Differentially private (DP) fine-tuning methods treat sparse Mixture-of-Experts (MoE) models as a single dense block, ignoring that shared layers see all data while experts only see routed records. We identify and formally characterize three resulting failure modes: global clipping suppresses expert gradients, batch-level normalization dilutes sparse expert updates, and fixed privacy noise degrades signal-to-noise ratio on low-load experts. We introduce RAPTOR - a Role-Aware Private Training framework, which alternates shared and expert optimization and targets each failure directly, using expert-specific clipping and noise together with a public expected-owner denominator and a count-independent update schedule that avoids conditioning on private, realized expert counts. We prove the resulting mechanism satisfies $(\varepsilon,\delta)$-DP: because each record is assigned to exactly one owner expert, per-expert mechanisms within a layer compose in parallel, so updating all $E$ experts costs no more, in privacy terms, than updating one, with shared and expert streams composing sequentially across training. We further derive a bias-variance decomposition of the public-denominator estimator showing its bias grows predictably with routing imbalance, yielding a privacy-free rule for selecting which layer to protect from routing entropy measured on a small public corpus. Experiments on Switch Transformer and OLMoE fine-tuning across GLUE tasks, and on DeepSeek-VL2-Tiny, show consistent gains over standard DP baselines across several privacy levels ($\varepsilon$), with the largest margins typically at the tightest budgets. Code and models are publicly available: https://github.com/leduckhai/RAPTOR
Duc Dm, Khai Le-Duc, D. Nguyen et al.· 0 citations
BACKGROUND
Mental health conditions such as depression, anxiety, and stress are commonly assessed using self-reported questionnaires and limited wearable physiological measures. However, reliance on subjective reporting, restricted sensor modalities such as heart rate variability and electrodermal activity, and small or homogeneous datasets may limit generalizability. We aimed to evaluate whether wearable optical sensing of microcirculation and tissue metabolism enables objective assessment of stress-related mental health states.
METHODS
We conducted a prospective observational study including 132 adults aged 18 to 94 years (58% female) from 19 countries. Participants underwent repeated fingertip measurements using a non-invasive wearable device combining laser Doppler flowmetry and fluorescence spectroscopy to capture microvascular perfusion and metabolic signals. Frequency-domain features were extracted using wavelet analysis. Depression, anxiety, and stress levels were assessed using a standardized 21-item questionnaire. Multiple machine learning models were evaluated under subject-wise validation, and model interpretability was assessed using Shapley-based feature attribution.
RESULTS
Here we show that ensemble-based models distinguish individuals with stress-related symptoms from those without with a receiver operating characteristic area under the curve of 0.72 and a precision-recall area under the curve of 0.89 under subject-wise validation. Microcirculatory variability and metabolic fluorescence features contribute substantially to prediction performance. Demographic variables, including sex, age, body mass index, and heart rate, are associated with increased stress-related risk.
CONCLUSIONS
Wearable optical sensing combined with interpretable machine learning provides physiological signatures associated with stress-related mental health conditions. This framework supports development of scalable and data-driven tools for objective mental health monitoring.
