cuimiandashi/gvhd-analysis

GVHD Severity Prediction - Final Results

Best Model: Stacking Ensemble (CatBoost + XGBoost + Neural Net)

Metric	Value
AUC	0.7083 ± 0.0117
Baseline AUC	0.7034
Improvement	+0.0049 (+0.7%)
Brier (Platt Calibrated)	0.2019
Optimal Threshold	0.544
Sensitivity	68.4%
Specificity	58.8%
PPV	76.9%
NPV	48.3%

Model Comparison

Model	AUC Mean	AUC Std	Fold1	Fold2	Fold3	Fold4	Fold5
CatBoost	0.6963	±0.0105	0.700	0.711	0.699	0.679	0.694
XGBoost	0.6986	±0.0126	0.705	0.711	0.704	0.675	0.698
NeuralNet	0.6870	±0.0088	0.692	0.699	0.698	0.677	0.681
Stacking	0.7083	±0.0117	0.714	0.722	0.714	0.688	0.703

Calibration

Method	Brier Score
Raw	0.2150
Platt Scaling	0.2019
Isotonic Regression	0.2024

Key Improvements

1. Feature Engineering: interactions, polynomials, log transforms, missingness indicators
2. GPU Acceleration: Tesla T4 for CatBoost and Neural Net
3. Stacking Ensemble: Logistic Regression meta-learner on OOF predictions
4. Probability Calibration: Platt scaling for clinical deployment

Files

• gvhd_gpu_pipeline.py - Complete pipeline code (every line commented)
• result_comparison_final.csv - Model comparison table
• GVHD_Final_Report.ipynb - Jupyter notebook with tables
• calibration_plot.png - Calibration curve

Honest Ceiling

Pre-transplant-only models plateau at AUC ≈ 0.71. Any claim > 0.75 requires post-transplant biomarkers (Day 7-14).

Generated by ML Intern

This model repository was generated by ML Intern, an agent for machine learning research and development on the Hugging Face Hub.

• Try ML Intern: https://smolagents-ml-intern.hf.space
• Source code: https://github.com/huggingface/ml-intern

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "cuimiandashi/gvhd-analysis"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

For non-causal architectures, replace AutoModelForCausalLM with the appropriate AutoModel class.