Abstract
Coronavirus disease 2019 (COVID-19) has infected more than one million individuals all over the world and caused more than 55,000 deaths, as of April 3 in 2020. Radiological findings are important sources of information in guiding the diagnosis and treatment of COVID-19. However, the existing studies on how radiological findings are correlated with COVID-19 are conducted separately by different hospitals, which may be inconsistent or even conflicting due to population bias. To address this problem, we develop natural language processing methods to analyze a large collection of COVID-19 literature containing study reports from hospitals all over the world, reconcile these results, and draw unbiased and universally-sensible conclusions about the correlation between radiological findings and COVID-19. We apply our method to the CORD-19 dataset and successfully extract a set of radiological findings that are closely tied to COVID-19.
Introduction
In the early stages of the COVID-19 pandemic, the medical community faced a deluge of reports from individual hospitals. While valuable, these studies often suffered from population bias, where radiological findings (observed via X-rays and CT scans) varied significantly due to local genetics, social factors, or ethnic characteristics. This led to a fragmented clinical picture with sometimes conflicting conclusions. To address this, researchers from UC San Diego developed an automated Natural Language Processing (NLP) pipeline to extract and aggregate findings from thousands of medical papers, aiming to identify universal clinical markers while filtering out local noise.
Methodology
The researchers built a multi-stage NLP pipeline designed to handle the lack of initial COVID-specific training data. The process involved:
- Sentence Classification: A BERT model (Bidirectional Encoder Representations from Transformers), pretrained on biomedical literature, was used to identify sentences describing radiological findings.
- Iterative Hard Example Mining: Initially, the model used general radiology reports from the MedPix database as positive samples. To refine this for COVID-19, the team used an iterative "hard example mining" approach. The model predicted findings, humans labeled the errors (hard examples), and the model was retrained. This process significantly boosted precision by forcing the model to learn finer distinctions.
- Constituent Parsing and Frequency Analysis: Once relevant sentences were identified, constituent parsing generated syntax trees to isolate specific noun phrases (e.g., "bilateral pleural effusion"). These phrases were then cleaned, lemmatized, and ranked by frequency across the CORD-19 dataset.
Results and Discussion
The iterative bootstrapping approach proved highly effective in improving the model's accuracy. Starting from a baseline precision of 47% using only general medical data, the precision jumped to 69% after three rounds of hard example mining.
The final model achieved a precision rate of 71.8% upon manual verification. By aggregating findings across the corpus, the system successfully identified the most critical COVID-19 markers recognized by the medical community, such as consolidation, pleural effusion, and ground glass opacity. This aggregation acts as a filter, allowing the true clinical "signal" of the disease to rise above the noise of individual hospital variations.
Conclusion
This research demonstrates that highly specific tools for emerging diseases can be built by leveraging general medical databases and refining them through targeted manual effort. While the current method focuses on noun phrases and requires some human-in-the-loop labeling, it provides a scalable way to reconcile disparate medical reports. Future work will aim to capture the temporal dynamics of these findings and the contextual relationships between different clinical markers.
Summary updated