educator-anchored-hitl-pubmedbert

PUBMEDBERT Educator-Anchored HITL : Educator-Anchored HITL Clinical Reasoning Classifier

This model is a fine-tuned version of microsoft/BiomedNLP-BiomedBERT-base-uncased-abstract-fulltext trained on the MedNLI dataset, used as a proxy for clinical reasoning scenarios in healthcare education.

It was refined using an educator-anchored Human-in-the-Loop (HITL) workflow across 17 rounds of incremental, simulated educator-guided corrections, as described in:

Educator-Anchored Human-in-the-Loop Learning: A Simulation Study of Transformer Models for Clinical Reasoning Assessment in Healthcare Education — [citation pending]

⚠️ Important: Simulation Study Disclaimer

This model is a research artefact from a simulation study, not a clinically validated or deployment-ready tool. Specifically:

• Training data (MedNLI) is a proxy for clinical reasoning, sourced from MIMIC-III clinical notes — it is not authentic learner submissions from a healthcare education context.
• HITL "educator corrections" during refinement were simulated (automated ground-truth relabelling of misclassified examples), not provided by real human educators.
• Do not use this model for actual clinical decision-making, patient safety assessment, or student grading without further validation by qualified healthcare educators.

This model is intended for reproducibility and further research only.

Labels

Usage

from transformers import pipeline

clf = pipeline(
    "text-classification",
    model="Nita200/educator-anchored-hitl-pubmedbert",
)
rresult = clf(
    "She was evaluated by neurosurgery, deemed to be intact neurologically. "
    "[SEP] Her neurological exam was normal and no deficits were identified."
)
print(result)

Training

• Base model: microsoft/BiomedNLP-BiomedBERT-base-uncased-abstract-fulltext
• Dataset: MedNLI (PhysioNet credentialed access required), original 80/10/10 train/validation/test split preserved
• HITL configuration: v3 (catastrophic-forgetting-mitigated)
• HITL rounds completed: 17 (stops early if the correction pool is exhausted before reaching the maximum of 20 rounds)
• Corrections per round: 50
• Replay buffer size: 100 (seed examples resampled each round to anchor prior representations and prevent catastrophic forgetting)
• Learning rate: 5e-6
• Epochs per round: 1
• Seed/pool split: 70% seed / 30% pool

This configuration was selected after a systematic three-version comparison (see paper Section 4.3 and 5.2) showing that a naive incremental fine-tuning configuration (higher learning rate, larger correction batches, no replay buffer) produces catastrophic forgetting. Five-fold cross-validation (paper Section 5.7) confirms that the AUC stability achieved under this configuration generalises across independent data splits, while the magnitude of accuracy improvement in any single run is split-dependent.