Refer to the project report for more details.

Detecting Hallucinations in Medical Question Answering: A Comparative Study of BERT and BioBERT

This project investigates hallucination detection in medical question answering systems using transformer-based models. I compare a general-purpose language model (BERT) with a biomedical domain-specific model (BioBERT) to evaluate their effectiveness in detecting hallucinated medical responses.

---

Project Overview

Large Language Models (LLMs) can generate fluent but factually incorrect outputs (hallucinations), especially in high-stakes domains like healthcare. Detecting such hallucinations is essential for safe deployment.

This project formulates hallucination detection as a binary classification problem:

Given a medical question, context, and answer: predict whether the answer is grounded (0) or hallucinated (1).

Evaluate both in-domain and out-of-domain generalization.

---

Datasets

MedHallu (Training + In-domain Evaluation)

Source: https://arxiv.org/abs/2502.14302

Dataset: https://huggingface.co/datasets/UTAustin-AIHealth/MedHallu

MedHallu is a medical hallucination detection benchmark with fine-grained annotations.

I use it to construct a binary classification dataset:

• (Question + Context, Ground Truth Answer) → Label 0 (non-hallucinated)
• (Question + Context, Hallucinated Answer) → Label 1 (hallucinated)

Categories include:

• Misinterpretation of Question
• Incomplete Information
• Mechanism and Pathway Misattribution
• Methodological and Evidence Fabrication

---

Med-HALT (Out-of-domain Evaluation)

Source: https://arxiv.org/abs/2307.15343

Dataset: https://huggingface.co/datasets/openlifescienceai/Med-HALT

Med-HALT is used to evaluate robustness under distribution shift and reasoning stress tests.

It includes three sub-tasks:

• FCT (False Confidence Test)
• NOTA (None of the Above Test)
• FQT (Fake Questions Test)

---

Models

I compare two transformer-based sequence classification models:

• BERT-base-uncased
• BioBERT-base-cased (biomedical pretraining)

Both models are fine-tuned using HuggingFace Transformers for binary classification.

---

Methodology

Input Representation

Each training example is constructed as: Question + Answer → Candidate Answer

---

Task Formulation

Binary classification:

• 0 → Grounded / correct answer
• 1 → Hallucinated / incorrect answer

---

Training Setup

• Framework: HuggingFace Transformers
• Loss Function: Cross-Entropy Loss
• Optimizer: AdamW
• Epochs: 5
• Batch Size: 8
• Max Sequence Length: 256

---

Results

In-Domain Performance (MedHallu)

Model	Accuracy	F1 Score
BERT	~0.94–0.95	~0.94
BioBERT	~0.95–0.96	~0.96

BioBERT consistently outperforms BERT in the medical domain.

---

Out-of-Domain Performance (Med-HALT)

Model	Accuracy	F1 Score
BERT	~0.43	~0.33
BioBERT	~0.56	~0.59

BioBERT shows significantly better generalization under distribution shift.

---

Med-HALT Task Breakdown

Task	BERT F1	BioBERT F1
FCT	~0.30	~0.49
NOTA	~0.35	~0.65
FQT	0.00	0.00

---

Key Findings

• Domain-specific pretraining (BioBERT) improves hallucination detection performance.
• Strong in-domain results do not guarantee out-of-domain robustness.
• Both models struggle significantly with Fake Question (FQT) tasks.
• Hallucination detection remains challenging under reasoning shifts and adversarial inputs.

---

Repository Structure

.
├── hallucination_classification_bert.ipynb
├── hallucination_classification_biobert.ipynb
├── hallucination_detection_in_medical_QA_biobert_vs_bert.pdf
└── README.md