Deep-learning analysis of speech using mel-spectrograms for the assessment of mild cognitive impairment and Alzheimer's disease

Background: Speech abnormalities are recognized as early indicators of Alzheimer's disease (AD) and mild cognitive impairment (MCI). Objective: To determine whether deep-learning models trained on mel-spectrograms of brief speech tasks can (i) discriminate individuals with MCI and AD from cognitively normal controls (NC) and (ii) estimate cognitive status with clinically useful accuracy. Methods: Speech from 594 participants (185 NC, 231 MCI, 178 AD) was recorded through a mobile application that included 11 cognitive-linguistic tasks. Audio was converted into mel-spectrogram images and processed using a VGG16-based deep-learning model with transfer learning and fine-tuning of block 5. Task-specific feature vectors were extracted, concatenated, and used to train a deep neural network. The dataset was split into training, validation, and test sets (3:1:1), and five-split cross-validation was performed. Results: The model demonstrated an overall accuracy of 72.4% in classifying NC from the abnormal group (MCI and AD), with sensitivity and specificity of 72.5% and 72.2%, respectively, a balanced accuracy of 72.4%, and an AUC of 0.997. In binary classifications, the model achieved 82.9% accuracy (balanced accuracy 82.9%, AUC 0.992) for NC versus AD, 70.7% accuracy (balanced accuracy 70.3%, AUC 0.956) for NC versus MCI, and 77.5% accuracy (balanced accuracy 78.9%, AUC 0.889) for MCI versus AD. Tasks such as serial subtraction, storytelling, and picture description contributed most to classification performance, indicating their effectiveness in capturing cognitive deficits. Conclusions: Mel-spectrogram-based deep-learning analysis of speech shows promise as a rapid, non-invasive, and language-independent screening tool for early cognitive impairment, with potential advantages over traditional assessments such as the Mini-Mental State Examination.