The study team applied the Olink Target 48 Neurodegeneration plasma panel alongside disease-specific Elastic Net classifiers to test whether one shared proteomic assay can discriminate three major neurodegenerative diseases. The locked hold-out test produced strong discrimination: hold-out AUCs were 0.974 for Alzheimer disease (AD), 0.926 for Parkinson disease (PD), and 0.995 for amyotrophic lateral sclerosis (ALS). These AUCs were accompanied by sensitivity and specificity values computed at a Youden threshold that had been fixed in the training set and then applied unchanged to the test set. The study team reported 95% confidence intervals for all key metrics in the source document.
Investigators enrolled a single-center diagnostic cohort at Keio University Hospital, Tokyo, and after quality control plus one consent withdrawal analyzed 302 participants in total. The diagnostic breakdown: healthy controls n = 62; AD continuum n = 93; PD n = 39; ALS n = 27; other neurological disorders n = 81. The dataset was allocated by diagnosis-stratified sampling into a discovery training set (n = 211) and a locked hold-out test set (n = 91).
Data used in model building were plasma normalized protein expression values from assays that passed protein quality control. The study team analyzed 42 assays after this quality-control step. Elastic Net penalized regression was the disease-specific modelling approach stated in the source; the authors provided analysis scripts for the classifiers at the linked zenodo repository.
Hold-out AUCs in the locked test set were reported as follows: AD 0.974 (95% CI, 0.942–0.995); PD 0.926 (95% CI, 0.820–1.000); ALS 0.995 (95% CI, 0.980–1.000). The authors evaluated sensitivity and specificity at Youden cutoffs determined in the training set and then locked for the test set. At those thresholds the reported performance was: for AD, sensitivity 0.929 (95% CI, 0.765–0.991) and specificity 0.921 (95% CI, 0.824–0.974); for PD, sensitivity 0.833 (95% CI, 0.516–0.979) and specificity 0.975 (95% CI, 0.912–0.997); for ALS, sensitivity 1.000 (95% CI, 0.631–1.000) and specificity 0.940 (95% CI, 0.865–0.980). All numeric values above are reported directly in the source preprint.
The analytic panel included assays retained after protein quality control (42 assays). The authors excluded plasma DOPA decarboxylase from all classifiers a priori because, as they stated, blood levels primarily reflect dopaminergic treatment exposure rather than underlying disease biology. The source did not present additional assay-level methodology beyond noting normalized protein expression inputs and quality-control filtering; specific preanalytic or analytic instrument settings were not reported in the abstract and main text excerpt provided in the preprint summary.
The study followed institutional and national ethical standards. Written informed consent was obtained from participants or legally authorized representatives as required. The dementia-stream cohort and ALS enrollments were approved by the appropriate Keio University review boards and registered in public trial registries; approval identifiers and registry numbers are listed in the source. Deidentified participant-level data are available from the corresponding author upon reasonable request and subject to institutional and ethics review. Analysis scripts for the Elastic Net classifiers are publicly available at the DOI the authors provided.
Competing interests disclosed in the source include provision of statistical analysis software by Kravis Labs without charge (K.O.), various honoraria and travel support for one author, and employment of several authors by Eisai Co., Ltd. The source includes a full competing interest and funding list.
The authors concluded that a single plasma neurodegeneration panel supported disease-selective classification of AD, PD, and ALS in internal validation. They explicitly stated that multicenter external validation is required before clinical application. The preprint did not report multicenter data or external test cohorts; therefore, generalizability beyond this single-center sample remains an open question. The source did not provide additional details on prospective clinical utility testing, cost-effectiveness, or how the panel would integrate with existing diagnostic workflows. Those points and multicenter replication are the next specific steps documented by the authors as necessary before translation to routine clinical use.