Application of Mass Spectrometry-Based Metabolomics and Machine Learning in the Diagnostics of Lyme Neuroborreliosis

TL;DR

This study uses mass spectrometry and machine learning to diagnose Lyme neuroborreliosis with high accuracy, even in cases where traditional antibody tests are unclear.

Contribution

Combines mass spectrometry-based metabolomics with machine learning to achieve high diagnostic accuracy for LNB, independent of antibody status.

Findings

Machine learning models achieved perfect classification between treated LNB and non-LNB controls.

Metabolomic profiling showed strong discriminatory performance across all group comparisons.

Results suggest metabolomics could serve as a complementary diagnostic tool for LNB.

Abstract

Lyme borreliosis (LB) and its disseminated nervous system
manifestation,
Lyme neuroborreliosis (LNB), presents diagnostic challenges, especially
in seropositive and ambiguous clinical cases. In this study, we applied
mass spectrometry (MS)-based metabolomics combined with machine learning
(ML) to analyze serum samples from patients with definite acute LNB
(n = 34), treated LNB (n = 34), together with Borrelia antibody-negative (non-LNB) controls (n = 62). Importantly, pre-
and post-treatment samples were collected from the same individuals,
enabling within-patient comparisons that enhance sensitivity to LNB-related
metabolic changes. The non-LNB control group was age- and sex-matched
(n = 34), and treated LNB patients served as a practical substitute
for postinfectious recovery. Strong discriminatory performance was
observed across all pairwise group comparisons. ML model classifiers
yielded accuracy rates significantly above those expected by chance,
with a perfect classification (1.00) achieved between treated LNB
and non-LNB controls. This high separation, independent of antibody
status, highlights the potential of MS-based metabolomics as a complementary
diagnostic strategy. Receiver operating characteristic curve (ROC)
analyses further supported robust performance, with high sensitivity
and specificity. Although variance explained in unsupervised ordination
was limited (PERMANOVA 4%), the supervised models demonstrated diagnostic
value. These findings support the feasibility of metabolomic profiling
combined with ML models for LNB diagnosis.

Figures