Towards Uncertainty-Aware Machine Learning for Cerebrospinal Fluid–Based Diagnostic Decision Support

In collaboration with clinicians in University Hospital Zurich, this project develops an end-to-end machine learning framework for diagnostic decision support in disorders affecting the central nervous system (CNS). Clinical interpretation of cerebrospinal fluid (CSF) biomarkers is central to neurological diagnostics but often challenging due to overlapping biomarker patterns across infectious, autoimmune, inflammatory, and malignant conditions.

The project investigates how machine learning models trained on routinely collected CSF biomarker panels alone can support etiologic differentiation across major CNS disease categories. In addition, it explores interpretability frameworks to better understand clinically meaningful relationships between CSF biomarker patterns and neurological health outcomes, enabling more transparent and clinically actionable AI-supported diagnostics.

Beyond predictive performance, the research focuses on improving the reliability of clinical AI by incorporating uncertainty-aware methods that quantify prediction stability under physiologically plausible variations in CSF profiles.

The developed algorithms are translated into a clinician-facing inference tool capable of automated CSF report ingestion, feature standardization, interpretable predictions, and uncertainty communication. By combining routine biomarker modeling with robust uncertainty quantification, the project aims to advance reliable and deployable AI-based decision support for neurological diagnostics.