AKG2/hybrid-readmission-classifier

Hybrid Readmission Classifier

Model Description

This is a hybrid model combining:

• Text embeddings: Extracted from emilyalsentzer/Bio_ClinicalBERT
• Structured features: Patient demographics and admission details
• Classifier: LightGBM gradient boosting

Performance

• ROC-AUC: 0.6602
• AUPRC: 0.3334
• Accuracy: 0.6386
• F1-Score: 0.3871

Model Artifacts

• lgbm_model.txt: LightGBM model
• scaler_combined.pkl: StandardScaler for combined features
• scaler_struct.pkl: StandardScaler for structured features
• label_encoders.pkl: Label encoders for categorical features
• metadata.json: Model metadata

Usage

import joblib
import numpy as np
from transformers import AutoTokenizer, AutoModel

# Load components
lgbm_model = joblib.load('lgbm_model.txt')
scaler = joblib.load('scaler_combined.pkl')
tokenizer = AutoTokenizer.from_pretrained('emilyalsentzer/Bio_ClinicalBERT')
embedding_model = AutoModel.from_pretrained('emilyalsentzer/Bio_ClinicalBERT')

# Extract embeddings
embeddings = embedding_model(tokenizer(text)['input_ids'])

# Combine with structured features and predict
X = np.hstack([embeddings, structured_features])
X_scaled = scaler.transform(X)
predictions = lgbm_model.predict(X_scaled)

License

MIT