Authors (including presenting author) :
Lui SM (1), Chan CH (1), Woo PS (1), Tsang CO (2), Tsui LH (1), You J (3), Yu LH (4)
Affiliation :
(1) Statistics & Data Science Department, Hospital Authority, (2) Department of Surgery, LKS Faculty of Medicine, The University of Hong Kong, (3) Department of Statistics & Actuarial Science, The University of Hong Kong, (4) Department of Mathematics and Information Technology, The Education University of Hong Kong.
Introduction :
In the past, the use of unstructured data such as clinical text in research studies has been greatly limited as manual extraction approach is time-consuming and costly, or even infeasible when large number of cases are involved. In this big data era, text analytics is the state-of-the-art technique to capitalize on these valuable data in huge volume, as illustrated by a research study [1] to build a deep learning model to predict large vessel occlusion stroke.
Objectives :
To develop a text analytics tool which can automatically abstract and transform into structured data those clinical notes’ free text related to stroke signs and symptoms in HA’s CMS, which are essential risk predictors for the core study during its model validation; informing its deployment and other research studies.
Methodology :
The study subjects were a sample of 645 acute ischemic stroke patients during 2016- 2018; admitted to HA’s hospitals and with a plain CT scan within 24 hours of A&E admission. A neurosurgeon independently reviewed all the clinical notes of study subjects and manually extracted GCS score and the side of limb/facial weaknesses as the reference standard. Respectively 300 and 345 sampled patients were randomly split to train and validate the model. The notes were first processed by text pattern matching algorithm to extract all corresponding values, followed by the XGboost classification model to determine which value represents the onset value closest to A&E admission timing. An agreement rate was calculated for each symptom to indicate the concordance between the model’s output and the reference standard.
Result & Outcome :
The text analytics model achieved 90%-96% of agreement with the reference standard based on the testing dataset. This model was then applied to a whole year cohort of 6,543 acute ischemic stroke patients for abstracting key risk predictors’ data to enable implementing the core study’s [1] deep learning model in HA Data Collaboration Laboratory. The data automatically extracted from clinical notes for this study and research results from the core study will inform planning of additional structured data items in CMS. These text analytics techniques can also be applied to other studies to mine past clinical free text. [1] TSUI E, TSANG A. Keynote Lectures: Data Analytics & Applications in Hong Kong Hospital Authority: Past, Present & Future at Health Research Symposium 2019; June 12, 2019. Hong Kong
