Joeky Senders

152 Chapter 8 Discussion In this study, we compared the learning curves of various natural language processing techniques for clinical text mining. We identified a deep learning architecture that learns two to eight timesmore efficiently than competing deep learning and regressions-based approaches. The underlying source code has been released on a publicly accessible repository, thereby allowing for external application, validation, and optimization. Fewother groupshaveexplored theuseof deep learning for natural languageprocessing of free-text clinical reports. 3–14 Although these studies demonstrate the strong potential of deep learning for clinical text mining, this has only been demonstrated on a static text corpus of clinical reports with a fixed number of training examples. The current body of literature therefore still lacks evidence on the learning capacity of natural language processing as it applies to clinical text mining. To the best of our knowledge, this is the first study that compares the learning curve of deep learning and other competing algorithms in their ability to process free-text clinical reports. Additionally, by removing the convolutional layer and combining the embedding layer with strong methods of regularization, we identified a model architecture that learns more efficiently compared to competing algorithms and even CNNs, thereby requiring less training examples to develop high performing clinical text mining models. Limitations Several limitations of the current study should be mentioned, which underline common barriers in natural language processing and machine learning modeling. The current models are trained on single institutional data and might not generalize well to data from external institutions, which may have different styles and routines in the language used in clinical reports. Instead of the resultant models, the underlying code pipeline was therefore made publicly accessible in order to promote the reproducibility and external generalizability of the current work. Labels were necessary for training and testing of the natural language processing models, and these were derived through manual chart review; however, manual chart review remains prone to error as well. In the current study, a trained clinical reviewer (R.W.) with over 20 years of experience has provided the required labels. The current study utilized a three-way classification problem (i.e., glioma, metastasis, or meningioma) to evaluate the learning curves. However, the number of diagnostic classes can vary widely dependent on the scope of interest, which could also impact the efficiency of the resultant learning curves.