Synheart Capacity: A Theory-Driven Physiological Representation of Cognitive Capacity Dynamics from Wearable Signals
Yisak Debele, Henok Ademtew, Israel Goytom, [0.5em] Synheart AI, Vancouver, Canada · 2026
Read Synheart Capacity: A Theory-Driven Physiological Representation of Cognitive Capacity Dynamics from Wearable Signals by Yisak Debele, Henok Ademtew, Israel Goytom, [0.5em] Synheart AI, Vancouver, Canada (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-24416.txt.
Introduction
Debele, Ademtew, and Goytom propose a theory-driven frame for tracking cognitive capacity from wearables, using effort and strain. The design encodes cardiac and electrodermal signals in two streams and estimates effort and stress separately. Tested on SWELL-KW with leave-one-subject-out validation, it reaches 70 percent on stress and about 72 percent on effort. Rather than one workload label, the effort-stress space separates productive engagement from overload strain. Predicted states shift with demand under controlled changes, with effort and stress responding differently to interruption and pressure.
Catalog record
| CREATOR | Yisak Debele, Henok Ademtew, Israel Goytom, [0.5em] Synheart AI, Vancouver, Canada |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 56650 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.24416
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Synheart Capacity: A Theory-Driven Physiological Representation of Cognitive Capacity Dynamics from Wearable Signals Yisak Debele, Henok Ademtew, Israel Goytom, [0.5em] Synheart AI, Vancouver, Canada Abstract Human cognitive performance is constrained by limited mental resources, yet continuous computational estimation of cognitive capacity dynamics remains an open challenge. We propose a theory-driven multimodal learning framework that models capacity-related cognitive state as a two-dimensional physiological representation defined by voluntary resource allocation (mental effort) and overload-related strain (stress). The proposed architecture combines dual-stream encoding of cardiac (IBI/HRV) and electrodermal (EDA) signals with late fusion and task-specific output heads that independently estimate probabilistic effort and stress states. Evaluation on the SWELL-KW dataset using strict leave-one-subject-out cross-validation demonstrates cross-individual generalization (stress: 70.0% balanced accuracy; effort: 72.2%), with significant gains from multimodal integration and theory-guided supervision. Rather than collapsing physiological dynamics into a single workload label, the proposed effort–stress state-space enables structured differentiation between distinct cognitive regimes, including productive engagement and overload-related strain. Predicted state trajectories exhibit significant demand-sensitive shifts under controlled workload manipulations, with effort and stress responding differentially across interruption and time-pressure conditions. These results suggest that physiologically grounded multidimensional state representations may provide a foundation for adaptive systems capable of continuous capacity-aware monitoring and human-centered interaction. Keywords capacity theory \cdot mental workload \cdot physiological computing \cdot deep learning \cdot multi-task learning \cdot real-time estimation \cdot cardiac signals \cdot electrodermal activity \cdot cognitive load \cdot adaptive systems 1 Introduction 1.1 Background and Motivation Human performance failures—whether in classrooms, workplaces, or safety-critical environments—often emerge when cognitive demands exceed available mental resources. Capacity theory proposes that individuals possess limited processing resources that must be dynamically allocated to task demands, and performance deteriorates when demand approaches or exceeds available capacity ( 16 ; 28 ) . Within this framework, voluntary resource allocation (mental effort) and involuntary strain responses (stress) represent two interacting mechanisms governing performance under load ( 14 ; 19 ) . In this work, we use the term capacity state to denote the dynamic configuration of effort (resource allocation) and stress (overload pressure) that reflects an individual’s functional cognitive regime. While capacity theory has guided decades of behavioral and neurocognitive research, modern intelligent systems remain largely unaware of users’ cognitive limits. Adaptive tutoring platforms, decision-support tools, and human–automation interfaces increasingly aim to personalize difficulty and assistance ( 7 ; 24 ) . However, these systems typically rely on behavioral metrics such as error rates or response times, detecting overload only after performance has degraded ( 2 ) . Without continuous estimation of cognitive state, AI systems cannot distinguish between productive engagement and emerging overload. To investigate this challenge, the present work draws on the SWELL Knowledge Work dataset ( 17 ) , a publicly available corpus of physiological recordings from 25 participants performing office tasks under controlled cognitive demand manipulations. A central challenge lies in the fact that capacity is a latent construct: it cannot be measured directly but must be inferred from allocation and strain processes. Traditional