Neural Architecture Search for Deep Reinforcement Learning-based Blood Glucose Prediction of Type-1 Diabetes

Deep Reinforcement Learning-based framework for single- or multi-step ahead predictions of time series data, e.g. (continuous) blood glucose values of type 1 diabetics. The framework uses Optuna for (additional) Neural Architecture Search and Hyperparameter Optimization.
Here is an overview of the proposed methodology:

Ohio T1DM dataset

Learn more about the Ohio T1DM dataset for blood glucose prediction here.

Dependencies

- tensorflow 2.8.0
- tensorboard 2.8.0
- tf-gents 0.12.0
- gym 0.23.0
- pandas 1.1.5
- numpy 1.21.5
- matplotlib 3.5.1
- absl-py 1.0.0
- gin-config 0.5.0

Install dependencies using the following command: pip install -r requirements.txt

Configuration options of the framework (see conig.gin)

General settings

- path_to_train_data [str]: specifiy path to csv training data set
- path_to_eval_data [str]: specify path to csv testing data set 
    Note: if path is empty, training data is used for evaluation
- normalization [str]: specify normalization method  (options: min_max, z_score, none)
- setup [str]: specify setup (currently single_step and multi_step)
- rl_algorithm [str]: specify RL algorithm to use   (options: ddpg, td3, reinforce, ppo, sac, dqn)
    Note: dqn only for single step; tf_agents only supoort scalar actions 
    Note: On-policy: reinforce, ppo ; Off-policy: ddpg, td3, sac, dqn
- env_implementation [str]: specify environment implementation to use   (options: tf, gym)
- use_nas [bool]; Specify to use Neural Architecture Search
- use_gpu [bool]: Specify to use GPU(s) 
- multi_task [bool]: If True multi-task setup (training on multiple patients simultaneously)

DRL training settings

- max_train_steps [int]: specify max number of training steps
- eval_interval [int]: specify how often to evaluate
- pretraining_phase [bool]: (off-policy algorithms) specify to use pretraining phase or not
- restore_dir [str]: path to directory with weights and biases to restore (no restoring if empty str) 
- layers_to_train [str]: if multi-task with resotring this str specifies which layers to train
   (options: last, dec_last, lstm_dec_last)

Single step settings

Gym environment settings 
------------------------------
- window_size [int]: input window size
- min_attribute_val [float]: minimum value of attribute, e.g., CGM
- max_attribute_val [float]: maximum value of attribute, e.g., CGM
- reward_def [str]: specify reward function to use    (options: abs_diff, linear, exponential)
- max_window_count [int]: specify the maximum number of windows to use per training iteration
    Note: specify -1 if you want to use as much windows as possible with random starting point


 TF environment settings
------------------------------
- window_size [int]: input window size
- min_attribute_val [float]: minimum value of attribute, e.g., CGM
- max_attribute_val [float]: maximum value of attribute, e.g., CGM
- max_window_count [int]: specify the maximum number of windows to use per training iteration
    Note: specify -1 if you want to use as much windows as possible with random starting point
- batch_size [int]: specify batch size
- state_type [str]: specify state type (options: skipping, no_skipping, single_step_shift)

Multi step settings

Gym environment settings

- window_size [int]: input window size
- forecasting_steps [int]: number of steps to forecast
- min_attribute_val [float]: minimum value of attribute, e.g., CGM
- max_attribute_val [float]: maximum value of attribute, e.g., CGM
- reward_def [str]: Specify reward function to use    (options: abs_diff, linear, exponential)
- max_window_count [int]: specify the maximum number of windows to use per training iteration
    Note: specify -1 if you want to use as much windows as possible with random starting point

TF environment settings

- window_size [int]: input window size
- pred_horizon [int]: number of steps to forecast
- min_attribute_val [float]: minimum value of attribute, e.g., CGM
- max_attribute_val [float]: maximum value of attribute, e.g., CGM    
- max_window_count [int]: specify the maximum number of windows to use per training iteration
    Note: specify -1 if you want to use as much windows as possible with random starting point
- batch_size [int]: specify batch size
- state_type [str]: specify state type (options: skipping, no_skipping, single_step_shift)

