A new deep learning-based model developed in South Korea has shown potential to reduce the time it takes to identify causes and predict risk factors of brain and spinal infections.
Researchers from the Yonsei University Health System (YUHS) teamed up on a Hyundai Motor foundation-backed project to develop and test an AI model for diagnosing and prognosing central nervous system (CNS) diseases, such as encephalitis and meningitis.
FINDINGS
They trained their AI model on nearly 1,500 3D images of cerebrospinal fluid (CSF) immune cells collected from 14 patients who presented with CNS infections at Severance Hospital.
The team claimed to be the first to utilise 3D cell images and analyse cell morphology in the diagnosis and prognosis prediction of brain and spinal infections.
In a study evaluating the AI model, the researchers found that its predictive performance increased in accuracy as more cell images were run through it.
Initially, running one cell image showed 89% accuracy in identifying the cause of infection and 79% accuracy in predicting the disease’s likely course. Feeding five cell images, the accuracy of causal pathogen identification and prognosis prediction reached 99% and 94%, respectively. A 100% accuracy in both tasks was achieved with less than 10 cell images.
Additionally, the authors also identified cell mass, volume, and protein density as important factors for predicting prognosis.
The study findings have been published in Wiley’s Advanced Intelligent Systems journal.
WHY IT MATTERS
The diagnosis and prognosis of central nervous system diseases may vary depending on the cause. Bacterial or tuberculous causes have the highest mortality rate, which, according to the YUHS researchers, critically requires early, rapid diagnosis and treatment.
Meanwhile, there are different diagnostic tests for each causal pathogen, and certain tests may take weeks or more to produce reports. The CSF test, for instance, would require further manual confirmation of cell shape and count.
The Yonsei University researchers sought to cut this entire process – from CSF collection to evaluation – down to an hour. “The method we propose is faster than the current diagnostic techniques, such as brain imaging, and the frequently employed clinical biomarkers such as blood levels of procalcitonin and C-reactive protein,” they said.
They also emphasised the potential of cell morphology, analysed through 3D holotomography, as a “highly accurate and effective” biomarker for brain and spinal infections. This cost-effective imaging technique produces real-time, label-free 3D cell images.
ON THE RECORD
“This study is the first case of using three-dimensional images of immune cells in cerebrospinal fluid to predict the cause and prognosis of patients with central nervous system infections. We expect that the deep learning model presented in the study will be helpful in shortening the time required for patient diagnosis and prognosis prediction,” said Park Yu-rang, one of the researchers and an associate professor at the Department of Biomedical Systems Informatics of Yonsei University College of Medicine.
