Summary
A single-centre preprint reports that disease-specific models using one plasma protein panel classified Alzheimer disease, Parkinson disease and ALS in a 302-person neurological cohort. The authors say multicentre external validation is needed before clinical application.
Researchers at Keio University Hospital in Tokyo report that one plasma protein panel, paired with separate statistical models for each disease, classified Alzheimer disease (AD), Parkinson disease (PD) and amyotrophic lateral sclerosis (ALS) in a 302-person neurological cohort. The study is a medRxiv preprint, and its findings come from an internal hold-out test set at a single centre; the authors say multicentre external validation is needed before clinical application.
How the panel was tested
The cohort included 62 healthy controls, 93 people in the AD continuum, 39 with PD, 27 with ALS and 81 with other neurological disorders. After quality control and one consent withdrawal, participants were divided by diagnosis into a training set of 211 and a locked test set of 91.
The researchers measured plasma proteins using the Olink Target 48 Neurodegeneration panel. Forty-two assays remained after protein quality control. Plasma DOPA decarboxylase was excluded from all classifiers in advance because, the researchers report, blood levels primarily reflect exposure to dopaminergic treatment.
For each disease, the team used an Elastic Net model. This statistical method can shrink or exclude predictors while fitting a model, helping it handle many candidate measurements. The researchers assessed each model on the held-out participants, using thresholds selected on the training set and then locked for the test.
Results in the held-out participants
The reported area under the receiver operating characteristic curve (AUC) was 0.974 for AD (95% confidence interval [CI], 0.942–0.995), 0.926 for PD (95% CI, 0.820–1.000) and 0.995 for ALS (95% CI, 0.980–1.000). AUC summarises how well a model separates the groups across possible thresholds; values closer to 1 indicate stronger discrimination in the tested data.
At the locked thresholds, AD sensitivity was 0.929 (95% CI, 0.765–0.991) and specificity was 0.921 (95% CI, 0.824–0.974). For PD, sensitivity was 0.833 (95% CI, 0.516–0.979) and specificity was 0.975 (95% CI, 0.912–0.997). For ALS, sensitivity was 1.000 (95% CI, 0.631–1.000) and specificity was 0.940 (95% CI, 0.865–0.980). Sensitivity measures the share of affected participants classified as positive; specificity measures the share of comparison participants classified as negative.
The results suggest that a shared panel can provide measurements for disease-specific classification models, rather than requiring a different panel for each condition. The estimates are from a test set drawn from the same single-centre cohort, however, and the confidence intervals—particularly for PD sensitivity and ALS sensitivity—are broad. The authors identify multicentre external validation as the next step before clinical use. The report is a preprint and has not completed peer review.