The Neuro-Cardiac Symbiotic Engine: A Multimodal Fusion Architecture for Cognitive State Decoding via High-Performance Computing Authors: Nayeli Bastidas-Benalcazar, Julián A. Calero-Apunte, Diego Almeida-Galarraga, Paulo Navas-Boada, Omar Alvarado-Cando, Andrés Tirado-Espín, Fernando Villalba-Meneses, Henry Carvajal Mora, Nathaly Orozco Garzón Published: 2026 ABSTRACT Robust decoding of latent cognitive states from non-stationary physiological time series is a challenging high-dimensional signal processing problem. Traditional unimodal frameworks based only on electroencephalography often show covariate shift and weak cross-task generalization. This study presents the Neuro-Cardiac Symbiotic Engine, a multimodal fusion architecture that combines high-frequency cortical EEG dynamics with low-frequency autonomic regulation derived from heart rate variability within a unified discriminative feature space. The pipeline integrates spectral decomposition and autonomic quadratic descriptors through a memory-optimized high-performance computing workflow on the CEDIA supercomputer. To reduce domain discrepancy between memory and piloting tasks, we design a few-shot calibration strategy based on affine manifold alignment and probabilistic ensemble inference. Validation on 29 subjects reaches a mean classification accuracy of 99.13 percent, far above the zero-shot baseline near 38 percent. Topological analysis also indicates phase-space contraction under high workload, where fused vagal and frontal-parietal biomarkers concentrate system dynamics into a low-entropy attractor. The results establish a mathematically grounded framework for passive brain–computer interfaces and show that orthogonal neuro-visceral integration is critical for reliable cognitive state estimation. 1. Introduction The quantitative decoding of Cognitive Workload (CWL) constitutes a central challenge in modern neuroergonomics, particularly within high-stakes operational environments such as air traffic control and robotic teleoperation where human error can be catastrophic. The capacity to dynamically adapt system autonomy based on the operator’s neural capacity is paramount for preventing failure, yet current methods often lack the robustness required for real-world deployment, even when evaluated on publicly available multi-session benchmarks such as the COG-BCI database [1]. CWL is not merely a subjective sensation but a measurable bio-energetic cost incurred during information processing; when this cost exceeds the operator’s channel capacity (overload), performance degrades non-linearly, precipitating critical errors [2]. While Electroencephalography (EEG) provides high-temporal-resolution access to cortical dynamics, identifying reliable correlates of mental effort remains fraught with challenges due to non-stationarity, sensor drift, and severe inter-session variability [3]. Consequently, a classifier trained on a controlled task (e.g., N-Back) typically fails to generalize to naturalistic environments, a limitation known as “Negative Transfer.” To transcend cortical limitations, this study adopts the Neurovisceral Integration Model proposed by Thayer et al. [4]. This theoretical framework posits that the Prefrontal Cortex (PFC)—the center of executive function—exerts inhibitory control over the amygdala and the autonomic nervous system via the Vagus Nerve. Consequently, Cognitive Workload is not solely a brain phenomenon but a systemic neuro-visceral process that requires holistic monitoring. According to Porges’ Polyvagal Theory [5], high cognitive demand requires the suppression of the “Vagal Brake,” leading to a measurable reduction in Heart Rate Variability (HRV). Specifically, the Root Mean Square of Successive Differences (RMSSD) serves as a robust, non-invasive biomarker of vagal tone and prefrontal inhibitory capacity [6,7]. By integrating these autonomic metrics with cortical oscillations, we hypothesize that the cardiac signal acts as a “physiological ground truth,” stabilizing the high-variance EEG feature space. Recent advances in 2024–2025 have begun to explore multimodal fusion as a solution to these stability issues. Salam et al. [8] demonstrated that combining ECG and EEG significantly improves stress classification accuracy compared to unimodal models by leveraging complementary physiological information. Similarly, Xiong et al. [9] demonstrated that combining EEG and ECG parameters provides complementary information, significantly improving the recognition of cognitive load conditions compared to