| dc.description.abstract | Stroke is one of the major health problems and the second leading cause of human death in the world. The disability sequelae caused by stroke is one of the main causes of adult disability in Taiwan. Stroke-associated pneumonia (SAP) is an important clinical problem in the prognosis of patients with acute stroke (AIS). Most stroke patients will suffer from varying degrees of mobility impairment. Dysphagia for example, seven times increased the risk of aspiration pneumonia. Thus, pneumonia is the most common respiratory complication which occurs in one- third of patients with AIS. SAP is closely associated with increased long-term mortality, prolonged hospital stays, increased healthcare costs, and decreased prognostic function.
The main purpose of this study is to add deep learning and text mining techniques to extract new risk factor that may affect SAP in unstructured electronic medical records (EMR). And then apply them to construct a model for predicting the risk of pneumonia complicated by AIS patients during hospitalization. It assists doctors in diagnosing accurately.
This study used Chinese EMR and stroke registration database from 2007-2017 from Chia- Yi Christian Hospital. In total, 941 eligible patients with AIS were used to build and evaluate the models. The unstructured data used six techniques for feature engineering, including TFIDF, Doc2Vec, MetaMap, Bidirectional Encoder Representations from Transformers (BERT), BERT for Biomedical Text Mining (BioBERT), BERT for Clinical Text Mining (Bio_Clinical Bert). Eight machine learning methods including support vector machine, simple Bayesian classifier, K-Nearest Neighbor algorithm, logistic regression, Decision Tree, Random Forest, Extreme Gradient Boosting, and Light Gradient Boosting Machine were implemented for developing models and comparing predictive result.
The results show that adding unstructured text features and combining structural features to construct prediction model achieved better performance than model constructed by simply using structural features or unstructured text features. In addition, Deep learning feature engineering technology achieved better embedding effect better than traditional feature engineering methods. Using the BioBERT and Bio_Clinical Bert model which is pre-trained based on biomedical corpus and medical clinical records as the unstructured EMR feature engineering technology for this task enabled subsequent modeling to achieve better performance than using BERT general corpus- based model. In all investigated classification techniques, extreme gradient boosting is the most suitable machine learning classifier for the prediction model of post-stroke pneumonia. | en_US |