🤖 machine learning

Mixed-Precision Information Bottlenecks for On-Device Trait-State Disentanglement in Bipolar Agitation Detection

The paper introduces MP-IB, a mixed-precision information bottleneck framework that leverages numerical precision asymmetry to effectively disentangle stable speaker traits from volatile agitation states on resource-constrained edge devices, achieving superior clinical performance and privacy protection compared to existing deep learning and hand-crafted methods.

Original authors: Joydeep Chandra

Published 2026-05-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Joydeep Chandra

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to listen to a radio station that is playing two different things at once: a permanent station ID (the voice of the person speaking) and a temporary weather report (their current mood, like agitation or calmness).

In the world of mental health monitoring, doctors want to hear the "weather report" (is the patient agitated?) without getting distracted by the "station ID" (who is speaking?). Usually, to do this, computers need massive, heavy brains (huge AI models) that run on powerful servers in the cloud. This is slow, expensive, and risky for privacy because the audio has to travel over the internet.

This paper introduces a clever new trick called MP-IB. Instead of building a bigger brain, the authors built a smart filter that runs on a tiny, cheap device (like a $20 Raspberry Pi) right next to the patient.

Here is how it works, using simple analogies:

1. The "Two-Headed" Strategy

Think of the AI model as having two different jobs, handled by two different "heads":

The Identity Head (The VIP): This head needs to remember who is speaking. It is given a high-precision memory (like a high-definition photo). It uses FP16 (16-bit) math, which is detailed and clear.
The Mood Head (The Reporter): This head only needs to know how the person is feeling right now. It is given a low-precision memory (like a rough sketch). It uses INT4 (4-bit) math, which is very coarse and blurry.

The Magic: By forcing the "Mood Head" to use such a tiny, low-resolution memory, the AI is physically unable to store the details of the person's voice. It's like trying to draw a detailed portrait of a person using only 4 crayons; you can capture the general shape (the mood), but you can't capture the specific features (the identity). This creates a natural "bottleneck" that separates the two automatically, without needing complex, expensive training tricks.

2. The "Tiny Device" Advantage

Most AI models for this task are like moving trucks—they are huge, require a lot of fuel (energy), and need a highway (the internet) to get to the destination.

MP-IB is a bicycle: It is incredibly small (only 617 KB, which is smaller than a single high-res photo).
It runs on a device smaller than a deck of cards (Raspberry Pi Zero 2W).
It processes sound in 23 milliseconds (faster than a human blink), allowing for real-time monitoring without waiting for the cloud.

3. The "Privacy Shield"

Because the device is so small and efficient, the audio never leaves the device.

The paper claims that the "Mood Head" is so blurry that it cannot identify the speaker. If you tried to guess who was speaking based on the mood data, you would be right only about as often as if you guessed randomly (like flipping a coin).
The authors added a "static noise" layer to the data, making it even harder for anyone to reverse-engineer the speaker's identity.

4. Why Bigger Isn't Better Here

The researchers tested their tiny bicycle against massive "moving trucks" (huge AI models with millions of parameters).

The Result: The massive models failed. They got confused by the small amount of medical data available and actually performed worse than random chance.
The Winner: The tiny, precision-focused MP-IB model won. It learned that in a small-data world, being smart about what you forget (the identity) is more important than being able to remember everything.

5. Real-World Performance

Accuracy: It successfully detected agitation (a state of high stress or restlessness) with a correlation score of 0.117. While this number sounds small, in this specific field of difficult medical data, it is significantly better than any other method tested (including hand-crafted rules and other AI models).
Transfer: When they tested it on a completely different dataset (actors pretending to be angry), it still worked well, proving it learned the concept of agitation, not just the specific voices in the training data.

Summary

The paper presents a new way to build AI for mental health: Don't make the model bigger; make it "blurrier" on purpose. By intentionally limiting the precision of the part of the AI that tracks mood, they forced it to forget the speaker's identity, creating a system that is fast, private, energy-efficient, and surprisingly accurate on tiny devices.

Technical Summary: Mixed-Precision Information Bottlenecks for On-Device Trait-State Disentanglement in Bipolar Agitation Detection

1. Problem Statement

The paper addresses the challenge of continuous, on-device monitoring of bipolar disorder (BD) agitation via voice biomarkers. A critical barrier in this domain is trait-state disentanglement: the need to separate stable speaker characteristics (traits, such as vocal tract anatomy and habitual prosody) from volatile affective signals (states, such as agitation-related pitch variation and speech rate changes).

Existing approaches face three primary limitations:

Privacy and Latency: Current systems rely on cloud-based inference with large pre-trained models, creating barriers regarding connectivity, latency, and patient privacy.
Computational Cost: Traditional disentanglement methods (adversarial training, mutual information minimization, vector quantization) are computationally expensive, training-unstable, and ill-suited for resource-constrained edge devices.
Small-Data Generalization: Foundation models (e.g., WavLM, ECAPA-TDNN) with millions of parameters fail to generalize on small clinical datasets (N=833 participants) due to overfitting, often performing worse than chance.

The authors pose three research questions: Can numerical precision asymmetry serve as a disentanglement mechanism? What is the optimal bit-width allocation for maximizing agitation prediction while suppressing identity leakage? Can the framework operate within the strict compute, memory, and privacy constraints of continuous on-device monitoring?

2. Methodology: MP-IB Framework

The authors propose MP-IB (Mixed-Precision Information Bottleneck), a framework that treats numerical precision itself as an information bottleneck to achieve trait-state separation without adversarial training.

