Skip to contents

Changelog

Source: NEWS.md

RHealth 0.2.0

Initial CRAN Release

This is the first CRAN submission of RHealth, a comprehensive deep learning toolkit for healthcare predictive modeling in R.

Major Features

Dataset Module
  • Support for multiple EHR database formats:
    • MIMIC-III and MIMIC-IV datasets
    • eICU Collaborative Research Database
    • OMOP Common Data Model
    • EHRShot benchmark datasets
  • Built-in data harmonization and caching with DuckDB
  • Dev mode for rapid prototyping with patient subsets
  • Flexible dataset splitting by patient, visit, or sample
  • YAML-based configuration for custom datasets
Model Module
  • Implementation of state-of-the-art deep learning architectures:
    • RNN (Recurrent Neural Networks)
    • CNN (Convolutional Neural Networks)
    • Transformer models with multi-head attention
    • AdaCare (Adaptive feature correlation and feature-aware attention)
    • ConCare (Personalized clinical predictions via attention mechanisms)
  • Embedding layers for medical code representations
  • Automatic dimension calculation and loss function selection
  • CPU and GPU support via torch
Task Module
  • Pre-built prediction tasks:
    • In-hospital mortality prediction (MIMIC-III, MIMIC-IV)
    • Next-period mortality prediction
    • 30-day readmission prediction (MIMIC-IV, OMOP)
    • Mortality prediction for eICU and OMOP datasets
    • EHRShot benchmark tasks
  • Extensible base classes for custom task definitions
  • Support for multiple input/output schemas (binary, multiclass, multilabel, regression)
Medical Code Module
  • Medical terminology lookup and hierarchical navigation
  • Cross-system code mapping:
    • ICD-9-CM ↔︎ ICD-10-CM
    • ICD-9-CM/ICD-10-CM → CCS (Clinical Classifications Software)
    • NDC → ATC (Anatomical Therapeutic Chemical)
    • RxNorm → ATC
  • Ancestor and descendant code retrieval
  • Drug-drug interaction checking
Trainer Module
  • Unified training interface with validation and testing
  • Multiple evaluation metrics (AUROC, AUPRC, accuracy, F1, calibration)
  • Early stopping and model checkpointing
  • Support for custom metrics and loss functions
  • Progress tracking and logging
Data Processing
  • Flexible processors for different data types:
    • Time series sequences
    • Binary, multiclass, and multilabel labels
    • Text data
    • Feature tensors
  • Automatic padding and batching with DataLoader integration

Infrastructure

  • R6-based object-oriented design
  • Integration with torch ecosystem
  • Comprehensive documentation and vignettes
  • Unit tests for core functionality
  • GitHub repository with CI/CD workflows

Funding

RHealth is funded by the ISC grant from the R Consortium.

RHealth is the R counterpart to the Python library PyHealth.