Sign in Subscribe
Coronaviruses

COVID-19 Diagnosis and Large Language Models: Causal Network Analysis

Strategic Insights

— 1 min read

Researchers from George Mason University have conducted a groundbreaking study on the diagnosis of COVID-19 using causal network analysis and large language models (LLMs). Their research, published in the journal Health Care Management Science, presents a method for converting a causal network into a LLM, demonstrating its potential to improve diagnosis accuracy. The study found that the causal network model outperformed the LLM when indirect information was available, emphasizing the importance of proactively asking patients about missing, indirect information.

Key Takeaways:

  • The researchers developed a method to convert a causal network into a LLM, enabling the use of large language models in COVID-19 diagnosis.
  • The causal network model achieved higher accuracy (AUROC = 0.91) than the LLM (AUROC = 0.88) when indirect information was available.
  • The accuracy of the LLM depended not only on direct predictors of the outcome but also on data not reported to the LLM.
  • The researchers concluded that conversational LLMs should proactively ask about missing, indirect information to improve diagnosis accuracy.
  • The study involved two databases: one with 822 patients reporting direct and indirect symptoms, and another with 80 patients reporting open-ended questions.
  • The accuracy of the causal network and Markov blanket was tested using the Area under the Receiver Operating Curve (AUROC).

Statistics:

  • 822 patients were surveyed, collecting 12 direct and 7 indirect symptoms of COVID-19.
  • 80 patients reported their symptoms in open-ended questions, often reporting direct predictors and rarely reporting indirect predictors.
  • The causal network model achieved an AUROC of 0.91 when indirect information was available.
  • The LLM model achieved an AUROC of 0.88 when indirect information was available.
  • The accuracy of the LLM depended on patterns among direct predictors and unreported indirect information.

Sources:

  • Causal networks guiding large language models: application to COVID-19. Health Care Management Science, 2025.
  • NewsRx. Researchers from George Mason University Discuss Findings in COVID-19 (Causal networks guiding large language models: application to COVID-19). Medical Letter on the CDC & FDA. November 2, 2025; p 224.

Read more