MedTextSeg: A Deep Dual Sequential Model for Section Segmentation in Medical Reports

Segmenting clinical texts into sections that uncover the underlying content structure is an important NLP task as it can facilitate robust information extraction in the medical domain. Recently, supervised neural models have been proposed for this task, however, they fail to capture the multi-granural sequential dependency characteristic within the texts effectively. In this light, we introduce the MedTextSeg model, which learns the sentence representation by modeling both the local and global contexts from the surrounding sentences. Through both quantitative and qualitative evaluations, we show that the proposed sequential encoding of sentences improves the overall section label prediction performance on several real-world datasets compared to state-of-the-art methods.

I. Introduction

A significant amount of patient information produced in hospitals is documented as free-text narratives (e.g., clinical notes/reports) by the caregivers. These clinical texts contain important insights and observations made by the caregiver about the patient’s health trajectory, and are arranged into different sections. For example, information about a patient’s past medical conditions can be found in the "Clinical History" section of the report, while the caregiver’s observations and interpretations are likely to appear in the "Findings" section, as shown in the sample report in Figure 1.