ratio-t2e 0.0.6

Linear left barrier loss and data augmention for survival anlysis

RATIO-T2E

This package contains 3 different elements:

1. suRvival Analysis lefT barrIer lOss (RATIO) + uniFormatIve fEatureS daTa Augmentation (FIESTA) + Ridge model
2. suRvival Analysis lefT barrIer lOss (RATIO) - loss function
3. uniFormatIve fEatureS daTa Augmentation (FIESTA) class

Prerequisites

Before you begin, ensure you have met the following requirements:

• Checked for Windows only
• python 3.8

Installing RATIO-T2E

pip install ratio-t2e

suRvival Analysis lefT barrIer lOss (RATIO) + uniFormatIve fEatureS daTa Augmentation (FIESTA) + Ridge model

This model solves a regression problem where the loss function is the Mean Square Error (MSE)
for the uncensored data and the RATIO loss for the wrong censored samples.
The user can moderate the relations between the censored and uncensored loss.
For extremely small datasets ( < 50),  Data augmentatuin (DA) should be added. The combination of RATIO and DA
is called FIESTA.

Using suRvival Analysis lefT barrIer lOss (RATIO) + uniFormatIve fEatureS daTa Augmentation (FIESTA) + Ridge model

To use the model follow these steps:

1. Divide your data to a censored dataframe and an uncensored dataframe, where a sample is condisered as censored when its time of a competing event preceeds the time to event (TTE), or when a sample did not have an event within the cohort's time. Make sure the dataframes consists of the following columns:

   • A TTE column for the uncensored dataframe and a competing times for the censored, that measures the times in days that have passes from the sample's date to the target date.

   • A column named "Date" of the date of the sample.

   • A column named "DateEnd" of the date of an event for uncensored, and date of competing event for censored.

   • A people column named similar to the people_col parameter, contains the identity of a patient.

   • A time column named similar to the time_col parameter, contains the order of samples for sequntial data.

   Example of the mendatory columns:

   

2. Load the uncensored and censored datafrmes.

censored = pd.read_csv("censored_data_file_name.csv", index_col=0)
uncensored = pd.read_csv("uncensored_data_file_name.csv", index_col=0)

3. Name the list of categorical features (name of columns).

list_of_categories = ['cat1', 'cat2', 'cat3', 'cat4']

4. Create the LBL class with the parameters you want.

 lbl = LBL("TTE_col", "people_col", "time_col", num_of_bact=881, feature_selection=20, categories=list_of_categories,
          with_microbiome=False, augmented_censored=False, gamma=0.0)

5. Divide the uncensored dataframe into a training set and a test set.

6. Merge the uncensored training dataframe with the censored dataframe, for training.

7. Fit the model on the training set.

8. Use the predict function for prediction.

9. Use the score function for evaluations (Spearman Correlation Coefficient (SCC), AUC and Concordance Index (CI).

suRvival Analysis lefT barrIer lOss (RATIO) - loss function

Since the RATIO loss is "model-free", there is an option to add RATIO loss to any model.

Using the RATIO loss

To use RATIO:

 import RATIO
 RATIO.RATIO(y,y_hat)

where y is a Tensor of shape (batch_size,2), its first dimension is a binary indicator of having the event (= 1) or lacking the event (= 0), and its second dimension is the TTE.

uniFormatIve fEatureS daTa Augmentation (FIESTA) class

uniFormatIve fEatureS daTa Augmentation (FIESTA) defines the augmented TTE of the censored samples by using high dimensional and not highly informative data. The DA process contains 2 steps:

1. Defining the augmented TTE as a weighted average of the uncensored samples based
on the difference in M (high dimensional data) between samples (as described in our paper).
There are 3 options for declaying functions Exponential (function_1), Hyperbolic (function_2) and Cauchy (function_3), which
is the default.
2. Computing the augmented TTE using Maximum Likelihood Estimation (MLE) on a model
where a constant censoring rate of lamda is assumed and the event is normally distributed around the previously computed in step 1.

Using FIESTA

To use FIESTA:

import MLE_augmentor
MLE_augmentor.FIESTA.implement_augment(censored_df,lamda)

such that censored_df has to contain the mendatory columns as was explained above, and lamda is the assumed censoring rate.

Contributors

Shtossel Oshrit

Contact

If you want to contact me you can reach me at [e-mail]

Citation