This is the source code for building a deep learning model using multi-task learning (MTL) approach. The model is designed to train and validate on data prepared from Electronic Health Records (EHRs).
How long a patient will stay in the hospital or patient's length of stay (LOS) is a necessary piece of information for hospital resource allocation. Having a good estimation of LOS can aid in hospital planning and decision making. Prediction of LOS at the admission time might be distorted because right at the beginning, we don't have adequate information to interpret patient's status. To address this, we have designed our model to learn and predict adaptively along the patient's stay. The model appends new data about the patient as it becomes available and update the prediction value accordingly.
Also, general hospitals where there are arrangements of multiple units for providing care, patients can get transferred from one unit to another due to a diverse set of underlying situations. Hospital logistics as well as the span of patient's stay might change depending on the unit he/she is moving. Considering the latent relationship between this two variable, we designed our model to predict these two variable simultaneously.
In this multi-task learning model, we train the model to predict:
- Task-1 --> patient's unit label in next time step
- Task-2--> patient's remaining length of stay
The model inputs get updated on a daily basis and generates prediction accordingly.
The dataset contains six years of health records starting from January 2016 and December 2021. We split the dataset for pre-COVID and post-COVID. We further split the dataset for adult and pediatric patients as they have separate care units inside the hospital.
The data folder holds small synthetic datasets for the purpose of running the program. We have added a few dummy patients and their corresponding bedflow data. The features of the toy data are similar to the original data. We have added dummy diagnostics and unit transfer sequence for the patients.
The code is specifically desinged to take input from the user based on the EHR data that we are using. Executing the code will ask the user whether she wants to train a model or wants to test the model that was previously trained. The user can also do both by selecting a dataset(train and test). Depending on the user input, the code will train and/or test model, save the model with its weight vectors and subsequently plot the evaluation metrics.
The starting point of the code is main.py. It takes the user input and then goes to run_experiment. The DataManager class prepares inputs depending on the user specified dataset. Then MultiTaskModel class handles the model building/training/testing. The default configurations are stored in config.py file.
Since there are two different types of tasks, our evaluation metrics are set accordingly. Prediction of next unit label is a classification task, our evaluation metrics for this task are accuracy and top-k accuracy (k=2). Prediction of remaining length of stay is a regression task, our evaluation metrics in this case is mean absolute error, mean squared error and root mean squared error. The MAE and Top-k accuracy values are plotted and saved in the image folder to show the model performance on test data. The history of the training time is saved in the hist_log folder.
When referring to this paper, please cite:
Kaniz Madhobi, Eric Lofgren, Ananth Kalyanaraman. A multi-task learning approach for predicting spatio-temporal patient variables. Proc. 15th ACM Conference on Bioinformatics, Computational Biology, and Health Informatics (ACM-BCB '24), Accepted/In Press, 2024.