Longitudinal Temporal NLP Models of Kidney Disease Progression Retrieved using Clinical Notes

Chronic Kidney Disease (CKD) is a progressive disease which needs early monitoring and regular monitoring to avoid complications and cure the patient. Conventional methods of CKD modeling might be premised on the most systematic data and overlook the plentiful time and text signals that can be found in patient narratives. In this paper, we introduce a time-dependent Natural Language Processing (NLP) framework, which is designed to capture progression of CKD with the use of unstructured clinical notes. Using temporal encoding strategies (Time2Vec, Bi-LSTM and time-aware Transformers) and incorporating domain-specific pretrained language models (Clinical BERT), we translate raw text into temporally-aligned representations of disease state at multiple points in time. The training and testing of the models occurred on an annotated set of MIMIC-III dataset that included the explicit and implicit CKD stages (temporal expressions). The models, which scored highest among those evaluated, is the Time-aware Transformer that had an F1-score of 0.915 and an AUC of 0.94, significantly better than baseline and interpretable models. The presented strategy shows how temporal reasoning plays a pivotal role in modelling disease dynamics and makes an effective and explainable framework available to monitor CKD using freetext clinical data. The present work establishes a framework that would be used in future multi-modal/real-time predictive systems of chronic disease management.