Add LabTOP model

[ruokun-niu]

UIUC ID: ruokunn2

Summary

This PR adds the LabTOP (Lab Test Outcome Prediction) model to PyHealth, enabling continuous numerical prediction of laboratory test values.

LabTOP uses digit-wise tokenization to represent numerical values as sequences of individual digits (e.g., 123.45 → ['1','2','3','.','4','5']), preserving exact precision while maintaining a compact vocabulary of ~20-50 tokens.

Paper Reference

• Title: LabTOP: A Unified Model for Lab Test Outcome Prediction on Electronic Health Records
• Authors: Sujeong Im, Jungwoo Oh, Edward Choi
• Conference: CHIL 2025 (Best Paper Award)
• arXiv: https://arxiv.org/abs/2502.14259
• Code: https://github.com/sujeongim/LabTOP

Implementation Details

• Architecture: GPT-2 transformer (12 layers, 768 dim, ~53M parameters)
• Classes Added:
  • DigitWiseTokenizer: Converts numbers ↔ digit sequences
  • LabTOPVocabulary: Manages complete vocabulary (special tokens + digits + lab codes)
  • LabTOP: Main model class inheriting from BaseModel

Files Modified

• ✅ pyhealth/models/labtop.py (new file, ~600 lines)

Performance (from paper)

• MAE: 0.064, SMAPE: 14.80%, NMAE: 0.042 on MIMIC-IV (44 lab types)

[Logiquo]

Could you add some test case for this model? thanks.