Multitask Learning and Benchmarking with Clinical Time Series Data

• Aram Galstyan  |  David C. Kale  |  Greg Ver Steeg  |  Hrant Khachatrian  |  Hrayr Harutyunyan

• University of Southern California

It is difficult to measure progress in machine learning for healthcare research as there are no publicly available benchmark data sets. To address this problem, four clinical prediction benchmarks are proposed using data derived from publicly available MIMIC-III database. This covers problems including modelling risk of mortality, forecasting length of stay, detecting physiological decline, and phenotype classification. The algorithms used for the four tasks are regression, LSTM and channel-wise LSTM. A subset of the MIMIC-III database is compiled containing events that cover 42,276 ICU stays of 33,798 unique patients. The four benchmark tasks are defined on this subset. For the benchmark tasks linear regression models and multiple neural architectures are developed. Experiments are performed with a basic LSTM-based neural network and also a modification of it, channel-wise LSTM is introduced. Hyperparameter search is performed to select the best model and those are evaluated on test sets of the corresponding tasks.

Input variables : Capillary refill rate, Diastolic blood pressure, Fraction inspired oxygen, Glascow coma scale eye opening, Glascow coma scale motor response, Glascow coma scale total, Glascow coma scale verbal response, Glucose, Heart Rate, Height, Mean blood pressure, Oxygen saturation, Respiratory rate, Systolic blood pressure, Temperature, Weight, pH
Output Variables : In hospital mortality, Length of remaining stay, Classification of Phenotype

Visit Model : github.com

Additional links : arxiv.org