FrESCO: Framework for Exploring Scalable Computational Oncology

Motivation

The National Cancer Institute (NCI) monitors population level cancer trends as part of its Surveillance, Epidemiology, and End Results (SEER) program. This program consists of state or regional level cancer registries which collect, analyze, and annotate cancer pathology reports. From these annotated pathology reports, each individual registry aggregates cancer phenotype information and summary statistics about cancer prevalence to facilitate population level monitoring of cancer incidence. Extracting cancer phenotype from these reports is a labor intensive task, requiring specialized knowledge about the reports and cancer. Automating this information extraction process from cancer pathology reports has the potential to improve not only the quality of the data by extracting information in a consistent manner across registries, but to improve the quality of patient outcomes by reducing the time to assimilate new data and enabling time-sensitive applications such as precision medicine. Here we present FrESCO: Framework for Exploring Scalable Computational Oncology, a modular deep-learning natural language processing (NLP) library for extracting pathology information from clinical text documents.

Documentation is available over at Read the Docs, and a software archive at Zenodo .

Quickstart Guide

Install from source

Clone the repo shell git clone https://github.com/DOE-NCI-MOSSAIC/FrESCO.git Load the working branch and pull in the subrepos shell cd FrESCO git checkout main Setup the conda environment (the default name for the environment is "ms39", this can be edited in the ms39.yml file) shell conda env create -n fresco python=3.11 conda activate fresco Install the FrESCO library and dependencies shell pip install .

For further PyTorch instructions or more details for pytorch, head over to the PyTorch docs.

Notebooks and Examples

We have supplied example notebooks for each of the sample datasets contained in this repository showing model training on each dataset. We have also supplied model_args files for each of the datasets contained within the repo to speedup the time to ge up and running with the codebase.

Data Preparation

We have supplied three different datasets as examples, each must be unzipped before any model training via the tar -xf dataset.tar.gz command from the data directory. The three datasets are: - imdb: binary sentiment classification with the imdb dataset, - P3B3: benchmark multi-task classification task, and - clc: case-level context multi-task classfication data.

To create the data files for the imdb dataset, you may run the bardi_data.py script within the scripts folder to create the necessary files for model training.

Model Training

The model_args.yaml file controls the settings for model training. Edit this file as desired based on your requirements and desired outcome. The "save_name" entry controls the name used for model checkpoints and prediction outputs; if left empty, a datetime stamp will be used.

The following commands allow setting your GPUs, if enabled, before training your model. shell nvidia-smi #check GPU utilization export CUDA_VISIBLE_DEVICES=0,1 #replace 0,1 with the GPUs you want to use echo $CUDA_VISIBLE_DEVICES #check which GPUs you have chosen To train the model for any information extraction task, multi-task calssification, simply run shell python train_model.py -m ie -args ../configs/model_args_bardi.yml from the scripts directory which will train a model on the imdb dataset.

We have supplied test data for each of the model types provided. Information extraction models may be created with either P3B3 or imdb data. To run with your own data, the following instructions explain the requirements for the training data.

Custom Datasets

Add the path to the desired dataset in the the data_path argument in the configs/model_args.yml file. The required data files are: - data_fold0.csv: Pandas dataframe with columns: - X: list of input values, of int type - task_n: output for task n, a string type (these are the y-values) - split: one of train, test, or val - id2labels_fold0.json: index to label dictionary mapping for each of the string representations of the outputs to an integer value, dict keys must match the y-values label - word_embeds_fold0.npy: word embedding matrix for the vocabulary, dimensions are words x embedding_dim. If no word embedding exists, the model will use randomly generated embeddings.

You will also need to set the tasks, these must correspond to the task columns names in the data_fold0.csv file and keys in the id2labels_fold0.json dictionary. For example, in the P3B3 data, the task columns are task_n, n = 1,2,3,4. Whereas the imdb data has the sentiment task. If using any sort of class weighting scheme, the keyword class_weights must be either a pickle file or dictionary with keys corresponding to the task and value with a corresponding list, or numpy array, of weights for that task.
If the class_weights keyword is blank, corresponding to None, no class weighting scheme will be used during training nor inference.

If working with hierarchical data, and the case-level context model is the desired output, then the dataframe in the corresponding data_fold0.csv must contain an additional integer-valued column group where the values describe the hierarchy present within the data. For example, all rows where group = 17 are associated for the purpose of training a case-level context model.

Case-Level Context Model

If you're wanting a case-level context model, there is a two-step process. See notebooks/clc_example.ipynb for a fully worked example.

