A deep learning model developed by researchers at Yonsei University has shown near-perfect performance in identifying the cause and predicting the prognosis of central nervous system (CNS) infections, using just a few images of immune cells extracted from cerebrospinal fluid (CSF).

The AI was trained on 3D holotomography images that quantify structural and biochemical features of live cells without requiring stains or labels. It reached 99 percent accuracy in identifying infection type, distinguishing among viral, bacterial, and tuberculosis cases. Prognosis prediction hit 94 percent accuracy. 

The researchers report that these results can be generated within an hour of CSF collection.

The study, published March 26 in the journal Advanced Intelligent Systems, was described by corresponding author Professor Park Yu-rang of the Department of Biomedical Systems Informatics at Yonsei University College of Medicine in a release issued last Thursday as the first to use 3D CSF immune cell morphology instead of protein or genetic markers for both diagnosis and outcome prediction in CNS infections. 

She added that the tool could “help shorten the time needed for diagnosis and treatment planning in patients with CNS inflammation.”

The prospective study enrolled 14 adults with confirmed CNS infections at Severance Hospital between January and October 2022. A total of 1,427 immune cell images were captured using holotomography, a label-free technique that measures the refractive index (RI) of live cells to reveal biophysical structure.

Patients were categorized by infection type and clinical outcome using the modified Rankin Scale (mRS) score at discharge. Of the 14, three had poor prognoses (mRS ≥4), and five had bacterial or tuberculosis infections.

The AI model, based on a modified DenseNet-169 architecture, was benchmarked against the commonly used ResNet-101. It scored an area under the ROC curve (AUROC) of 0.89 in distinguishing viral from non-viral infections, outperforming ResNet’s 0.82. For prognosis prediction, the model achieved an AUROC of 0.79, which improved to 0.94 when five cells per patient were analyzed.

When five immune cell images per patient were used, performance improved. The AUROC rose to 0.99 for identifying infection type and 0.94 for predicting clinical outcome, with improved consistency and reduced variability across samples.

Cell morphology appeared to carry strong predictive weight. Immune cells from viral cases had larger nuclei and higher protein density. Those from patients with poor outcomes showed greater dry mass but lower protein density, a pattern also seen in non-viral infections. These features were extracted directly from holotomography-based RI measurements.

To better understand what the model was seeing, the researchers used gradient-weighted class activation mapping (Grad-CAM) to highlight cell regions influencing predictions. Variations in refractive index near the nucleus were key. In viral infections, the effective region was confined to the shell of the inner cell. In poor-prognosis cases, nuclear components expanded laterally and density decreased in the outer regions.

While previous AI models in infectious disease have relied on clinical data or molecular tests, many require electronic health records or lab-based assays that delay results. The Yonsei team’s approach uses cell shape and structure, allowing for rapid, label-free analysis with minimal lab infrastructure.