Mixed-Precision Information Bottlenecks for On-Device Trait-State Disentanglement in Bipolar Agitation Detection
Abstract
Continuous monitoring of bipolar disorder agitation via voice biomarkers requires disentangling stable speaker traits from volatile affective states on resource-constrained edge devices. We introduce MP-IB, the first framework to treat mixed-precision quantization as an information bottleneck for clinical trait-state separation. The core insight is that numerical precision itself controls capacity: an FP16 trait head (1,024 bits) encodes speaker identity, while an INT4 state head (128 bits) captures agitation, yielding 8x information asymmetry without adversarial training. We augment this with Dynamic Precision Scheduling and Multi-Scale Temporal Fusion. On Bridge2AI-Voice (N=833, 4 sessions/participant, strict speaker-independent CV), MP-IB achieves rho = 0.117 (95\% CI: [0.089, 0.145], p=0.003 vs. chance), outperforming 94M-parameter WavLM-Adapter with in-domain SSL continuation (rho = -0.042), beta VAE disentanglement (rho = 0.089), and hand-crafted prosody (rho = 0.031) by 2.8--15.9 points absolute. Zero-shot transfer to CREMA-D achieves AUC=0.817. Identity leakage is suppressed to near-random (EER=0.42, MIA-AUC=0.52). End-to-end latency is 23.4 ms with a 617 KB footprint, enabling real-time monitoring on sub 20 dollar devices.
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
The paper presents a pioneering approach to using mixed-precision quantization for trait-state disentanglement in bipolar agitation detection, significantly advancing the field of clinical speech analysis and edge deployment. The combination of theoretical insights and practical applications positions this work as a valuable contribution to machine learning and mental health monitoring.
Methodology
The paper introduces a novel framework, MP-IB, that leverages mixed-precision quantization as an information bottleneck for disentangling speaker traits from emotional states in voice data. The methodology is grounded in information theory, specifically utilizing numerical precision to control information capacity, which is a unique approach in the context of clinical speech applications. The integration of Dynamic Precision Scheduling and Multi-Scale Temporal Fusion enhances the model's performance while maintaining low latency and memory requirements, making it suitable for edge devices.
Results & evaluation
The experiments are robust, utilizing a well-defined dataset (Bridge2AI-Voice) with a significant number of participants (N=833) and employing rigorous statistical methods to validate results. The framework outperforms existing methods, including adversarial training and traditional disentanglement approaches, across multiple metrics, indicating strong empirical performance. The zero-shot transfer capability to CREMA-D further demonstrates the model's generalizability.
Reproducibility
The paper provides detailed implementation information, including training protocols, model architecture, and evaluation metrics, which enhances reproducibility. However, the absence of a publicly accessible code repository limits the ability for external validation.
Limitations
The study's limitations include potential overfitting due to the relatively small labeled dataset (166 participants) and the lack of formal differential privacy guarantees despite empirical privacy analysis. Additionally, the method's performance may vary across different populations and settings, which could affect its generalizability.
Broader impact
The proposed framework has significant implications for real-time mental health monitoring, particularly in low-resource settings. By enabling continuous monitoring of bipolar disorder agitation through voice analysis on inexpensive edge devices, it could improve patient care and provide timely interventions. The privacy-preserving aspects of the model also address critical concerns in sensitive health data applications. The paper presents a pioneering approach to using mixed-precision quantization for trait-state disentanglement in bipolar agitation detection, significantly advancing the field of clinical speech analysis and edge deployment. The combination of theoretical insights and practical applications positions this work as a valuable contribution to machine learning and mental health monitoring.
Author of this paper? Send a correction.