MedPython 1.1.0 Release Notes

Relase Notes - 1.1.0

• Modularized Package Structure
  • Decoupled medpython-etl into a standalone package.
  • medpython now acts as a meta-package, including medpython-etl as a core dependency for user convenience without introducing direct code coupling.
• Enhanced med.Model Functionality
  • Runtime Model Updates: Introduced apply_model_changes to modify models on the fly (e.g., removing explainability or lowering batch counts to save memory) without requiring retraining.
  • Post-Processor Integration: Added add_post_processors_json_string_to_model to attach and train new components such as calibrators, explainers, or fairness adjustments onto existing/trained models.
  • Component Management: New train/remove/print commands for rep-processors and post-processors, allowing direct execution of training functions for specific model components.
  • Predictor Export: Added functionality to retrieve the core predictor in standard scikit-learn or XGBoost formats (for supported predictors)
  • Architecture Visualization: Added the ability to retrieve the MedModel architecture in a human-readable, printable format.
  • Usage Validation: Implemented strict state checks to prevent logical errors (e.g., ensuring apply cannot be called before the model is trained).
  • Signal Debugging: Introduced debug_rep_processor_signal to inspect how rep-processors transform raw signals, outlier, etc.
• Platform & Build Improvements
  • macOS Compatibility: Resolved all compilation issues for macOS environments using Conda, ensuring a smoother installation process for Apple Silicon and Intel-based Macs.