Skip to content

Wikimedial

Overview

This is an infrastructure developed by Medial EarlySign to streamline the creation of predictive models using EMR data for clinical applications. Existing tools often fall short for clinical use many Python libraries are not optimized for sparse time series analysis, leading to high memory consumption and, in some cases, performance that is 10–100 times slower than necessary.

Medial Infrastructure is designed to turn the Electronic Medical Record (EMR)-a complex, semi-structured time-series dataset, into a machine-learning-ready resource. Unlike images or free text, EMR data can be stored in countless formats, and its "labels" (the outcomes or targets you want to predict) aren’t always obvious. We address this by standardizing both the storage and the processing of time-series signals. We can think about this infrastructure as "TensorFlow" of medical data machine learning.

Using a Prebuilt Model

Available models:

Model Name Model description Contact Details for Usage
LGI/Colon-Flag Detects colon cancer using age, sex, and CBCs Roche
LungFlag Detects lung cancer using age, sex, smoking infromation, and common blood tests Roche
GastroFlag Detects gastric cancer using age, sex, and common blood tests Roche
AAA Predicts AAA events Geisinger/TBD
FluComplications Predicts flu followed by complications such as pneumonia, hospitalization, or death TBD
Pred2D Predicts progression from prediabetes to diabetes Planned to be open source
FastProgressors Predicts rapid decline in eGFR Planned to be open source
MortatlityCMS Predicts mortality using CMS claims data TBD
Unplanned COPD Admission Prediction Model Predicts COPD hospitalization using CMS claims data TBD

There are two options for using a model:

Creating a new model

If you want to create a new model, please follow those steps:

  1. Set up your environment: Compile and clone the necessary tools
  2. Create a data repository: Follow the ETL guide
  3. Define your cohort: Prepare a list of patient IDs, prediction times, and outcome. See MedSamples for details on labeling. In this step you will create a file where each patient has a timestamp for prediction and a label in a CSV format.
  4. Specify the model architecture in JSON. Refer to: Model Json Format, infrastucture explaination
  5. Run our tools to train the model with the model architecture, samples and data repository - python or Flow, Optimizer to train the model, or python code
  6. Apply and test the model using Flow or python code
  7. Analyze model performance using bootstrap and test kit/AutoTest or write your own tests
  8. Wrap the model for deployment - follow - AlgoMarker Deployment

Other Documentation Pages