podfastio/hith-eligibility-clinicalbert

Hospital-in-the-Home (HITH) Eligibility Classifier

A fine-tuned Bio_ClinicalBERT model that predicts whether a patient is suitable for Hospital-in-the-Home care based on clinical features.

Model Description

This model uses a pseudo-note fusion approach (inspired by the MEME paper) to combine structured and unstructured clinical data into a single text input for classification:

• Presenting complaint → text
• Ward → text
• Length of stay → numeric, serialized to text
• Primary diagnosis → text
• Nurse handover notes → free text
• Allied health notes → free text

All features are serialized into a structured pseudo-note format and classified using Bio_ClinicalBERT with a classification head.

Input Format

The model expects a single text string in pseudo-note format:

PRESENTING COMPLAINT: [complaint]. WARD: [ward]. LENGTH OF STAY: [days] days. PRIMARY DIAGNOSIS: [diagnosis]. NURSE HANDOVER: [nurse notes] ALLIED HEALTH: [allied health notes]

Performance (Synthetic Data)

Metric	Score
AUROC	1.000
AUPRC	1.000
F1	1.000

⚠️ Note: These results are on synthetic training data with clear class separation. Real-world performance will differ and the model should be retrained on actual clinical data before any clinical use.

Training Details

• Base model: emilyalsentzer/Bio_ClinicalBERT (110M params)
• Training data: 2,000 synthetic clinical records
• Train/Val/Test split: 70/15/15
• Max sequence length: 512 tokens
• Learning rate: 2e-5
• Batch size: 16
• Epochs: 4 (early stopping, patience=3)
• Optimizer: AdamW with weight decay 0.01
• Hardware: NVIDIA T4 GPU

Intended Use

This is a prototype/research model for exploring HITH eligibility prediction. It is trained on synthetic data and is NOT suitable for clinical decision-making without:

1. Retraining on real clinical data
2. Extensive clinical validation
3. Regulatory approval
4. Integration with clinical governance frameworks

Key Features (SHAP Analysis from XGBoost baseline)

The most predictive features identified by SHAP analysis:

• Pro-HITH: "stable", "improving", "drinking", "mobilising", "independent", "cellulitis"
• Anti-HITH: "cardiac monitoring", "ICU", "ventilated", "sedated", "high acuity", "bed rest"

Two-Model System

This repository contains the ClinicalBERT model. A companion XGBoost + TF-IDF baseline model is also available, providing:

• Faster inference
• Better interpretability via SHAP feature importance
• Comparable performance on this task

Citation

If you use this model, please cite:

• Bio_ClinicalBERT: Alsentzer et al. (2019) "Publicly Available Clinical BERT Embeddings"
• MEME approach: Yang et al. (2024) "MEME: Machine Learning Enhanced Medical Embeddings for ED Disposition Prediction"