Longitudinal Speech Biomarkers for Automated Alzheimer's Detection

• DOI: 10.3389/fcomp.2021.624694
• Publication Date: 2021-04-08T13:11:44Z
• Authors: Jordi Laguarta; Brian Subirana
• Keywords: FOS: Computer and information sciences; Sound (cs.SD); graph neural-networks; transfer learning; Quantitative Biology - Quantitative Methods; Computer Science - Sound; 03 medical and health sciences; 0302 clinical medicine; Audio and Speech Processing (eess.AS); FOS: Electrical engineering, electronic engineering, information engineering; Quantitative Methods (q-bio.QM); multimodal deep learning; QA75.5-76.95; brain model; I.2.m; I.2.0; AI diagnostics; 3. Good health; explainable speech recognition; Electronic computers. Computer science; FOS: Biological sciences; I.2.0; I.2.m; Electrical Engineering and Systems Science - Audio and Speech Processing

We introduce a novel audio processing architecture, the Open Voice Brain Model (OVBM), improving detection accuracy for Alzheimer's (AD) longitudinal discrimination from spontaneous speech. also outline OVBM design methodology leading us to such which in general can incorporate multimodal biomarkers and target simultaneously several diseases other AI tasks. Key our is use of multiple complementing each other, when two them uniquely identify different subjects disease we say they are orthogonal. illustrate by introducing 16 biomarkers, three orthogonal, demonstrating simultaneous above state-of-the-art apparently unrelated as AD COVID-19. Inspired research conducted at MIT Center Minds Machines, combines biomarker implementations four modules intelligence: The brain OS chunks overlaps samples aggregates features sensory stream cognitive core creating multi-modal graph neural network symbolic compositional models task. apply it AD, achieving 93.8% on raw audio, while extracting subject saliency map that longitudinally tracks relative progression using reported ultimate aim help medical practice detecting onset treatment impact so intervention options be tested. Using OBVM methodology, lung respiratory tract created 200,000+ cough pre-train model discriminating cultural origin. This dataset sets new benchmark largest health with 30,000+ participating April 2020, first-time bias.