From Biosignals to Adaptive Systems

Research question. What has to happen before physiological signals can become reliable inputs for adaptive systems rather than isolated measurements?

Biosignals become useful for adaptive systems only after they are transformed into stable, interpretable summaries of state. For heart rate variability, for example, one common time-domain feature is RMSSD:

HRV Feature

\mathrm{RMSSD} = \sqrt{\frac{1}{N-1}\sum_{i=1}^{N-1}(RR_{i+1} - RR_i)^2}

The point is not the formula itself, but what it represents: a choice about windowing, preprocessing, and physiological interpretation. The classical HRV standards made clear long ago that measurement choices directly affect the meaning of the resulting metric [1].

That issue shows up immediately in multimodal comfort and office studies such as HERO and HEROx, where physiological channels cannot be interpreted independently from protocol stage, thermal exposure, and subjective report. A feature only becomes useful when it remains meaningful after synchronization, quality control, and contextual annotation.

The same logic applies to electrodermal activity. A practical decomposition writes the observed signal as:

EDA Decomposition

s(t) = tonic(t) + phasic(t) + \varepsilon(t)

Separating tonic and phasic components is not cosmetic. It determines whether a model is seeing a slow baseline shift, a rapid response, or sensor noise. Modern decomposition methods emphasize that EDA analysis is fundamentally a signal-separation problem, not just a scalar feature problem [2].

Why feature quality matters more than feature count

Adaptive systems need representations that are stable under change. If HRV is extracted on inconsistent windows, if EDA is left entangled with motion artifacts, or if EEG is filtered without documenting frequency assumptions, then the downstream model is learning preprocessing quirks as much as physiology. EEG signal-processing reviews make this explicit: feature quality depends on artifact handling, transform choice, and domain-specific interpretation [3].

In practice, I think the feature layer should be treated as a mapping from synchronized multimodal windows to interpretable state variables:

Feature Mapping

\phi_t = \Phi(EEG_t, HRV_t, EDA_t, env_t, q_t)

Here $$env_t$$ captures environmental conditions and $$q_t$$ captures data quality. If $$q_t$$ is ignored, the system confuses signal with artifact. If $$env_t$$ is ignored, the same physiological pattern can be interpreted as the wrong state.

This is why I think biosignal-adaptive systems should be framed as systems problems rather than pure classification problems. Schultz and Maedche describe biosignal-adaptive systems as closed-loop systems that interpret implicit human signals and feed them back into adaptation [4]. That framing is useful because it forces the pipeline to be judged end to end: sensing, preprocessing, feature extraction, classification, and adaptation.

From signal to adaptation

For a biosignal to support adaptation, three conditions usually have to hold. First, the signal must be synchronized and quality-controlled. Second, the feature must remain interpretable under variation in context and time. Third, the decision layer must preserve uncertainty rather than pretending that every feature map is equally trustworthy. This is one reason datasets such as HERO, HEROx, and CLTR matter: they make it possible to study biosignals as part of a complete multimodal system rather than as detached measurements.

From biosignals to adaptive systems

Why feature quality matters more than feature count

From signal to adaptation

References