ProMUSE: Progressive Multi-modal Uncertainty-guided Staged Evidential Alzheimer Disease Classification

A new machine-learning framework called ProMUSE could cut reliance on costly brain scans for early Alzheimer’s diagnosis by 50 to 90 percent, according to research submitted on 11 June 2026, while maintaining accuracy on par with full-modality systems [1]. Alzheimer’s disease progressively destroys memory and cognitive skills in older adults, and most available treatments are effective only in its early stages, driving demand for earlier detection [1]. Clinicians increasingly turn to multimodal data — clinical assessments, structural MRI, and PET imaging — to make a diagnosis, but MRI and PET scans remain expensive and not universally accessible, limiting their use in routine screening [1]. The proposed system, called ProMUSE, addresses that barrier by starting with low-cost clinical data and only ordering imaging when the model’s uncertainty exceeds a learned threshold [1]. It first performs evidential classification on clinical inputs and quantifies uncertainty through a Dirichlet-based subjective logic model. If uncertainty is high, the network progressively incorporates MRI or PET features, fusing modality-wise belief and uncertainty via Dempster-Shafer theory to produce a calibrated multimodal prediction [1]. Across three benchmark datasets — ADNI, AIBL, and OASIS — ProMUSE matched or exceeded the accuracy of baselines that always used all modalities, while reducing MRI and PET usage by 50 to 90 percent [1]. The researchers tested the model on three diagnostic tasks: distinguishing cognitively normal individuals from Alzheimer’s patients, cognitively normal from mild cognitive impairment, and mild cognitive impairment from Alzheimer’s disease [1]. The cost implications are significant. Structural MRI and PET scans can run to thousands of dollars per session and require specialized equipment and personnel, factors that have slowed the deployment of AI-assisted screening in community clinics and low-resource settings. By adaptively deciding when imaging is necessary, ProMUSE offers a staged acquisition strategy that preserves diagnostic performance while yielding what the authors describe as “substantial cost savings” [1]. The work arrives as health systems globally are under pressure to expand dementia screening. The World Health Organization has identified dementia as a public-health priority, and the number of people living with Alzheimer’s is projected to triple by 2050. Tools that lower per-patient costs without sacrificing accuracy could make population-level early detection more feasible, though the model has not yet been validated in prospective clinical trials [1].