Exploring the performance and explainability of fine-tuned BERT models for neuroradiology protocol assignment

Salmonn Talebi, Elizabeth Tong, Anna Li, Ghiam Yamin, Greg Zaharchuk, Mohammad R. K. MofradBMC Medical Informatics Decis. Mak 2024 · OpenReview

Abstract

Deep learning has demonstrated significant advancements across various domains. However, its implementation in specialized areas, such as medical settings, remains approached with caution. In these high-stake environments, understanding the model's decision-making process is critical. This study assesses the performance of different pretrained Bidirectional Encoder Representations from Transformers (BERT) models and delves into understanding its decision-making within the context of medical image protocol assignment.

Introduction

In clinical radiology, "protocoling" is the essential task of reviewing a patient's medical history to determine the specific technical parameters for an MRI or CT scan. This manual review process is time-intensive, accounting for approximately 6% of a radiologist's workload and contributing significantly to professional burnout. While traditional machine learning models like Random Forests have been applied to this task, they often struggle with the idiosyncratic, shorthand-heavy language found in physician notes.

This study utilizes a dataset of 88,052 clinical notes from Stanford University to automate the assignment of the ten most common neuroradiology protocols (e.g., MR Stroke, MR Seizure). By leveraging Transformer-based models, the researchers aim to bridge the gap between automated efficiency and the nuanced reasoning required for patient safety.

Methodology and Interpretability

The researchers evaluated four Transformer-based architectures: BERT, RoBERTa, BioBERT, and ClinicalBERT. To ensure the models were not merely "black boxes," two interpretability techniques were employed:

  1. Integrated Gradients: This method assigns importance scores to individual words, highlighting which terms most heavily influenced the final protocol selection.
  2. Erasure (Stress Testing): This technique involves systematically removing high-importance words to observe the resulting degradation in model performance, thereby validating the reliability of the importance scores.
The end-to-end workflow from physician notes through the BERT model to an explainable decision output.

The overall pipeline, illustrated above, demonstrates how raw clinical text is processed by the BERT architecture to produce a protocol decision accompanied by a visualization of word importance.

Key Results

BioBERT, pre-trained on biomedical literature, emerged as the top performer with a weighted average F1 score of 0.92. This significantly outperformed baseline models like XGBoost (0.84). Notably, BioBERT achieved parity with a senior human neuroradiologist in specific categories such as "MR seizure" and "MR moya-moya," though the human expert maintained a higher overall accuracy of 0.97.

Analysis of the model's errors revealed that 52% were due to "multifarious choices" where the clinical question was overly complex, while 15% were related to the model's failure to account for patient age—a critical factor in protocol selection.

Insights into Model Reasoning

The use of Integrated Gradients provided a window into the model's "logic." While the models were highly effective at identifying "anchor" keywords, they lacked the deep anatomical and hierarchical reasoning of human experts.

Visualization of word importance using Integrated Gradients, showing how terms like 'breast cancer' and 'numbness' drive the model's decision.

As shown in the figure above, the model places high importance on specific clinical findings. However, error analysis indicated that the models often prioritize generic procedural terms over specific anatomical structures (e.g., "hippocampus"). Furthermore, when faced with notes containing multiple conflicting clinical indications (e.g., both tumor and stroke symptoms), the models struggled to determine which protocol should take precedence—a task human radiologists handle through established medical hierarchies.

Conclusion and Future Directions

While fine-tuned BERT models show promise in automating radiology protocoling, their current "statistical" understanding of text remains a hurdle for full clinical integration. The reliance on single-center data from Stanford suggests a need for multi-institutional validation to ensure generalizability across different clinical "dialects."

Future research will focus on labeling individual imaging acquisitions rather than broad protocols and exploring the reasoning capabilities of larger generative models like GPT-4 to address the identified gaps in anatomical and hierarchical understanding.

Summary updated

Abstract

Deep learning has demonstrated significant advancements across various domains. However, its implementation in specialized areas, such as medical settings, remains approached with caution. In these high-stake environments, understanding the model's decision-making process is critical. This study assesses the performance of different pretrained Bidirectional Encoder Representations from Transformers (BERT) models and delves into understanding its decision-making within the context of medical image protocol assignment.
BibTeX
@inproceedings{talebi2024exploring,
  author = {Salmonn Talebi and Elizabeth Tong and Anna Li and Ghiam Yamin and Greg Zaharchuk and Mohammad R. K. Mofrad},
  title = {{Exploring the performance and explainability of fine-tuned BERT models for neuroradiology protocol assignment}},
  booktitle = {BMC Medical Informatics Decis. Mak},
  year = {2024},
  url = {https://openreview.net/forum?id=SY2zsnbc0F},
}
Figure