single-modality systems. However, most existing pBCI systems (e.g., [10]) operate within single-task paradigms and lack rigorous cross-domain validation, leaving an important gap in our understanding of how these models behave under distribution shift. Our work addresses this gap by introducing a multimodal neuro-cardiac fusion framework that explicitly models brain–heart coupling and is evaluated under both zero-shot transfer and few-shot adaptation from memory tasks to piloting simulations. The primary objective of this study is to engineer and validate a generalizable multimodal artificial intelligence framework for robust cognitive workload level (CWL) classification under non-stationary and cross-domain conditions. By leveraging a multimodal fusion architecture that integrates cortical EEG dynamics with autonomic cardiac regulation through early feature concatenation followed by ensemble inference, the proposed system aims to overcome the limitations of unimodal and task-specific models. Particular emphasis is placed on achieving stable performance across heterogeneous cognitive contexts while maintaining interpretability and computational tractability. This objective is grounded in the neurovisceral integration paradigm, which conceptualizes cognitive workload as a systemic neuro-cardiac process rather than a purely cortical phenomenon, ensuring that the resulting model is both mathematically sound and physiologically plausible. This study is guided by a central research question on neuro-visceral invariance: does the fusion of vagal tone (RMSSD) and fronto-parietal beta-band EEG power define a domain-invariant biomarker of cognitive workload across heterogeneous tasks? We examine whether this coupling remains stable across controlled memory paradigms and complex piloting simulations, under the hypothesis that autonomic regulation provides a metabolic anchor that stabilizes high-variance cortical features. Demonstrating this invariance is essential for mitigating covariate shift and improving cross-task generalization in passive brain–computer interface systems, where EEG-only models are known to degrade under domain transfer [3,10]. The Neurovisceral Integration Model posits that cognitive control is mediated by prefrontal–autonomic coupling via vagal pathways, making heart rate variability—particularly RMSSD—a robust marker of executive regulation and cognitive effort [4,11]. Prior multimodal studies have shown that combining EEG and cardiac signals improves classification accuracy, yet they lack evidence of cross-domain invariance [9]. Therefore, investigating whether neuro-cardiac fusion yields a stable, task-independent physiological manifold is essential for establishing generalizable cognitive workload decoding. A second research question addresses whether few-shot calibration based on a multimodal neuro-cardiac signature can align feature distributions between memory and piloting domains. Specifically, we evaluate whether affine manifold alignment and probabilistic ensemble inference can transfer representations from a source memory task to a target piloting simulation with minimal labeled data. This question targets a critical translational bottleneck: rapid deployment in operational settings where extensive recalibration is impractical. Successful few-shot adaptation would support scalable, real-time cognitive state decoding across heterogeneous domains. Domain adaptation remains a central challenge in passive brain–computer interfaces, as models trained on controlled laboratory tasks often fail in ecologically valid environments due to distributional mismatch [12,13]. Few-shot learning has emerged as a promising paradigm for reducing calibration costs by leveraging compact, informative representations that require minimal labeled data [14]. From a physiological perspective, autonomic markers such as RMSSD provide a low-frequency, metabolically grounded reference signal that may act as an anchor for aligning high-dimensional cortical feature spaces [4,5]. Consequently, evaluating whether few-shot affine alignment of multimodal neuro-cardiac features can bridge memory and piloting domains is critical for enabling rapid, scalable deployment of cognitive monitoring systems in real-world operational settings. To contextualize the proposed architecture within the broader landscape of neuroergonomics, we conducted a systematic review of recent methodologies. As delineated in Table 1, the field has progressively shifted from unimodal, single-task validations toward complex deep