A Lightweight, Low-Latency System Architecture for Deployable Depression Screening using Acoustic Biomarkers

Major Depressive Disorder (MDD) requires continuous, longitudinal monitoring, yet current diagnostic methods rely on intermittent clinical visits. Ambient Assisted Living (AAL) environments offer a solution by enabling passive screening via acoustic biomarkers. However, a critical research gap exists in the deployability of state-of-the-art detection models within these environments. Existing Machine Learning (ML) approaches often rely on heavy GPU acceleration, making them computationally prohibitive for widespread deployment on standard medical or home-server infrastructure. To address this, this paper proposes a lightweight, service-oriented system architecture designed for integration into privacy-centric on-premise environments. We present a low-latency inference microservice that prioritizes engineering constraints, specifically CPU efficiency and transparency, over marginal accuracy gains. Leveraging the eGeMAPSv02 feature set and an optimized Random Forest classifier (F1=0.696), our system achieves an end-to-end inference latency of approximately 7.3s on standard commodity hardware, with the core inference engine executing in less than 120ms. This efficient architecture allows the service to be deployed as a containerized local network service, enabling scalable, privacy-conscious depression screening without the reliance on specialized hardware accelerators or cloud connectivity.