Speech AI detects ALS with 0.73 F1 using graph neural networks

A new paper accepted at Interspeech2026 demonstrates how graph neural networks and self-supervised speech embeddings can detect Amyotrophic Lateral Sclerosis (ALS) and predict its progression from voice recordings. The researchers, Behrad TaghiBeyglou, Fatemeh Bagheri, and Ervin Sejdic, built a subject-level graph framework that aggregates multiple phonation recordings into a k-nearest-neighbor graph using pretrained SSL embeddings from 2-second segments. They tested four front-ends (wav2vec 2.0, HuBERT, data2vec-audio, and UniSpeech-SAT) paired with five GNN architectures (GCN, residual GCN, GAT, GraphSAGE, and GIN) on the SAND dataset containing 205 ALS patients and 134 controls.

The best configuration—HuBERT for feature extraction combined with a Graph Isomorphism Network (GIN)—achieved macro-F1 scores of 0.73 for the 5-class dysarthria severity task and 0.69 for the 4-class ALSFRS-R progression prediction. These results significantly outperform the official SAND validation baselines of 0.61 and 0.58, respectively. The work highlights the potential of combining graph-based learning with cross-lingual self-supervised speech representations for low-resource ALS detection and progression monitoring, enabling non-invasive, remote assessment of disease severity through voice analysis.