Abstract
Emotions play an important role in human behavior, and their analysis is often a necessary step in correct interpretation of said behaviors, providing context and depth. In active assisted living (AAL) applications, affective computing methods are used to understand a user’s affective state, as well as provide systems interacting with the user with improved communication capabilities. In this chapter, we provide an overview of affective computing and its uses in AAL, discuss issues of trust and responsible AI, as well as provide a short overview of the AI Act as an example legal framework that regulates the use of affective computing in AAL.
Introduction
Traditional Active Assisted Living (AAL) systems are highly effective at detecting physical events, such as falls or changes in walking speed. However, human behavior analysis is fundamentally incomplete without emotional context. To truly support independent living for the elderly or those with chronic conditions, systems need to understand "affect"—moods, feelings, and stress levels.
The core challenge is that emotional data is deeply personal and often communicated unconsciously. This creates a tension between the benefits of proactive care and the risks of privacy invasion, algorithmic bias, and the loss of human dignity. This paper seeks to bridge the gap between technical sensor capabilities and the emerging "Responsible AI" frameworks required for safe deployment.
Sensing Modalities and Methods
The paper categorizes Affective Computing (AC) methods based on a specific taxonomy of modalities, distinguishing between wearable and ambient implementations.
Visual and Physiological Modalities
Visual analysis is moving from traditional pipelines (face detection → landmarks → classification) to end-to-end deep learning, utilizing tools like OpenFace. Physiological sensing uses wearables to track heart rate variability (HRV) and skin conductivity (GSR). Ambient alternatives, such as Doppler radar for gait analysis, offer less intrusive monitoring that maintains a higher level of user confidentiality than cameras.
Audio and Text
Researchers distinguish between linguistic (what is said) and paralinguistic (how it is said, e.g., prosody) features. A key technical insight is that acoustic features are generally better at discriminating arousal (the energy of an emotion), while linguistic features are superior for determining valence (whether the emotion is positive or negative).
Key Insights and Regulatory Landscape
The "High-Risk" Reality
Under the EU AI Act, most AC systems in AAL are classified as "high-risk" because they use biometric data to infer emotions. This imposes strict transparency and data security requirements. While the Act prohibits emotion recognition in workplaces and schools, it carves out exceptions for medical and safety reasons, protecting the primary use cases of AAL.
Performance and Interpretability
Fine-tuning self-supervised models like Wav2Vec 2.0 has led to unprecedented accuracy in predicting emotional valence. Furthermore, the paper emphasizes a vital distinction: Interpretability involves designing models that are inherently transparent, which is preferred over Explainability (post-hoc attempts to explain "black box" models).
Limitations and Future Work
Despite progress, several challenges remain:
- The "In-the-Wild" Gap: Systems still struggle with complex activities outside laboratory settings.
- Demographic Bias: There is a significant need for data from older adults to reduce age-related bias, as most current models are trained on younger populations.
- The Human Substitution Risk: Technology should not substitute for human contact, which could lead to increased social isolation.
- Informed Consent: Developing better methods for seeking consent in "smart environments" where ambient sensors might pick up data from visitors or caregivers is essential.
Summary updated