learning frameworks. However, a significant gap remains regarding the mathematical formalization of cross-domain transfer without incurring prohibitive computational costs. While the progression from unimodal to multimodal systems is evident, the architectural choices differ significantly in mathematical philosophy. Recent Transformer-based approaches, such as the framework proposed by Sasi et al. (2026) [20], achieve high accuracy (∼96%) by modeling long-range temporal dependencies via self-attention mechanisms. However, these Deep Learning architectures inherently suffer from high variance and lack interpretability (“Black Box” phenomenon). In contrast, our Neuro-Cardiac Symbiotic Engine favors a Random Forest Ensemble approach. By leveraging the Law of Large Numbers through Bootstrap Aggregating (Bagging), our model reduces the variance of the estimator by a factor of 1/B (where B is the number of trees), achieving superior stability (99.13% accuracy) with a fraction of the computational complexity (O(NlogN) vs. O(N2) for Transformers), making it viable for real-time edge computing. Furthermore, the treatment of non-stationarity distinguishes our work from geometric adaptation methods. Studies like Liu et al. (2023) [16] and Cai et al. (2024) [19] employ Riemannian Manifold Learning to align feature distributions. While mathematically elegant, projecting data onto the tangent space of the Riemannian manifold is computationally expensive and sensitive to noise. Our approach simplifies this via an Affine Z-Score Diffeomorphism, which aligns the marginal distributions of the cortical and autonomic subspaces efficiently. As demonstrated in our results, this linear alignment, anchored by the physiological stability of Vagal Tone (RMSSD), suffices to mitigate the covariate shift between memory and piloting tasks without the need for complex geodesic computations. In the realm of multimodal fusion, Salam et al. (2026) [8] and Xiong et al. (2020) [9] have set strong benchmarks using feature-level integration and sequential feature selection, respectively. However, their models typically treat ECG and EEG as additive components. Our methodology diverges by treating the cardiac signal as a Topological Anchor. Following the Neurovisceral Integration Model [4], we posit that the autonomic vector constrains the high-dimensional EEG manifold, reducing the entropy of the system during high-load states. This effectively collapses the search space for the classifier, allowing for Few-Shot calibration with only 30% of target data, whereas Deep Subdomain Adaptation Networks (e.g., Sun & Li, 2025 [13]) often require extensive unsupervised pre-training. Finally, recent topological-data-analysis frameworks such as the Hodge-FAST spectral analysis of Roy et al. (2025) [17] emphasize higher-order interactions within brain networks. While promising for connectivity analysis, such approaches often underrepresent the broader physiological context. In our results, the contraction of the joint neuro-cardiac state space provides complementary evidence that peripheral biomarkers are informative for resolving ambiguous brain states in unconstrained environments. To place these findings in context, Table 2 benchmarks the proposed model against representative recent architectures. In line with recent advancements in multimodal physiological computing, the integration of heart rate metrics has proven essential for robust cognitive state decoding. For instance, Salam et al. (2026) [8] emphasized that the feature-level fusion of EEG and ECG provides a superior index for cognitive stress than unimodal approaches, attaining highly accurate classification in stressful environments. Similarly, recent comprehensive reviews on Topological Data Analysis (TDA) emphasize that extracting topological invariants provides robust feature representations that are highly resistant to noise and artifacts in EEG signal processing [21]. Our work builds upon these findings by formalizing the neuro-cardiac coupling not just as a combined feature vector, but as a symbiotic system where the cardiac signal anchors the volatile cortical dynamics, ensuring mathematical consistency during domain transfer. The primary innovations and scientific contributions of this study can be summarized as follows. First, we introduce the Neuro-Cardiac Symbiotic Engine, a multimodal fusion framework in which high-frequency cortical spectral features (ΦPSD) and low-frequency autonomic quadratic descriptors (ΦHRV) are concatenated into a unified discriminative feature