A Knowledge Distillation Ensemble Framework for Predicting Short and Long-term Hospitalisation Outcomes from Electronic Health Records Data

The ability to perform accurate prognosis of patients is crucial for\nproactive clinical decision making, informed resource management and\npersonalised care. Existing outcome prediction models suffer from a low recall\nof infrequent positive outcomes. We present a highly-scalable and robust\nmachine learning framework to automatically predict adversity represented by\nmortality and ICU admission from time-series vital signs and laboratory results\nobtained within the first 24 hours of hospital admission. The stacked platform\ncomprises two components: a) an unsupervised LSTM Autoencoder that learns an\noptimal representation of the time-series, using it to differentiate the less\nfrequent patterns which conclude with an adverse event from the majority\npatterns that do not, and b) a gradient boosting model, which relies on the\nconstructed representation to refine prediction, incorporating static features\nof demographics, admission details and clinical summaries. The model is used to\nassess a patient's risk of adversity over time and provides visual\njustifications of its prediction based on the patient's static features and\ndynamic signals. Results of three case studies for predicting mortality and ICU\nadmission show that the model outperforms all existing outcome prediction\nmodels, achieving PR-AUC of 0.891 (95$%$ CI: 0.878 - 0.969) in predicting\nmortality in ICU and general ward settings and 0.908 (95$%$ CI: 0.870-0.935) in\npredicting ICU admission.\n