15-Protein Blood Test Uses AI to Diagnose Four Major Brain Diseases at Once
A new AI classifier using just 15 proteins achieves 95.5% AUC accuracy across Alzheimer's, Parkinson's, FTD, and Lewy body dementia.
Summary
Diagnosing neurodegenerative diseases is notoriously difficult, especially when patients have overlapping conditions. Researchers at Washington University developed GPND-AI, an artificial intelligence classifier that analyzes 15 proteins measured via a highly sensitive blood-based assay called NULISA. The system can distinguish between Alzheimer's disease, Parkinson's disease, frontotemporal dementia, dementia with Lewy bodies, and healthy controls — all from a single panel. Trained and tested on clinical cohorts and validated against neuropathologically confirmed diagnoses, the classifier achieved an AUC of 0.955 and 92.3% accuracy across five categories. Crucially, it can also detect mixed pathologies, a common real-world complication. This represents a significant step toward accessible, blood-based diagnostic tools for neurodegenerative disease.
Detailed Summary
Distinguishing between neurodegenerative diseases in living patients is one of neurology's hardest problems. Alzheimer's, Parkinson's, frontotemporal dementia, and dementia with Lewy bodies share overlapping symptoms, and many patients carry mixed pathologies that confound clinical diagnosis. Definitive confirmation has historically required post-mortem neuropathology, leaving clinicians and patients without certainty during life.
Researchers from Washington University School of Medicine developed GPND-AI, a generalizable protein-based AI classifier built on the NULISA (NUcleic acid-Linked Immuno-Sandwich Assay) central nervous system panel. The system was trained using proteomic and clinical data from the Knight Alzheimer's Disease Research Center and a Movement Disorder Clinic, then externally validated in a Banner Sun Health Research Institute cohort with neuropathologically confirmed diagnoses.
The classifier identified a 15-protein signature that achieves an area under the curve of 0.955 and 92.3% accuracy across five diagnostic categories — four diseases plus healthy controls. Notably, the model also predicted co-pathologies, and these predictions significantly correlated with clinical characteristics in the validation cohort, suggesting the tool can map overlapping disease processes rather than forcing a single diagnosis.
The implications are substantial. A blood-based, proteomics-driven diagnostic tool could transform how neurodegenerative diseases are identified in clinical practice — enabling earlier intervention, more precise trial enrollment, and better patient stratification. For clinicians, it offers a potential adjunct to imaging and CSF biomarkers that is less invasive and more scalable.
Caveats remain. This summary is based on the abstract only, so full methodological details, cohort demographics, and protein identities are unavailable. The classifier requires prospective validation in diverse, real-world populations before clinical deployment. Mixed-pathology detection, while promising, needs further benchmarking against comprehensive neuropathological ground truth.
Key Findings
- A 15-protein NULISA panel achieved 95.5% AUC and 92.3% accuracy across four neurodegenerative diseases plus healthy controls.
- The AI classifier can detect co-existing pathologies, not just a single primary diagnosis.
- External validation against neuropathologically confirmed cases supports real-world diagnostic utility.
- The approach is blood-based, making it less invasive than CSF sampling or PET imaging.
- Model was trained and validated across multiple independent clinical cohorts.
Methodology
GPND-AI was trained on proteomic and clinical data from the Knight ADRC and Movement Disorder Clinic at Washington University. External validation used neuropathologically confirmed samples from the Banner Sun Health Research Institute cohort plus additional Knight-ADRC cases. The NULISA CNS panel provided high-sensitivity protein quantification across the diagnostic groups.
Study Limitations
This summary is based on the abstract only, as the full paper is not open access; detailed methodology, protein identities, and cohort characteristics are unavailable. The study requires prospective validation in broader, more diverse populations before clinical adoption. Performance in early or preclinical disease stages has not been reported here.
Enjoyed this summary?
Get the latest longevity research delivered to your inbox every week.