space before classification. Second, to mitigate the domain discrepancy between heterogeneous cognitive tasks (N-Back → MATB-II), we propose a few-shot manifold-alignment strategy based on affine Z-score normalization. This approach partially compensates for covariate shift and improves performance from a near-chance cross-task baseline to approximately 94% when only 5% of target-domain data are incorporated. Third, stochastic variability is reduced by exploiting ensemble averaging effects consistent with the Law of Large Numbers through a memory-optimized Random Forest implementation, leading to a marked decrease in performance variance (σ≈1.2%) and improving the stability of workload estimation across runs. Fourth, beyond heuristic modeling, we examine the relationship between empirical performance and theoretical expectations by assessing variance reduction, feature-space projection, and entropy-related behavior; the results suggest a phase-space contraction effect in which elevated cognitive load is associated with a progressive concentration of system dynamics into lower-entropy regions, providing a potential topological indicator of cognitive overload. Finally, reproducibility at scale is addressed through a memory-efficient processing pipeline implemented on the CEDIA supercomputer, incorporating lazy loading and file-system sanitation to support large multimodal datasets. The remainder of this article is organized as follows: Section 2 describes the mathematical formalization and computational workflow; Section 3 presents the empirical results and validation analyses; Section 4 discusses the physiological implications and limitations; and Section 5 concludes the study. 2. Methodology This study formalizes a mathematically explicit and reproducible computational pipeline for multimodal EEG–ECG analysis that directly confronts the stochastic non-stationarity, cross-modal coupling, and high dimensionality characteristic of neuro-cardiac biosignals [22]. The methodological exposition is organized into coordinated subsections that operationalize robust data engineering, rigorous stochastic signal-space formulation, and principled feature fusion. To provide a holistic view of the system’s topology—ranging from the stochastic acquisition of physical signals (XEEG,XECG) to the probabilistic inference of cognitive states—the complete architectural blueprint is delineated in Figure 1. To visualize the complete engineering journey from raw signal acquisition to final classification, Figure 2 details our end-to-end methodological architecture. This five-stage pipeline begins by rigorously sanitizing raw data artifacts to ensure signal purity before initializing the High-Performance Computing environment. We then implement a dual-stream processing approach to synthesize EEG and ECG dynamics, refining the model through a transfer learning calibration strategy. This systematic progression culminates in the validation phase, where the architecture demonstrates its robustness with a 99.13% accuracy rate. To operationalize the theoretical fusion of cortical and autonomic dynamics within a scalable computational environment, we engineered a comprehensive end-to-end processing architecture. This pipeline orchestrates the transformation of high-entropy raw biosignals into a low-dimensional, discriminative feature manifold through a sequence of five strictly controlled phases. The lifecycle begins with the ingestion of heterogeneous BIDS-formatted datasets, requiring a recursive sanitation protocol to eliminate non-signal artifacts such as macOS metadata and hidden resource forks, which can precipitate parsing errors during high-throughput analysis. Given the scale of the dataset (29 subjects, 64-channel EEG), the workload is migrated from local workstations to a High-Performance Computing (HPC) environment on the CEDIA Supercomputer, leveraging NVIDIA DGX A100 nodes. To bypass the physical constraints of the 32GB RAM quota, we implemented a “Lazy Loading” evaluation strategy via the MNE-Python backend, ensuring that only necessary data segments reside in volatile memory, thereby maintaining system stability below a 4GB peak usage during the extraction of spectral power densities and autonomic quadratic forms [23,24]. Following feature synthesis, the architecture addresses the distribution shift between source and target domains through a specialized Few-Shot Transfer Learning strategy. By pre-training the model on a source memory task and injecting a calibrated subset (30%) of target piloting