Step 1: Create an information extraction model specifying the data/clc data directory in the file configs/clc_step1.yml. Then run shell python train_model.py -m ie -args ../configs/clc_step1.yml from the scripts/ directory.

Step 2: To train a clc model, set the model_path keyword arg to the path of the trained model trained from step 1 step in the configs/clc_step2.yml file. Then run shell python train_model.py -m clc -args ../configs/clc_step2.yml from the scripts/ directory to train a case-level context model.

Note that the case level context model requires a pre-trained information extraction model to be specified in the configs/clc_args.yml file. The default setting, if the -m argument is omitted, is information extraction, and which task is specified in the configs/model_args.yml file.

Deep-Abstaining Classifier and Ntask

Both information extraction and case-level context models have the ability to incorporate abstention or Ntask. The deep abstaining classifier (DAC) from the CANDLE repository, allows the model to not make a prediction on a sample if the softmax score does not meet a predetermined threshold specified in the minacc keyword of the `modelargs.yml` file. It can be tuned to meet a threshold of accuracy, minimum number of samples abstained, or both through adapting the values of alpha through the training process. This adapation is automated in the code, only requires the user to specify the initial values, tuning mode, and scaling.

Ntask is useful for multi-task learning. It creates and additional 'task' that predicts if the softmax scores from all of the tasks do not meet a specified threshold within the model_args file . It has its own parameters that are tuned during the training process to obtain a minimum of abstained samples and maximum accuracy on the predicted samples. Ntask may not be enabled without abstention being enabled as well. The code will throw an exception and halt if such configuration is passed.

Expected Results

Training a model with the supplied default args, we see convergence within 50-60 epochs with 0.80-0.85 accuracy on the imdb data set and in excess of 0.90 accuracy across all tasks for the P3B3 dataset within 60 epochs or so. NOTE: P3B3 has a known issue training with mixed precision and with DAC and NTask enabled. Ensure these keywords are all False for all runs with the P3B3 dataset.

Bring Your Own (Torch) Model

To try out your favorite torch model with our training and evaluation methods, your torch model must subclass the torch.nn.Module and provide a forward call within the model class. The forward method is expected to take a pytorch torch.utils.data.DataLoader class batched input, a map-style dataset, see the docs describing this class. The forward method is expected to return 'logits', the unnormalized outputs fro mthe network that have not been passed through a function mapping the logits the interval [0, 1].

The machinery provied here is designed to work with neural network models, and is overkill for a traditional sklearn type model, as most of the necessary functions to train and evaluate a model are provided within sklearn. If you wish to try our sample pre-processing steps with your favorite sklearn model, an example is provided within the notebooks directory. Using the generated word embeddings and tokenized documents generated by our preprocessing script are incompatibe with the sklearn library, which expects feature vectors as input, not higher dimensional, ie matrix nor tensor, inputs.

Contributing

Get in touch if you would like to help in writing code, example notebooks, and documentation are essential aspects of the project. To contribute please fork the project, make your proposed changes and submit a pull request. We will do our best to sort out any issues and get your contributions merged into the main branch.

If you found a bug, have questions, or are just having trouble with the library, please open an issue in our issue tracker and we'll try to help resolve it.

How to Cite

If you use our software in your work please cite this repository. The bibtex entry is: @misc{osti_1958817, title = {FrESCO}, author = {Spannaus, Adam and Gounley, John and Hanson, Heidi and Chandra Shekar, Mayanka and Schaefferkoetter, Noah and Mohd-Yusof, Jamaludin and Fox, Zach and USDOE}, abstractNote = {The National Cancer Institute (NCI) monitors population level cancer trends as part of its Surveillance, Epidemiology, and End Results (SEER) program. This program consists of state or regional level cancer registries which collect, analyze, and annotate cancer pathology reports. From these annotated pathology reports, each individual registry aggregates cancer phenotype information and summary statistics about cancer prevalence to facilitate population level monitoring of cancer incidence. Extracting cancer phenotype from these reports is a labor intensive task, requiring specialized knowledge about the reports and cancer. Automating this information extraction process from cancer pathology reports has the potential to improve not only the quality of the data by extracting information in a consistent manner across registries, but to improve the quality of patient outcomes by reducing the time to assimilate new data and enabling time-sensitive applications such as precision medicine. Here we present FrESCO: Framework for Exploring Scalable Computational Oncology, a modular deep-learning natural language processing (NLP) library for extracting pathology information from clinical text documents.}, url = {https://www.osti.gov//servlets/purl/1958817}, doi = {10.11578/dc.20230227.2}, url = {https://www.osti.gov/biblio/1958817}, year = {2023}, month = {3}, note = }