AI-based staging, causal hypothesis and progression of subjects at risk of Alzheimer's disease: a multicenter study

Simona Aresta^1,2Raffaello Nemni³Moreno Zanardo⁴Graziella Sirabian³Dario Capelli³Marco Alì³Paolo Vitali^4,5Enrico Giuseppe Bertoldo⁶Valentina Fiolo⁶Lilla Bonanno⁷Giuseppa Maresca⁷Petronilla Battista²Francesco Sardanelli^4,8Francesca Benedetta Pizzini⁹Isabella Castiglioni^10*Christian Salvatore^1,11

• ¹Department of Science, Technology and Society, University School for Advanced Studies IUSS Pavia, 27100 Pavia, Italy
• ²Istituti Clinici Scientifici Maugeri IRCCS, Laboratory of Neuropsychology, Institute of Bari, 70124 Bari, Italy
• ³Centro Diagnostico Italiano, Milan, Lombardy, Italy
• ⁴Unit of Radiology, IRCCS Policlinico San Donato, San Donato Milanese, Italy
• ⁵Department of Biomedical Sciences for Health, Università degli Studi di Milano, Milan, Italy
• ⁶Clinical Psychology Service, IRCCS Policlinico San Donato, San Donato Milanese, Italy
• ⁷IRCCS Centro Neurolesi Bonino Pulejo, Messina, Italy
• ⁸Lega Italiana per la Lotta contro i Tumori (LILT), Milano Monza Brianza, Milan, Italy
• ⁹Department of Engineering for Innovation Medicine, University of Verona, Verona, Veneto, Italy
• ¹⁰Department of Physics “G. Occhialini”, Università degli Studi di Milano-Bicocca, Milan, Italy
• ¹¹DeepTrace Technologies SRL, Milan, Lombardy, Italy

In 2024, eleven European scientific societies/organizations and one patient advocacy association have defined a patient-centered biomarker-based diagnostic workflow for memory clinics evaluating neurocognitive disorders. We tested the performance of an artificial intelligence (AI) tool applied to neuropsychological and magnetic resonance imaging (MRI) assessment for staging and causal hypothesis, which are the two recommended workflow steps guiding the next one recommending optimal biomarkers to be used for a biological diagnosis of neurocognitive disorders, according to intersocietal recommendations. Moreover, we assessed the AI performance in predicting the progression to Alzheimer's disease (AD)-dementia. For the three-class classification of staging (n patients = 426), the inter-rater AI-humans agreement was substantial for both healthy subjects/subjective cognitive impairment/worried-well vs. all the remaining groups (rest) (Cohen's κ = 0.81) and mild cognitive impairment/mild dementia vs. rest (κ = 0.70) classification, almost perfect for moderate/severe dementia vs. rest (κ = 0.90) classification. For the three-class classification of causal hypotheses (n = 112), the AI performance vs. biomarker-based diagnosis was: positive predictive value 91% [95% CI: 84-96%]; negative predictive value 100%, and accuracy 91% [84-96%]. For the binary classification of progression or not progression to AD-dementia at 24-month, with clinical conversion as a reference standard (n = 341), the AI performance was: sensitivity 89% [84-94%], specificity 82% [77-87%]; accuracy 85% [81-89%]; and area under the receiver operating characteristic curve 83% [79-87%]. The AI tool showed high agreement with human assessment for staging, high accuracy with biomarkers for causal hypotheses of neurocognitive disorders and predicted progression to AD at 24-month with 89% sensitivity and 82% sensitivity.