data, the system successfully aligns the multimodal manifold to the operational reality of the flight simulation [12,25]. This process involves a three-stage discriminative filtration pipeline where ANOVA F-tests isolate the top 300 most informative features, which are then processed by a Random Forest ensemble of 100 trees to execute probabilistic soft voting. The final phase of the workflow subjects the calibrated model to a rigorous 50-iteration Monte Carlo Bootstrap framework, ensuring that the achieved 99.13% accuracy is not a stochastic artifact but a reproducible result of the neuro-cardiac coupling. This comprehensive methodology establishes a mathematically grounded blueprint for deploying reliable passive Brain–Computer Interfaces in safety-critical environments where data stationarity is rarely guaranteed [26,27]. 2.1. Data Engineering and Computational Infrastructure Formalization of the Signal Space. The empirical basis of this study is the COG-BCI dataset [1], which is formally defined as a collection of multimodal observations D={(XEEG(i),xECG(i),y(i))}i=1N acquired from a cohort of N=29 subjects. In accordance with established multimodal neurophysiological acquisition protocols [26], the physiological signal space is modeled as the Cartesian composition of two mutually orthogonal subspaces, embedded within a structured experimental manifold. The cortical component is represented as a multivariate time-series matrix XEEG∈RC×T, where C=64 denotes the number of scalp electrodes and T the temporal extent of the recording, sampled at the aforementioned frequency [23]. Complementarily, the autonomic component is described by a univariate electrocardiographic signal xECG∈R1×T, recorded in a standard Lead I configuration and temporally synchronized with the EEG stream via hardware-level triggering (Δt→0), thereby ensuring negligible cross-modal latency [11]. The resulting multimodal observations are further contextualized within an experimental manifold that partitions the state space into three discrete cognitive domains grounded in contemporary cognitive taxonomy [28]: a baseline resting condition (Sbase), a source domain associated with controlled working-memory engagement through the N-Back task (Ssrc), and a target domain corresponding to ecologically valid multitasking demands as instantiated by the NASA-MATB-II piloting simulation (Stgt) [29,30]. 2.1.1. HPC Architecture and Space Complexity Optimization The spectral feature extraction from terabytes of raw neurophysiological data necessitated a migration to the CEDIA Supercomputing Cluster. The computational workload was distributed across NVIDIA DGX A100 nodes, leveraging the high-throughput bandwidth of dual AMD EPYC processors. A critical boundary condition was the partition-enforced memory quota Mmax=32 GB. Under a standard eager-loading paradigm, the space complexity S(n) for a single continuous recording session scales linearly with recording duration T:(1)SRAM(T)≈C·fs·T·δbits+ϵ where δbits is the floating-point precision and ϵ represents system overhead. For extended sessions, SRAM>Mmax, triggering kernel-level termination. Optimization Strategy: To satisfy the constraint SRAM≪Mmax, we engineered a Memory-Mapped I/O (Lazy Loading) architecture via the MNE-Python backend [23]. By initializing ingestion pointers with a lazy evaluation strategy, the signal matrix is mapped directly to the physical address space on the Lustre filesystem rather than volatile memory. This transformation reduces the asymptotic space complexity of the ingestion phase:(2)O(N)⟶O(1) This architectural decision ensures that Random Access Memory usage remains bounded (<4 GB peak), guaranteeing stability regardless of the temporal dimension T. 2.1.2. Data Integrity and Recursive Artifact Sanitation A critical divergence between the acquisition file system (macOS/APFS) and the processing environment (Linux/CentOS) introduced binary incompatibility risks [31]. Specifically, the injection of AppleDouble resource forks (prefixed metadata artifacts) mimics valid signal headers but lacks the tensor structure required for deserialization. To enforce deterministic execution, we implemented a Pre-Ingestion Sanitation Algorithm (Aclean) operating at the file system level. Let Fraw be the set of all files in the directory tree. The algorithm applies a recursive inode traversal to filter the set such that:(3)Fvalid=Fraw∖{f∈Fraw∣is_artifact(f)} By procedurally removing non-signal binary objects prior to pipeline initialization, this operation guarantees that the input set Iinput forms a bijective mapping to the valid signal set Ivalid. 