Title: Ensuring Ethical Equity in General Health Monitoring Systems Without Relying on Known Attributes

Abstract:

Computational models in ubiquitous and mobile health ecosystems typically deduce human conditions by analyzing data streams from wearable, behavioral, and physiological sensors. In these environments, achieving high predictive accuracy is not enough; models are required to operate ethically and equitably across a wide spectrum of individuals, contexts, and hardware devices. However, traditional fairness techniques that depend on demographic or heterogeneous attributes during the training phase face significant implementation hurdles, as such data is frequently inaccessible, subject to privacy concerns, heavily regulated, or simply unwanted to gather. Furthermore, standard parity-based fairness approaches may conflict with ethical standards by compromising the performance of specific subgroups.

To overcome these limitations, we introduce Flare (Fisher-guided LAtent-subgroup learning with do-no-harm REgularization), a framework designed to be agnostic to demographic and heterogeneous attributes. This approach aligns human-centric fairness with ethical principles tailored for ubiquitous and mobile sensing. Flare utilizes optimization geometry, specifically leveraging Fisher Information, to regularize curvature and detect latent disparities in model behavior without needing explicit demographic data. By synthesizing signals from representation, loss, and curvature, the method identifies hidden performance layers and refines them through collaborative optimization that adheres to a "do-no-harm" principle, thereby boosting subgroup outcomes while maintaining ethical equilibrium.

Additionally, we propose BHE (Beneficence-Harm Avoidance-Equity), a comprehensive metric suite that extends the definition of ethical fairness beyond mere statistical parity. Evaluated across various mobile physiological, behavioral, and clinical sensing datasets—including EDA, OhioT1DM, IHS, and Percept-R—Flare demonstrates superior ethical fairness compared to current state-of-the-art baselines. Analyses of loss landscapes, interpretability, and ablation studies indicate that these improvements stem from flatter optimization geometry, more straightforward decision rules, and latent-subgroup adaptation that avoids harm. Runtime evaluations further confirm Flare’s viability for practical deployment in resource-constrained sensing environments.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC