Adverse Drug Event Information Extraction from Medical Narratives using Ensemble Learning & Deep Learning

Wunnava, Susmitha

Etd

Adverse Drug Event Information Extraction from Medical Narratives using Ensemble Learning & Deep Learning

Public

An adverse drug event (ADE) is an injury resulting from medical intervention related to a drug. Many ADEs are detected only during the post-marketing phase of the drug when it is used by a more diverse population than during clinical trials. Early detection of the ADE incidents is crucial for timely assessment, mitigation and prevention of future occurrences of ADEs. Natural Language Processing (NLP) techniques towards ADE information detection from medical narratives provides an effective way of post-marketing drug safety monitoring and pharmacovigilance. My dissertation studies the problem of detecting ADE information from medical narratives at different levels of granularity: word-level, sentence-level and multi-grained (word-level + sentence-level) using supervised machine learning techniques. In this dissertation research, we first propose an Ensemble learning approach for fine-grained word-level information detection. Existing supervised machine learning approaches towards biomedical Named Entity Recognition (NER) are limited in their ability to identify certain entity types and result in significant performance difference in terms of accuracy. Another critical problem faced by NER in the biomedical context is that the data is highly skewed for these challenging entity types. We propose a novel methodology called Tiered Ensemble Learning System with Diversity (TELS-D) to address the above challenges in NER. We propose a balanced, under-sampled bagging strategy that is dependent on the level of imbalance to overcome the class imbalance problem. Next we propose an ensemble of heterogeneous recognizers approach that leverages a novel ensemble combiner. Second, we propose Sequence labeling for word-level information detection using deep learning. Although Electronic health records (EHR) contain valuable ADE information, the EHR text tends to be noisy and comprised of medical and non-medical abbreviations, acronyms, numbers, misspelled words and semantic type ambiguity among certain named entities - making it difficult to detect critical information. We propose the Dual-Level Embedding for Adverse Drug Event Detection framework (DLADE) by adapting a three-layered, deep learning RNN architecture of (1) Bi-directional Long Short-Term Memory (Bi-LSTM) for character-level word representation to encode the morphological features of the medical terminology, (2) Bi-LSTM for capturing the contextual information of each word within a sentence, and (3) Conditional Random Fields for the final label prediction by also considering the surrounding words. In addition, we propose a rule-based EHR text preprocessor for transforming the EHR text into clean tokenized text input essential for the success of the subsequently applied computational detection method. Our proposed NER system was ranked first in the MADE1.0 NLP Challenge for Detecting ADE information from EHR. Third, we propose a multi-grained joint modelling approach for word-level and sentence-level information detection using deep learning. Existing ADE detection from text can be either fine-grained (ADE entity recognition) or coarse-grained (ADE assertive sentence classification), with limited efforts leveraging inter-dependencies among these two granularities. Moreover, in most attention-based neural network models for sentence classification only a single round of attention focusing on simple semantic information is applied for learning the importance of words and the overall representation of the sentence. We design a multi-grained joint deep network model MGADE to concurrently solve both ADE tasks MGADE takes advantage of their symbiotic relationship, with a transfer of knowledge between the two levels of granularity. Our dual-attention mechanism constructs multiple distinct representations of a sentence that capture both task-specific and semantic in-formation in the sentence, providing stronger emphasis on the key elements essential for sentence classification. In several comprehensive experimental studies, namely, one for each part of this dissertation, we demonstrate the superiority of the proposed strategies over the state-of- the-art techniques with respect to precision, recall and F1-measure.

Creator