Run in terminal

Set configuration options in config.gin and run the following command in the terminal: python main.py

Run in Google colab

• main.py: change path to config.gin (absolute path, e.g., '/content/rl_time_series_forecasting/config.gin')
• config.gin: change path to datasets (absolute path, e.g., '/content/rl_time_series_forecasting/data/540-ws-training.csv')
• Jupyter notebook code
  • !pip install -r /content/BG-Prediction-for-Type-1-Diabetics-using-DRL/requirements.txt
  • !python /content/BG-Prediction-for-Type-1-Diabetics-using-DRL/main.py

Visualization in Tensorboard

Navigate to logging directory of interest and type (in the terminal)

tensorboard --logdir .

Or specify the absolute path to the directory, e.g.,

tensorboard --logdir /home/my_project/logs/my_log_dir

Stand-alone scripts (in evaluation directory)

1. metric_eval_csv.py

   • Script to calculate the standard metrics (MAE, MSE, RMSE, ...) for arbitrary data sequences
   • Stand-alone script that can be called directly from the terminal, e.g. python metric_eval_csv.py
   • It has the following configuration parameters (arguments of the python call)
     • csv_path [str]: path to csv data set
     • metrics [str]: metrics to evaluate, e.g., 'mse, rmse'
     • indices [str]: 'min_index, max_index' of samples to evaluate
     • strategy [str]: 'consecutive' or 'random' samples
     • setup [str]: 'single_step' or 'multi_step' scenario
   • Thus, a call of the script can look the following (including all arguments)
     python metric_eval_csv.py --csv_path="./data.csv" --metrics="mse,rmse" --indices="0,100" --strategy="consecutive" --setup="multi_step"

2. uq_visualization.py

   • Script to visualize UQ (variance) samples of training iterations
   • Stand-alone script that can be called directly from the terminal, e.g. python uq_visualization.py
   • It has the following configuration parameters (arguments of the python call)
     • csv_path [str]: path to csv data set
     • setup [str]: 'single_step' or 'multi_step' scenario
     • vis_type [str]: different types of plots ('vis_eval_samples', 'vis_avg_training', 'vis_avg_forecasting')
     • vis_steps [int]: number of steps to visualize (multiple of eval freq, e.g. vis_steps * 500)
     • vis_std [bool]: if True visualize std otherwise visualize variance
     • vis_forecasting_error [bool]: if True additionally visualize forecasting error
     • error_metric [str]: metric to visualize if vis_forecasting_error is True
     • y_lim [int]: limit of y-axis for vis_avg_forecasting
     • dataset_path [str]: path to (training) dataset to visualize windows with min. / max. error
     • save_fig [str]: "True" or "False"
     • save_path [str]: path to store figure
   • Thus, a call of the script can look the following (including all arguments)
     python uq_visualization.py --csv_path="./data.csv" --setup="multi_step" --save_fig="True", --save_path="."

3. seg_visualization.py

   • Script to calculate and visualize the Surveillance Error Grid (SEG)
   • Stand-alone script that can be called directly from the terminal, e.g. python seg_visualization.py
   • It has the following configuration parameters (arguments of the python call)
     • csv_path [str]: path to csv data set
     • img_path [str]: path to seg image (superimposed with prediction results)
     • ph [int]: prediction horizon, e.g., 30 min
     • patient [int]: patient id
     • save_path [str]: path to store figure
   • Thus, a call of the script can look the following (including all arguments)
     python uq_visualization.py --csv_path="./data.csv" --img_path="./seg.png" --ph=30, --patienth=570 --save_path="."
   • Example:

Citation

If you like our work, please cite the following paper

@article{domanski2024advancing,
  title={Advancing blood glucose prediction with neural architecture search and deep reinforcement learning for type 1 diabetics},
  author={Domanski, Peter and Ray, Aritra and Lafata, Kyle and Firouzi, Farshad and Chakrabarty, Krishnendu and Pfl{\"u}ger, Dirk},
  journal={Biocybernetics and Biomedical Engineering},
  volume={44},
  number={3},
  pages={481--500},
  year={2024},
  publisher={Elsevier}
}