https://github.com/soda-inria/predictive-ehr-benchmark

Exploring a complexity gradient in representation and predictive algorithms for EHRs
https://github.com/soda-inria/predictive-ehr-benchmark

Last synced: 1 day ago
JSON representation

Exploring a complexity gradient in representation and predictive algorithms for EHRs

• Host: GitHub
• URL: https://github.com/soda-inria/predictive-ehr-benchmark
• Owner: soda-inria
• License: other
• Created: 2023-09-13T15:12:28.000Z (over 1 year ago)
• Default Branch: main
• Last Pushed: 2023-09-14T11:59:16.000Z (over 1 year ago)
• Last Synced: 2024-11-06T21:19:16.271Z (about 2 months ago)
• Language: Python
• Homepage: https://soda-inria.github.io/predictive-ehr-benchmark/
• Size: 76 MB
• Stars: 0
• Watchers: 7
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# Predictive algorithms from Electronic Health Records 

This repository hosts code for the working paper: *Exploring a complexity gradient in representation and predictive algorithms for EHRs* 

[**Documentation**](https://soda-inria.github.io/predictive-ehr-benchmark/)

[**Source Code**](https://github.com/soda-inria/predictive-ehr-benchmark)

[**Working Paper repository**](https://github.com/strayMat/predictive_ehr_paper)

### Abstract

Electronic Health Records contain time-varying features with high cardinality.

Current state-of-the-art predictive models build on increasingly elaborated

pipelines --based on transformers-- to handle the complexity of these data.

Acknowledging the complexity to deploy, transfer and adapt these models on local

care environments, we explore a complexity-benefit tradeoff by comparing them to

simple aggregation of events. We use three clinical tasks involving time-varying

structured Electronic Health Records (EHRs) and increasingly clinically relevant

problems. We show that these benchmarking tasks display heterogeneous predictive

difficulties. We introduce a simple aggregation of static embeddings

--transferred from national claims and publicly available--, showing that it

outperforms transformer-based models on simple tasks with medium sample sizes.

We highlight the sample and computing resource efficiency of these models.

Finally, clinically relevant problems generally present a strong class

imbalance, which complicates models development and undermines their

performances. Further work is needed to understand if transformer-based models

perform well in these scenarios where the number of cases requires good sample

efficiency.

# Usage

See the [usage page on the documentation](https://soda-inria.github.io/predictive-ehr-benchmark//usage.html)