operationalizations—dual-task paradigms, subjective workload ratings (e.g., NASA-TLX), or neuroimaging—provide indirect and discontinuous measures ( 10 ; 26 ) . These methods lack ecological validity and are unsuitable for real-time adaptive deployment. Effort and stress are theoretically related but not equivalent constructs. High effort may reflect adaptive engagement under manageable demand, whereas high stress may indicate that task demands are approaching or exceeding available cognitive resources ( 14 ; 19 ) . As a result, collapsing these processes into a single workload label obscures an important distinction between sustainable engagement and overload-related strain. This distinction is particularly important for adaptive human-centered systems, which may require fundamentally different interventions depending on whether a user is productively challenged or approaching cognitive overload. Physiological signals offer a pathway toward continuous, objective inference of cognitive dynamics. Cardiac variability reflects autonomic regulation associated with executive control and effort mobilization ( 21 ; 27 ) , while electrodermal activity indexes sympathetic arousal linked to stress and strain ( 4 ; 8 ) . With the proliferation of wearable sensors and edge-capable machine learning, these signals can now be acquired and processed unobtrusively in real-world contexts ( 25 ; 5 ) . Figure 1: Illustration of capacity theory with and without monitoring. In the absence of capacity awareness, cognitive demands exceed available resources, resulting in overload. Wearable-based monitoring provides early detection and intervention, allowing recovery before capacity saturation. 1.2 Related Work Research on physiological measurement of cognitive states has progressed along three primary directions. Classical psychophysiology. 21 demonstrated that cardiac measures—particularly heart rate variability—track task difficulty under laboratory conditions. 4 showed that electrodermal responses increase under elevated cognitive demand. These foundational studies linked physiological signals to resource mobilization and strain, but relied on predefined epochs and manual feature engineering, treating workload as a unidimensional construct without explicitly distinguishing voluntary effort from overload-induced stress. Machine learning approaches. 15 combined EEG and peripheral measures to classify workload levels, while 9 demonstrated wrist-worn stress detection using convolutional networks. 13 reported smartwatch-based cognitive load prediction during office work. These approaches confirm that automated detection from wearable signals is feasible, but most models optimize for discrete workload categories, treating stress, workload, and effort as interchangeable labels. Deep learning for temporal state modeling. 18 applied LSTMs for multimodal emotion recognition; 29 and 30 explored CNN-based representations for cognitive load classification. While these architectures capture dynamic patterns, they are designed to maximize classification performance rather than operationalize theoretical constructs. Existing work on SWELL-KW. Using supervised classifiers with HRV features and k-fold cross-validation, prior work achieved accuracies up to 74.8% ( 6 ) . Under strict subject-independent leave-one-out evaluation, a Random Forest trained on combined HRV and EDA features achieved only 42.5\%\pm 19.9\% balanced accuracy ( 22 ) . Critically, no existing approach has simultaneously estimated both effort and stress as structurally distinct outputs within a theory-consistent capacity state-space. 1.3 The Critical Gap Empirical inspection of the SWELL-KW dataset illustrates the cost of single-dimensional modeling concretely. Across 25 participants, subjective mental effort and perceived stress exhibited only weak covariation ( r=0.37 , p=0.001 ), sharing less than 14% of their variance. The interruption condition (c2) elicited effort levels statistically equivalent to those under time pressure (c3; paired t -test: \Delta=-0.21 , p=0.49 ), yet produced substantially lower stress ( \Delta_{\text{stress}}=+0.27 relative to neutral, vs. +0.94 under time pressure). Among participants in c2, 15 of 25 exhibited a high-effort/low-stress profile—a state that a single stress classifier collapses with the low-demand baseline. Figure 2: Empirical evidence for effort–stress dissociation in SWELL-KW ( n=25 participants). (a) Subjective effort ( U ) and stress ( O ) ratings across all subject \times condition observations ( r=0.37 ). Substantial scatter populates the Motivated Engagement quadrant (high U , low O ). (b) Condition profiles (mean \pm SEM). The interruption condition (c2) elevates effort significantly ( p=0.011 ) while stress remains low, producing a 3.37-point gap. Time pressure (c3) raises both dimensions. (c) Quadrant occupancy per condition. In c2, 15 of 25 participants occupy the Motivated Engagement quadrant—working hard without overload. Figure 3: Capacity theory framework illustrating cognitive resource utilisation as a sigmoid function of task demand. Three zones are delineated — underutilisation, optimal engagement, and overload — bounded by a fixed capacity ceiling. Effort (voluntary resource allocation) and stress