Reviewed by Lexie CornerMar 13 2024
A novel AI language model can identify clinical symptoms in medical summaries and correlate them with brain tissue data from donors of the Netherlands Brain Bank.
This process offers fresh perspectives on individual disease progression and enhances comprehension of prevalent misdiagnoses of brain disorders. Ultimately, the model holds promise for aiding the diagnostic process, potentially leading to more precise outcomes.
Creating new treatment options for brain diseases can be difficult because the underlying molecular mechanisms are often poorly understood. Investigating these mechanisms is challenging due to the complex relationship between tissue abnormalities and a patient's symptoms. Some symptoms can occur in multiple conditions, and the clinical presentation can vary significantly from patient to patient, leading to a high rate of misdiagnosis (up to 30 %). However, a newly developed AI language model could change this situation by providing valuable insights.
The Netherlands Brain Bank houses brain tissue obtained from 3,042 donors who suffered from various brain disorders. What sets the institution apart is that, along with the tissue, it also maintains records of the donors' medical histories and symptoms. However, this data was not easily quantifiable as it was transcribed in a text format, thus posing a challenge in terms of analyzing and processing it.
Language Model
To unlock valuable medical information, Inge Huitinga and her team at the Netherlands Institute for Neuroscience collaborated with Inge R. Holtman and her team at the University Medical Center Groningen to utilize a new AI language model. This model can analyze the text in medical records and detect specific symptoms as defined beforehand.
A second AI prediction model was created to diagnose based on the patient's clinical presentation.
First, the records had to be thoroughly examined to identify symptoms that regularly occur in donors with different brain diseases. We eventually identified 90 different symptoms in five different domains: psychiatric symptoms (such as depression and psychosis), cognitive symptoms (such as dementia and memory problems), motor issues (such as tremors), and sensory symptoms (such as feeling things that are not there). We then manually labeled 20,000 sentences to train the classification model.
Inge R. Holtman, Department of Biomedical Sciences, University Medical Center Groningen
After considering all the factors, the final model was able to determine the symptoms that occurred in donors annually. Despite the prediction model's effectiveness in making accurate diagnoses, it could not effectively diagnose rare disorders.
Upon analyzing the diagnoses made by the prediction model, a group of donors who had received incorrect diagnoses emerged. Upon further investigation, it was discovered that a considerable number of these donors had also been misdiagnosed by their doctor at some point during their lifetime.
Subtypes
Inge Holtman added, “It seems that there is a group of people suffering from a certain condition, such as Alzheimer’s disease, but exhibiting symptoms more reminiscent of Parkinson’s disease. Or a subtype of Frontotemporal Dementia manifesting as Alzheimer’s disease. It is often challenging to diagnose these groups properly, which makes sense since these individuals show a clinical pattern that does not align with their condition. We strive to continuously improve the prediction model, hoping to make diagnoses of brain diseases more accurate.”
Understanding individual factors contributing to symptoms in brain diseases is crucial, as the reality is that many people have a combination of different conditions. Molecular markers to guide treatment are the future. Our ultimate goal is to develop a molecular atlas of symptoms of brain diseases. Such an atlas precisely shows which cells and molecules in the brain change with symptoms such as anxiety, forgetfulness, and depression.
Inge Huitinga, Netherlands Institute for Neuroscience
Inge Huitinga added, “We expect the impact of this molecular atlas to be enormous. When we map out the molecular changes, we hope to identify the first biomarkers that can predict the correct diagnosis during a person’s lifetime. This opens doors to the development of new therapies. We are laying the foundation.”
The study was funded by the Friends Foundation from the Netherlands Institute for Neuroscience.
Journal Reference:
Mekkes, J. N., et al. (2024) Identification of clinical disease trajectories in neurodegenerative disorders with natural language processing. Nature Medicine. doi.org/10.1038/s41591-024-02843-9