Development and validation of AI‐based tools for brain amyloid‐β detection using MRI. (20th December 2022)

• Record Type: Journal Article
• Authors: Devanarayan, Viswanath; Charil, Arnaud; Nelson, Todd; Qi, Xin; Murali, Leema; Reyderman, Larisa; Irizarry, Michael C; Koyama, Akihiko; Dhadda, Shobha
• Is Part Of: Alzheimer's & dementia. Volume 18(2022)Supplement 5
• Page Start: n/a
• Page End: n/a
• Publication Date: 2022-12-20
• Subjects: Alzheimer's disease -- Periodicals; Alzheimer Disease -- Periodicals; Dementia -- Periodicals; Démence; Maladie d'Alzheimer; Périodique électronique (Descripteur de forme); Ressource Internet (Descripteur de forme); 616.83
• DOI: 10.1002/alz.068231
• Languages: English
• ISSNs: 1552-5260

Abstract

Abstract: Background: Non‐invasive methods for brain amyloid‐b (Ab) detection are needed for effective patient care, and for faster and more cost‐effective screening of patients in clinical trials. Using MRI and other patient screening data from internal clinical trials, we employ artificial intelligence (AI) algorithms to develop brain Aβ positivity detection tools and validate their performance in independent datasets. Method: Training dataset included structural MRI and several cognitive assessments obtained during screening of 1960 patients from a clinical trial. Volumetric measures from 207 brain regions were generated by BioClinica using their proprietary imaging pipeline. Over 85% of these patients had mild cognitive impairment (MCI), and approximately 50% were Aβ positive based PET visual read (mostly using Florbetaben). Signatures for detecting brain Aβ positivity were derived via several AI algorithms such as Bayesian elastic‐net, neural network, regularized random forests, and stochastic gradient boosting. These models included demographics and clinical features, and the added value of ApoE and MRI features were assessed. Performance accuracy was first assessed via 10‐fold cross‐validation within the training set. Signatures from the best performing algorithm were evaluated in two independent test sets: a clinical trial cohort (n=1853) and an ADNI cohort (n=1092). Result: Signatures using cognitive assessments and demographics achieved 62.2% accuracy for detecting brain Aβ positivity when evaluated via cross‐validation, which improved to 75.2% when adding ApoE, and to 77.9% when including MRI. When tested in two independent cohorts, signatures with cognitive assessments and demographics achieved accuracy of 64.5% and 65.8% respectively, which improved to 71.9% and 74.3% when adding ApoE, to 76.7% and 77.5% when including MRI, and to 83.6% and 84.8% when considering 2/3 of the test‐sets that were predicted with high‐confidence (i.e., subjects predicted with > 70% probability as brain Aβ positive or negative). Conclusion: These results demonstrate the potential of non‐invasive AI tools based on cognitive assessments, ApoE and MRI for detecting brain Aβ positivity in clinical trials and patient care. Results reported here may be refined further using impending additional data prior to final presentation.