Major Refactor: Unified Event Stream, YAML Config, Multimodal Processor, Simplified Model

[zzachw]

This PR delivers a full-stack refactor of PyHealth’s data and model pipeline.

• Unified Event Stream

  • All patient data is now represented as a unified event stream (Polars DataFrame).
  • Simplifies data handling across modalities and supports flexible, on-demand querying.
  • Patient object redesigned to work directly with the event stream for efficient event access and filtering.

• YAML Configuration

  • Dataset loading is fully driven by YAML config files.
  • Clean, declarative specification of table paths, attributes, and data sources.
  • Enables ease of extension to new datasets.

• Multimodal Processor Framework

  • Introduced a modular processor system for handling diverse data types (sequences, signals, images, labels, time series).
  • Processors are easily extendable and managed through a centralized registry.
  • Standardized interface for preprocessing, feature extraction, and data transformation.

• Simplified Model Structure

  • Refactored model classes for clarity and modularity.
  • Added EmbeddingModel to unify input embedding across models.
  • Cleaned up BaseModel and RNN class implementations, removing deprecated logic and improving readability.

• Additional Improvements

  • Adopted Polars backend for faster data manipulation.
  • Unified dataset APIs and removed legacy code.
  • New tasks added: InHospitalMortalityMIMIC4, Readmission30DaysMIMIC4.
  • Updated .gitignore and utility functions for YAML parsing.
  • Improved internal typing and documentation.

⚠️ Breaking Changes & Follow-Up

• This is a major refactor that may break existing functionalities.
• Extensive testing and integration with downstream modules are still required.
• Documentation is outdated and will need a full revision to match the new pipeline.
• Reviewers should focus on overall structure and architecture.
• Follow-up PRs will address testing, doc updates, and additional validation.

[Copilot]

Copilot reviewed 30 out of 30 changed files in this pull request and generated no comments.

Comments suppressed due to low confidence (2)

pyhealth/processors/image_processor.py:75

• [nitpick] The repr method in ImageProcessor returns 'ImageLoadingProcessor', which is inconsistent with the class name. Consider updating it to 'ImageProcessor' to reflect the correct class.

f"ImageLoadingProcessor(image_size={self.image_size}, to_tensor={self.to_tensor}, normalize={self.normalize}, mean={self.mean}, std={self.std})"

pyhealth/data/data.py:35

• The filtering of dictionary keys in Event.from_dict assumes that attribute keys are prefixed with the event_type. This behavior might be fragile if keys are not uniformly prefixed; consider adding validation or documentation to ensure consistency.

for k, v in d.items() if k.startswith(event_type)

[jhnwu3]

Looks good tbh

[jhnwu3]

I see no red flags.

[jhnwu3]

Signal processor may need Jathurshan to look at later.