Machine learning classification of glioma molecular markers from intraoperative fluorescence spectra

Source article: Intraoperative fluorescence spectroscopy for IDH classification in glioma: a feasibility study

Background Molecular features are fundamental to both the diagnosis and prognosis of gliomas. Specific molecular features are generally unavailable at the time of index surgery; however, upfront knowledge of molecular characteristics could meaningfully inform surgical strategy. Aims/objectives To determine the feasibility of a machine learning model trained on fluorescence spectral features to classify IDH mutation and MGMT promoter methylation status in glioma tissue. Methods Tissue samples were collected durin…

By the Truvace deskOctober 17, 20262 min read1 sourcePeer-reviewedTRV-2026-1678· cite & verify

Intraoperative fluorescence spectroscopy for IDH classification in glioma: a feasibility study — "Brain Microbleed" by National Institutes of Health (NIH) is marked with Public Domain Mark 1.0. To view the terms, visit https://creativecommons.org/publicdomain/mark/1.0/.

Trace impact readingPositive state

P 64

Both sides are scored from claims and sources, not community votes.

G 72

In brief

Researchers recorded fluorescence spectra from glioma tissue taken during 5-ALA-guided surgery and used the spectra to train a machine learning algorithm. The model was tested on a held-out half of the patients and distinguished IDH-mutant samples with 87% accuracy, but it did not reliably separate MGMT methylation status.

Molecular features such as IDH status normally become known only after the initial operation, so a real-time prediction could inform surgical strategy. The study is a proof of concept with a small cohort, so it remains unclear how well the method would perform in larger and more varied groups of patients.

Main points

Twenty-seven patients were recruited and more than 8000 spectra were measured from tumour tissue.
Samples were collected during 5-ALA-guided surgery and probed with a fibre probe to record fluorescence spectra.
The fluorescence data were blinded and randomised and trimmed to relevant wavelengths before analysis.
The model was trained on half of the patients for each label and tested on the other half.
The authors call for further work integrating IDH and 1p/19q status in larger cohorts.

The gain

In a 27-patient pilot, a machine learning model trained on 5-ALA fluorescence spectra identified IDH-mutant glioma samples with 87% accuracy, suggesting possible real-time intraoperative molecular guidance.

The problem

The same model could not reliably classify MGMT promoter methylation status, so the approach did not extend to this second molecular marker.

The rundown

The study addresses a gap in neurosurgery: molecular characteristics are fundamental to glioma diagnosis and prognosis, yet they are generally unavailable at the time of the index surgery. Upfront knowledge of them could meaningfully inform surgical strategy.

Tissue specimens were each linked to recorded IDH and MGMT methylation data, which were incorporated into the machine learning algorithm. Training used fluorescence data from half of the patients with each label, and testing used the other half.

The results were mixed. IDH-mutant tissue was identified with 87% accuracy, while MGMT methylation could not be reliably classified. The authors conclude the approach shows promise for real-time intraoperative IDH prediction and describe the work as a feasibility study.

What this doesn’t fix

This is a proof-of-concept study in only 27 patients, and the authors say larger cohorts and integration of 1p/19q status are still needed; MGMT methylation could not be reliably classified.