Joeky Senders

175 General discussion semantic relationships). By scrutinizing the full complexity of the data, however, deep learning algorithms were computationally inefficient and perhaps prone to overfitting to statistical noise. In Chapter 8 , we compared the learning curves of various algorithms in determining the histopathological diagnosis of brain tumor patients based on free-text pathology reports. In this study, we developed a modified version of the generic convolution network model equipped with stronger methods of regularization. The resultant model was able to model the semantic complexity of text documents without overfitting to statistical noise. The number of required training samples to reach the predetermined performance thresholds (an AUC of 0.95 and 0.98) was two to eight times lower for the modified deep learning model, ClinicalTextMiner, compared to regression and conventional deep learning-based architectures. The steep learning curve can be valuable for natural language processing tasks with a limited set of training examples available (e.g., rare diseases and events or institutions with lower patient volumes). Utilizing natural language processing in healthcare could have profound implications for clinical research and even patient care. Currently, clinical research endeavors are restricted significantly by the need for financial and human resources to gather, process, and analyze clinically generated data. Observational studies are therefore limited to data sets that can be collected by hand, often a mere fraction of the entire population. Yet, their results are generalized to the entire population. The automatic nature could accelerate retrospective chart review to an unprecedented scale, such as the entire population, and allow for the assembly of large, continuously updated clinical registries. The deterministic nature can make data collection less subject to inter- and intra-reviewer inconsistencies but rather based on a consensus label from clinical experts. The impact of natural language processing in clinical care could even bemore profound. Although the bulk of biomedical information is increasing in volume and complexity, the human physician brain that has to comprehend this information is and will remain the same. Information overload, therefore, constitutes a significant problem in the digital age of healthcare and plays a key role in diagnostic errors, near misses and patients’ safety, as well as the stress and work satisfaction perceived among healthcare workers. 5,6 Natural language processing algorithms might assist exposing relevant information in a patient’s chart without multiple clicks or relieve the administrative burden on clinicians. The process of viewing and entering the clinically most useful data frictionless is essential for clinicians, not just for their convenience, but to spend more time with their patients and provide the best possible care. 7