Predicting Patient Outcomes using Temporal Standardized Metrics and Natural Language Processing Models
An increasing amount of crucial information is stored in Electronic Health Records (EHRs). Therapeutic records often contain unstructured free text, making labeling and processing time-consuming, while the insights labeled data provides are immense. Natural language processing models are growing in clinical practice and are already used to identify disease onset and medication information from clinical notes. One specific NLP model, BERT, has been trained on other language datasets (giving a greater understanding of words and their relationships), allowing us to extract desired values from therapeutic note language. This study employs Natural Language Processing (NLP) to extract standardized clinical scales from therapeutic EHRs, aiming to build a tool for predicting and tracking patient outcomes from note histories. We exported, deidentified, and labeled a random subset of clinical notes from the Massachusetts General Hospital provider network using the AM-PAC 6 click score (for functional mobility), splitting into more specific scales by each component of the larger scale. We then used the BERT transformer model to automate the labeling of these notes. We generated one NLP model for each metric, which was determined to be optimal for generating the appropriate standardized metric from therapeutic notes. The final models had an 83% validation accuracy, indicating the automation of standardization of notes is feasible. Implementing this model will provide numerical data for each therapeutic note, allowing us to work towards the final goal of predicting therapeutic patient outcomes.
