A deep learning model that reads roughly 7,000 plasma proteins from a single blood draw and returns simultaneous probabilities for six conditions of ageing was published in Nature Medicine on 31 March 2026 by Lijun An, Jacob Vogel and colleagues at Lund University with the Global Neurodegeneration Proteomics Consortium (GNPC). The model, Proteomics-based Artificial Intelligence for Dementia Diagnosis (ProtAIDe-Dx), was developed on 17,187 patients and controls (mean age 70.3 years, 53.2 per cent female) drawn from 19 contributing sites in the GNPC dataset. The six categories are Alzheimer's disease, Parkinson's disease, frontotemporal dementia, amyotrophic lateral sclerosis, previous stroke or transient ischaemic attack, and cognitive unimpairment. Across all six, the paper reports cross-validated balanced classification accuracy of 70 to 95 per cent and area under the curve above 78 per cent. This is retrospective modelling on existing cohort samples, not a prospective clinical trial.
The stated reason for the architecture is co-pathology. The authors write that sensitive, specific and scalable biomarkers for living patients are unavailable for most neurodegenerative pathologies, and that one-disease-at-a-time testing sits badly with brains that carry more than one process. Rather than forcing a single label, the model outputs a probability per condition, and the authors report that those probabilities flagged subgroups with more than one pathology. In an external memory clinic sample, the BioFINDER-2 cohort (N = 1,786, a site held out of model fitting), Parkinson's probability was raised not only in Parkinson's patients but also in people diagnosed with dementia with Lewy bodies, and stroke or transient ischaemic attack probability was raised in patients with vascular dementia. Vascular dementia itself was not a label: because the GNPC held no vascular dementia diagnoses, the preprint states that the stroke and TIA group was used to represent documented cerebrovascular disease.
The authors are direct about what the model cannot yet do. Performance dropped in leave-one-site-out and site-to-site testing, the settings that most resemble real clinical translation, and worst for conditions concentrated in a single site: nearly all the ALS cases came from one contributor. The paper states that predictive proteomics as a field still faces insufficient diagnostic accuracy for clinical deployment, limited generalisability across sites and potential confounding. The reference labels are a further boundary. The preprint version of the study says plainly that ProtAIDe-Dx was trained exclusively on clinical diagnosis, and that most of those diagnoses were made without biomarker support; none were confirmed at autopsy. The version of the preprint posted in October 2025 lists 6,693 recruited controls and 5,772 patients, of whom 2,249 had Alzheimer's disease, 2,287 Parkinson's disease, 435 ALS, 317 stroke or TIA and 175 frontotemporal dementia, with 309 carrying two or more diagnoses.
What is new here is the design argument rather than a test families can ask for. No regulator has reviewed ProtAIDe-Dx, it has not been studied in primary care, and the paper reports no evidence that acting on its probabilities changes what happens to a patient. Code for the model has been posted publicly by the Lund group. The authors position the work as a benchmark for future multi-disease plasma tools, and note that data harmonisation and standardisation across sites will probably be needed before such a model travels.