Unifying Cardiovascular Modelling with Deep Reinforcement Learning for Uncertainty Aware Control of Sepsis Treatment.
This repository conatins code for Reinforcement Learning based Dynamic Treatment learning, respecting partial observability for Sepsis Treatment. Associated with the paper Unifying Cardiovascular Modelling with Deep Reinforcement Learning for Uncertainty Aware Control of Sepsis Treatment.
We use a novel Cardiovascular Physiology driven autoencoder, with learns patient specific Cardiovasular states, this structure is expected to convey important Physiological information which can't be directly observed. We also use another denoising Autoencoder to represent the history of the labs, and then use Offline Reinforcement Learning to learn optimal treatment strategies. Further we define model uncertainty and we instruct our RL agent to consider it's own uncertainty. The Figure below shows the overall decision structure. Given a patient state, we ask the agent to produce value distributions associated with each (discretized) action, how uncertain it is on it's own results and what a clinican may do (based on behavior cloning). Then all these information is then combined into making the decision.
.png "Overall")
Open mimic database was used as the data source. Relationships of the schema can be found here. You need permission to access MIMIC Data, more information on this is given on their website. After getting access I would recommend using Big Query MIMIC project to quickly access already processed schema.
Most of the preprocessing was done on Google Big Query Mimic-iii project,and preprocessed and cleansed using SQL on Big Query and Pandas. Pivoted Labs, Sofa Score related scores and Vitals are all available on Big Query in Derived Tables.
Vasopressor and Fluid based csv files were derived from Komorowski et. al, for the code included here, you would need hourly bins for the Metavision actions. Carevu only has the specific time.
To Run the analysis we need, Pivoted Vitals,Sofa Scores and hourly vasopressor and fluid treatments. The RL cohort is included in terms on icustay ids for convenience.
We propose an inference structure which could infer patient and intervention specific cardiovascular states, by a Sequential AutoEncoder, which is implicitly regularized by constraining the latent representation to have phyisilogical meaning and the decoder to be a mathematical model. This has the benefit from a simplicity, but has the potential to give useful Cardiovascular information in the ICU settings. The folder Physiology Aware Sequential Autoencoder, contains the training and modules.
Since the labs are much more sparse, we use a denoising stacked sequential autoencoder structure with low dimensional final hidden layer to encode the history of the labs, This is supposed to give a much better representation which then could be used as a state component in the (PO)MDP. The code can be found under Lab Models
For the RL implemntations, for convenience, I had saved states (which includes derived states), actions and rewards in (a rather large) csv file. This makes the replay buffer/batching straightforward application of PyTorch Dataset and Dataloader.
Initially tried Continous Batch Constrained Deep Q Learning (BCQ) and Discrete BCQ (with some minor modifications) as the DRL alogrithms. BCQ is an offline batch RL algorithm with as the name suggests regulaizes state actions pairs based on what is included in the batch.
Final work uses Distributinal RL (C51) algorithm, further we use uncertainty Quantification to quatify epistemic uncertainty.
To reproduce our results, you would first need to get access to MIMIC-iii data, and by using SQL queries in SQL folder (or otherwise), derive equivalent vasopressor and fluid treatments. Careview and Metavision tables have different formats of input events. Once this is done you could use the PyTorch dataset, dataloader classes to train the representation learning modules.
Finally to run RL, you need to way to generate (s,a,s',reward,done) tuples, where s includes the inferred hidden states. For this work, we save the states in a csv file, and then used a Pytorch Dataloader to generate batches for RL. A basic/naive way of this process is shown in the RL/RL_csv.ipynb notebook.
Refer RL folder for RL modules. distRL has all the RL modules.
SQL queries regarding vasopressors and fluids were derived from Komorowski et. al. Sepsis score realted queries were derived from Sepsis 3