is higher if the sentence is longer, placed closer to the title, and with less word mover distance (WMD) from the title than the other sentences.