Artificial Intelligence-Enabled End-To-End Detection and Assessment of Alzheimer's Disease Using Voice

No accessible, non-invasive screening tool exists for identifying Alzheimer's disease in community settings. Current diagnostic pathways rely on costly, specialized medical procedures that remain unavailable to many patients. This limitation creates a significant barrier for early detection. Prior research has shown that speech patterns often change during cognitive decline. However, existing computational approaches frequently require complex manual feature extraction. That uncertainty drove the need for automated systems that process raw audio data directly. This study addresses the gap by utilizing self-supervised learning architectures. Such models offer a path toward more efficient, scalable diagnostic screening. No prior work had resolved the challenge of end-to-end severity prediction using only spontaneous speech samples.

The study aims to develop an artificial intelligence-powered system for detecting Alzheimer's disease using only voice recordings. Current diagnostic methods are often invasive, costly, and restricted to specialized clinical facilities. This lack of accessible screening tools creates a significant barrier for early patient identification. The researchers sought to create a solution that functions directly from raw audio data. They intended to predict both the presence of the disease and its cognitive severity. This effort addresses the need for a scalable, community-based diagnostic alternative. By leveraging self-supervised learning, the authors aimed to improve upon existing speech-based diagnostic approaches. The project explores the potential of vocal biomarkers as a reliable, non-invasive indicator of neurological health.

The researchers developed an end-to-end system powered by self-supervised learning to analyze vocal biomarkers. They utilized the pre-trained data2vec algorithm as the primary computational engine for this task. The team performed internal evaluation using the ADReSSo dataset, which contains speech from individuals describing a specific visual scene. External validation occurred using a separate test set sourced from DementiaBank. This approach avoids manual feature engineering by processing raw audio signals directly. The design focuses on predicting both disease presence and cognitive severity scores. Statistical validation included the Hosmer-Lemeshow test to assess the calibration of the model outputs. This methodology ensures that the system remains robust across different patient populations and recording conditions.

The model achieved an average area under the curve of 0.846 on held-out data. External validation on the DementiaBank dataset resulted in an area under the curve of 0.835. The system demonstrated strong calibration, evidenced by a Hosmer-Lemeshow goodness-of-fit p-value of 0.9616. These metrics confirm the high performance of the algorithm in distinguishing between healthy subjects and those with dementia. The system reliably predicted cognitive testing scores using only raw voice input. No significant performance degradation occurred when applying the model to external data sources. The findings indicate that the architecture effectively captures subtle vocal changes associated with cognitive impairment. This performance level supports the feasibility of using speech-based screening for early detection.

The authors demonstrate that their system effectively identifies Alzheimer's disease using only raw audio input. Their findings suggest that self-supervised algorithms provide a viable framework for cognitive health screening. The model achieves high diagnostic accuracy across both internal and external validation datasets. Statistical calibration confirms the reliability of the system for clinical interpretation. The researchers propose that this technology could facilitate widespread, low-cost monitoring in community environments. Their work highlights the potential for speech-based biomarkers to supplement traditional diagnostic assessments. The study confirms that cognitive scores can be predicted without requiring manual linguistic feature engineering. These results support the integration of advanced machine learning into routine neurological health evaluations.