2.1. Core Insight: Precision as Information Bottleneck

Grounded in rate-distortion theory, MP-IB assigns different bit-widths to different representation heads to control information capacity at the hardware level:

Trait Head (FP16): Encodes stable speaker identity with high capacity (1,024 bits for a 64-dimensional vector).
State Head (INT4): Encodes volatile agitation with low capacity (128 bits for a 32-dimensional vector).
This creates an 8× information asymmetry. The INT4 constraint forces the state head to discard identity-related information, while the FP16 head preserves it. This operates at the arithmetic level, enabling packed-integer SIMD acceleration and unifying compression with disentanglement.

2.2. Architecture and Components

Shared Encoder: A MobileNetV3-Small (quantized to INT8) processes 96×64 log-Mel spectrograms.
Temporal Masked Autoencoder (T-MAE): To address small-data scarcity, the shared encoder is pre-trained on 12,000 hours of unlabeled Bridge2AI-Voice data using a T-MAE objective (reconstructing masked spectrogram patches). This provides a significant performance boost over baselines that lack in-domain pretraining.
Orthogonal Precision Loss (OPL): A deterministic loss function that enforces orthogonality between the trait and state embeddings. Unlike gradient reversal layers, OPL operates on upcasted embeddings to provide stable gradients without adversarial dynamics.
Dynamic Precision Scheduling (DPS): An adaptive mechanism that adjusts the effective precision of the state head based on input uncertainty (estimated via Monte Carlo dropout variance). This triggers higher precision (INT6) for challenging inputs (e.g., vocal pathology, whispers) to maintain accuracy.
Multi-Scale Temporal Fusion (MSTF): Aggregates state embeddings across three temporal resolutions (0.5s, 2s, 10s) using learned attention, while using mean pooling for trait fusion.

2.3. Privacy Mechanism

To prevent speaker re-identification from stored trait embeddings, the system employs output perturbation during the one-time onboarding phase. Calibrated Gaussian noise ( $\sigma = 25.3$ ) is injected into trait embeddings, reducing membership inference attack (MIA) success to near-random levels while maintaining clinical utility.

3. Key Contributions

Precision-Based Disentanglement: The first framework to use mixed-precision quantization (FP16 vs. INT4) as a principled information bottleneck for clinical trait-state separation, validated by a formal information-theoretic capacity bound (Theorem 3.1).
Privacy-Preserving Edge Deployment: An asymmetric deployment protocol achieving a 617 KB footprint and 23.4 ms end-to-end latency on a Raspberry Pi Zero 2W (sub-$20 device). The system suppresses identity leakage to near-random levels (EER=0.42, MIA-AUC=0.52).
Rigorous Empirical Validation: Systematic ablation studies, statistical significance testing (paired Wilcoxon), and comprehensive leakage metrics (EER, Top-k, MI estimation) demonstrating superiority over foundation models and classic disentanglement methods on small clinical data.

4. Experimental Results

Evaluated on Bridge2AI-Voice (N=833, 4 sessions/participant) with strict speaker-independent cross-validation:

Agitation Prediction: MP-IB achieves a Spearman correlation of $\rho = 0.117$ (95% CI: [0.089, 0.145], p=0.003 vs. chance).
- Outperforms 94M-parameter WavLM-Adapter ( $\rho = -0.042$ ) and ECAPA-TDNN-Adapter ( $\rho = -0.031$ ).
- Outperforms $\beta$ -VAE ( $\rho = 0.089$ ) and MI-Minimization ( $\rho = 0.095$ ) by 2.8–2.8 points.
- Outperforms hand-crafted prosody features ( $\rho = 0.031$ ) by 8.6 points.
Identity Leakage: The state head achieves an Equal Error Rate (EER) of 0.42 (near random 0.50) and a Top-1 speaker identification accuracy of 0.083 (10× above chance but significantly lower than baselines). With noise injection, MIA-AUC drops to 0.52.
Zero-Shot Transfer: On the CREMA-D dataset (anger detection as an agitation proxy), MP-IB achieves an AUC of 0.817, outperforming $\beta$ -VAE and MI-minimization baselines.
Efficiency: The system consumes ~318 Wh/year for continuous monitoring, representing a 140× energy reduction compared to a realistic cloud-mobile pipeline.

Ablation Findings:

T-MAE Pretraining is the largest single contributor (+0.083 $\rho$ ), mitigating overfitting on small data.
Precision Asymmetry contributes +0.052 $\rho$ compared to uniform quantization.
OPL contributes +0.036 $\rho$ .
DPS and MSTF provide marginal benefits and can be disabled for latency reduction without significant performance loss.

5. Significance and Claims

The paper claims that MP-IB demonstrates that precision-aware architectures with structured bottlenecks can surpass both scale-driven foundation models and classic disentanglement approaches when data is scarce and privacy is paramount.

Key claims include:

Small Models Win: On small clinical datasets, compact models with appropriate inductive biases (precision asymmetry) outperform massive foundation models that suffer from severe overfitting.
Hardware-Aware Learning: Numerical precision is a "first-class design choice" for representation learning, offering a mechanism for information allocation that is computationally efficient and theoretically grounded.
Clinical Utility: While the correlation ( $\rho = 0.117$ ) is modest, it enables clinically meaningful monitoring with a median time-to-detection of 4.2 hours, outperforming the 8–12 hour latency of inpatient nurse observations. The authors position the system as a clinician decision aid (flagging potential episodes) rather than an autonomous alert system, noting that further validation is required for autonomous deployment.

The work establishes a new direction for edge-based psychiatric monitoring, proving that high-fidelity trait-state separation is achievable on sub-$20 hardware without compromising privacy or requiring cloud infrastructure.