lyprox.riskpredictor

package documentation

Risk predictions for involvement patterns, given personalized diagnoses.

The predictions displayed in the risk predictor are not actually computed by the web app. Instead, LyProX reconstructs one of the lymph.models with the help of config files - parsed by the lyscripts package - that can be found in any repos that define a DVC pipeline.

This is probably the most complex interplay between the different libraries, so we will try to break it down as complrehensively as possible:

The lymph Models

The code implementing the mathematical models for lymphatic tumor progression are provided by the lymph library. They implement the different lymph.models.
DVC Pipelines in Paper Repos

Our research group has, in the past, published a number of papers where we inferred the optimal parameters of various such models, given a cohort of patients. These sets of inferred parameter samples are referenced by a tool called DVC in different repositories with instructions on how to fetch them from a remote storage container.
Reproducing Models with lyscripts

However, these references are worthless without unambiguous instructions on how to reconstruct the exact model that was used for inference. For this, we use the lyscripts package. Its module lyscripts.configs provides a bunch of pydantic.BaseModel subclasses for defining basically every important aspect of the modelling process.

What happens in this "little" app now, is that one defines where it should look for the YAML files that define a model and a sampling pipeline. This definition is stored as a CheckpointModel in the SQLite database. From instances of this class, one can then validate the config files, fetch the samples from the remote storage using DVC, and precompute prior state distributions from the trained model. See the docs of the CheckpointModel for a list of all the available methods and how they cache their intermediate results using joblib.

Consequently, a user can then head to the risk predictor's dashboard, select a CheckpointModel and have the web app compute specific risks defined via the RiskpredictorForm from the precomputed quantities. This actual risk predictions happens in the predict module of the app.

Module	`admin`	Boilerplate code to register the CheckpointModel with the Django admin interface.
Module	`apps`	Boilerplate Django configuration for the riskpredictor app.
Module	`forms`	Forms used in the `riskpredictor` app.
Package	`management`	Undocumented
Package	`migrations`	Undocumented
Module	`models`	Models for the `riskpredictor` Django app.
Module	`predict`	Module for functions to predict the risk of lymphatic progression.
Module	`tests`	Undocumented
Module	`urls`	URLs related to the `riskpredictor` prediction app.
Module	`views`	Module for the views of the riskpredictor app.