An AD detection system using topic enhancement and dynamic differential privacy for intelligent elderly care

To address the challenges of low detection accuracy and high privacy leakage risk in Alzheimer’s disease (AD) detection for intelligent elderly care systems, this paper proposes ADDETECTOR, an AD detection system based on topic enhancement and dynamic differential privacy. The system adopts a three-layer architecture comprising a user layer, client layer, and cloud layer, enabling full-process AD detection from speech collection to model aggregation. A language feature enhancement method based on topic modeling is introduced, which constructs an AD-related topic lexicon and designs a semantic deviation measurement model to effectively capture subtle linguistic anomalies in AD patients. Furthermore, a dynamic differential privacy mechanism combined with a federated security framework is developed, employing Laplace noise injection, elliptic curve Diffie–Hellman (ECDH) key exchange, and secret sharing to ensure secure parameter transmission and gradient aggregation while protecting user voice data privacy. Experimental results on the ADReSS dataset demonstrate that ADDETECTOR achieves significant improvements in detection accuracy and F1 score compared to baseline models such as HAN-AGE and BiGRU, while maintaining low system response time. Parameter sensitivity tests show good scalability across 2–9 clients and a favorable privacy-utility balance when the privacy budget ε is set between 0.8 and 5. The results demonstrate the feasibility of ADDETECTOR for privacy-aware speech-based AD screening under the controlled ADReSS experimental setting; further validation on independent datasets and real-world elderly-care environments is required before practical deployment