A patient specific forecasting model for human albumin based on deep neural networks.

BACKGROUND AND OBJECTIVES: Hypoalbuminemia can be life threatening among critically ill patients. In this study, we develop a patient-specific monitoring and forecasting model based on deep neural networks to predict concentrations of albumin and a set of selected biochemical markers for critically ill patients in real-time.

METHODS: Under the assumption that metabolism of a patient follows a patient-specific dynamical process that can be determined from sufficient prior data taken from the patient, we apply a machine learning method to develop the patient-specific model for a critically ill, poly-trauma patient. Six representative biochemical markers (albumin (ALB), creatinine (Cr), osmotic pressure (OSM), alanine aminotransferase (ALT), total bilirubin (TB), direct bilirubin (DB)) were collected from the patient while scheduled exogenous albumin injection was administered to the patient for the total of 27 consecutive days. A sliding window of data in 11 consecutive days were used to train and test the neural networks in the model.

RESULTS: The obtained dynamical system model represented by neural networks is used to forecast the biochemical markers of the patient in the next 24 h. The relative error between the predictions and the clinical data remains consistently lower than 2%.

CONCLUSIONS: This study demonstrates that a patient-specific dynamical system model can be established to monitor and forecast dynamical behavior of concentrations of patients' biochemical markers (including albumin) using deep learning methods on neural networks.