Sign in Subscribe
Machine learning

Unlocking Critical Details in Medical Records with AI

Strategic Insights

— 1 min read

Researchers at the MUSC Hollings Cancer Center have developed a high-accuracy AI model using natural language processing (NLP) to extract critical details from medical records. The model can help doctors determine important information for patient treatment, particularly in cases of metastatic tumors affecting the brain. By automating data extraction from unstructured notes, the researchers aim to improve cancer classification, enable more precise care, and accelerate research efforts.

Key Takeaways:

  • The researchers developed an NLP model that can "read" doctors' notes and identify key words and phrases indicating the primary cancer type, with an accuracy rate of over 90% in 82,000 clinical notes.
  • The model was tested on doctors' notes from the medical records of more than 1,400 patients treated with SRS for brain metastases and correctly identified the primary cancer in more than 90% of cases.
  • The NLP model outperformed traditional medical codes, such as ICD codes, in determining a patient's original cancer diagnosis, particularly for common cancers like lung, breast, and skin cancer, where classification was nearly perfect at 97%.
  • The model was designed to be simple and efficient, requiring minimal resources and avoiding many of the ethical concerns associated with larger, generative AI models.
  • The research team is now working on a study using a similar NLP approach to identify patients at risk for radiation necrosis, a rare but serious side effect of too much radiation.
  • Future researchers could use the NLP model with other health systems, other cancer types, or add in health data, such as imaging scans or lab tests.

Statistics:

  • The NLP model was tested on 82,000 clinical notes from the medical records of more than 1,400 patients treated with SRS for brain metastases.
  • The model correctly identified the primary cancer in more than 9,000 cases (out of 10,287 total cases tested, including 82,000 notes).
  • The accuracy rate for common cancers like lung, breast, and skin cancer was nearly perfect at 97% (945 out of 968 cases tested).
  • The model could even identify lung cancer subtypes, which ICD codes were unable to do.
  • The model requires minimal resources and can be scaled for use in other hospitals with limited resources.

Sources:

  • Obeid, J., and Fugal, M. (2023) "Unlocking Critical Details in Medical Records with AI." JCO Clinical Cancer Informatics.
  • MUSC Hollings Cancer Center.

Read more