Deep neural networks to forecast cardiac and respiratory deterioration of intensive care patients

Deep neural networks have proven valuable in several applications. The availability of electronic health records at high frequency has made it possible to provide realtime prediction to stay relevant to the user’s immediate and changing context. This thesis implements deep neural networks for the prediction of short term cardiac and respiratory deterioration. It is based on the cardiac and respiratory SOFA sub-scores to define the event of deterioration, and it uses convolutional neural networks, long short-term memory and multitask learning to construct models that alert if the patient is prone to deterioration. Data from the FINNAKI study was used in training the predictive models, and a subset of the MIMIC III clinical database was used to investigate the applicability of those models in intensive care units from different locations. In terms of area under the ROC curve, the predictive models could achieve an area under score of 0.7812 from the FINNAKI data and 0.6816 for a subset of MIMIC III. Those results confirm that short-term deterioration is predictable which could help caregivers in focusing more on the patients at risk of deterioration in the short term