# Targeted CSF metabolomics and conformal prediction improve diagnostic accuracy of normal pressure hydrocephalus

## TL;DR

This study shows that analyzing cerebrospinal fluid metabolites and using a machine learning method called conformal prediction can help diagnose normal pressure hydrocephalus more accurately.

## Contribution

The study introduces a novel combination of targeted CSF metabolomics and conformal prediction to improve iNPH diagnosis.

## Key findings

- Eight metabolites were consistently reduced in iNPH patients, independent of other factors.
- An integrated model achieved high diagnostic accuracy (AUC = 0.97) for distinguishing iNPH from other conditions.
- Conformal prediction provided reliable confidence estimates for individual diagnoses.

## Abstract

Idiopathic normal pressure hydrocephalus (iNPH) is a progressive but treatable neurological disorder. Yet, diagnosis is often confounded by overlapping symptoms and biomarker profiles with Alzheimer’s disease (AD), mild cognitive impairment (MCI), and frontotemporal dementia (FTD). We aimed to determine whether cerebrospinal fluid (CSF) metabolomic profiling, combined with uncertainty-aware machine learning using conformal prediction (CP), could improve diagnostic differentiation of iNPH.

CSF samples were collected from 120 patients with iNPH, 44 healthy controls, and 152 individuals with AD, MCI, or FTD. Targeted metabolomics of 59 metabolites was performed using liquid chromatography–high-resolution mass spectrometry. Group differences were assessed using age- and sex-adjusted regression models. Multivariate classification with partial least squares discriminant analysis (PLS-DA) incorporated metabolites, demographics, and conventional biomarkers (amyloid-β42, tau, phosphorylated tau). CP was applied to address individual-level diagnostic uncertainty.

Eight metabolites (proline, threonine, histidine, tyrosine, tryptophan, isobutyrylcarnitine, citric acid, and dehydroascorbic acid) were consistently reduced in iNPH (q < 0.05), independent of ventricular volume and cortical tau or amyloid-β pathology. An integrated PLS-DA model combining metabolomic, demographic, and AD-biomarker data achieved excellent discrimination (AUC = 0.97). CP provided calibrated case-level confidence, identifying clear-cut and uncertain cases while maintaining high accuracy (94% for iNPH, 97% for not-iNPH).

iNPH exhibits a distinct CSF metabolomic signature reflecting altered amino acid metabolism, mitochondrial function, and oxidative stress. Integrating metabolomic data with established biomarkers enhances diagnostic accuracy, while CP adds individualized uncertainty estimates to improve diagnostic confidence and guide treatment decisions.

The online version contains supplementary material available at 10.1186/s12987-026-00771-z.