Joeky Senders

174 Chapter 10 Part III: Natural language processing in neurosurgical oncology Part III encompasses the application of machine learning to a higher dimensional problem. Various natural language processing approaches were developed to automate the processing and analysis of narratively written clinical reports. In Chapter 6 , we have developed a pipeline for automated clinical chart review by analyzing a corpus of free-text radiology reports of brain tumor patients. In this study, we utilized a bag- of-words approach with a classical statistical algorithm known for its strong method of regularization, LASSO regression. The developed pipeline was able to extract 15 distinct radiographic features with high to excellent discriminatory performance (AUC 0.82-0.98). Model performance was correlated with the interrater agreement, which underlines the importance of expert consensus in generating ground truth training labels. However, expert consensus can also be used as a potential indicator for the complexity of the natural language processing task at hand. In Chapter 7 , we compared various statistical (logistic regression, LASSO regression), classical machine learning (fully connected artificial neural networks), and deep learning (convolutional neural networks, gated recurrent unit, and long short-term memory) techniques in their ability to classify radiology reports of brain metastases patients into reports that describe solitary versus multiple metastases. Both the LASSO regression and convolutional neural networks model demonstrated to outperform other competing statistical and machine learning models. Although these algorithms are on the opposite ends of the machine learning spectrum, their performance were highly comparable. The LASSO regression model focused merely on the relative frequency of words or word combinations but ignored the order or semantic properties of individual words. In contrast, the deep learning model (i.e., convolutional neural networks) were able to accommodate to higher-level lexical complexity. This sequence- based approach also modeled the order of the words and paragraphs, as well as the semantic relationships among words and thus the statistical properties of a language. Despite the advantages of modeling these sequential and semantic attributes, the deep learning model in this project did not outperform the less complex LASSO regression model. Perhaps due to its simplicity, LASSO regression demonstrated the most robust performance across different metrics. This implies that the underlying signal for this particular text classification task was found in primitive (i.e., relative word frequencies) rather than complex patterns within the data (i.e., sequential and