Evaluation of a Data Annotation Platform for Large, Time-Series Datasets in Intensive Care: Mixed Methods Study

TL;DR

This study evaluates a specialized annotation platform for large clinical time-series datasets in intensive care units, highlighting challenges, participant engagement strategies, and the potential for semi-automated annotation methods to improve data labeling efficiency.

Contribution

The paper introduces and assesses a bespoke annotation tool with dual modes, providing insights into clinician engagement and semi-automated annotation effectiveness in a complex clinical setting.

Findings

Significant recruitment and engagement challenges were identified.

Interventions improved participant engagement during the study.

Participants showed preferences for different parameter types and agreement in semi-automated annotations.

Abstract

Intensive Care Units are complex, data-rich environments where critically ill patients are treated using variety of clinical equipment. The data collected using this equipment can be used clinical staff to gain insight into the condition of the patients and provide adequate treatment, but it also provides ample opportunity for applications in machine learning and data science. While this data can frequently be used directly, complex problems may require additional annotations to provide context and meaning before it could be used to train the machine learning models. Annotating time-series datasets in clinical setting is a complex problem due to a large volume and complexity of the data, time-consuming nature of the process and the fact that clinicians' time is in both high demand and short supply. In this study, we present an evaluation of a bespoke tool designed to annotate large, clinical time-series datasets with staff from intensive care units. The software incorporates two modes for annotation: by annotating individual admissions and by generating rulesets which are applied to the entire dataset. Our study was split into two stages focusing on individual and semi-automated annotation and included 28 annotators across both stages who utilised 50 clinical parameters to guide their annotations. We experienced significant challenges in recruitment and engagement of the participants in the annotation activities and developed interventions which improved the participation over the course of the study. During the semi-automated annotation, we observed preferences for different parameter types (measured vs. observed), as well as relative agreement of participants across shared admissions to the decision-tree model trained using their rulesets.