2.2. Overview of the Multimodal Neuro-Cardiac Inference Pipeline This inference engine is constructed as a sequential transformation pipeline, Ψ:(Vctx×Vaut)→Y, designed to preserve the orthogonality of the input sources while mitigating the curse of dimensionality via statistical filtration. The complete anatomical structure of this hybrid classifier—ranging from the early concatenation of spectral and metabolic vectors to the probabilistic soft-voting mechanism of the Random Forest ensemble—is rigorously schematized in Figure 3. 2.3. Signal Conditioning and Stochastic Feature Extraction Let Sraw={XEEG,xECG} denote the raw biosignal set. We formalize the processing pipeline as a composition of measurable operators mapping the non-stationary time domain RT to a stationary feature manifold. 2.3.1. Spectral Transformation Operator (Phi1) The EEG time-series XEEG undergoes a zero-phase FIR bandpass filter (0.5–45 Hz). To estimate the energy distribution, we utilize Welch’s Method. We formalize this not merely as an algorithm, but as an averaged periodogram estimator. For a window function w[n] with energy normalization U, the estimator S^c(f) for channel c is defined as:(4)S^c(f)=1KU∑k=0K−1∑n=0L−1w[n]xc[n+kD]e−j2πfn2 where K is the number of overlapping segments. The feature vector vEEG is derived by integrating the spectral density over canonical bands B:(5)vc,b=∫f∈bS^c(f)df∀c∈{1…64},∀b∈B Fix channel c and band b. There exists a symmetric positive semi-definite matrix Qc,b∈RT×T such that the extracted band-power admits the quadratic form representation: (6) vc,b=xc⊤Qc,bxc Consequently, the spectral features are strictly non-negative convex functionals of the time-domain signal, preserving the signal energy structure [32]. A detailed visual decomposition of this operator, from the windowed Fourier transform to the final band-power projection, is provided in Figure S1 of the Supplementary Information. 2.3.2. Stochastic Modeling of Autonomic Regulation (Phi2) To mitigate computational overhead, we implemented a Native Vectorized Peak Detection algorithm. From the resulting R-R interval series R={ri}i=1N, we derive robust metrics defined as statistical moments. We formalize these metrics as quadratic forms over the vector space RN: Let D be the forward-difference matrix and H the centering matrix (H=IN−1N11⊤). (7) RMSSD(R)2=R⊤D⊤DN−1R (8) SDNN(R)2=1NR⊤HR Assuming the inter-beat interval process is weak-sense stationary and ergodic, as N→∞, the sample metrics converge almost surely to their population counterparts: RMSSD2→a.s.E[(Δr)2] This ensures that our 2-min windowing strategy provides a consistent estimator of Vagal Tone [33]. 2.4. The Neuro-Cardiac Integration Architecture The architectonic pivot of this framework is the synthesis of a composite feature manifold. 2.4.1. High-Dimensional Feature Synthesis (Early Fusion) We employ a feature-level concatenation strategy. The composite state vector xi is constructed via the direct sum of the feature spaces:(9)xi=vEEG(i)⊕vHRV(i)⇒xi∈R323This procedure generates a high-dimensional embedding in which cortical and autonomic descriptors coexist as complementary coordinates of the same physiological state space. 2.4.2. Isotropic Manifold Alignment (Standardization) A fundamental pathology in multimodal learning is the Physical Scale Discrepancy (μV2 vs. ms). Direct ingestion would result in an ill-conditioned optimization landscape. To enforce feature invariance, we apply Z-score standardization, formalized here as an Affine Diffeomorphism. Let μ∈RD be the sample mean and Σ=diag(σ1,…,σD) the scaling matrix. The map is defined as:(10)Z(x)=Σ−1(x−μ) Since σj>0, Z is a bijective affine mapping with Jacobian determinant detJZ=∏σj−1. Information-theoretically, this implies that the differential entropy shifts by a constant, but the Mutual Information structure I(Z;Y)=I(X;Y) remains invariant. This ensures no discriminative information is lost during normalization [34]. A visual formalization of this geometric interpretation as a topology-preserving map is provided in Figure S2 of the Supplementary Information. 2.5. Discriminative Subspace Projection via ANOVA The fused manifold RD contains sparse and redundant dimensions. To extract the optimal discriminative subspace, we formalize feature selection as a geometric Orthogonal Projection Operator. 2.5.1. Statistical Formulation of Class Separability We decompose the total variance of the feature space into Between-Class (SB) and Within-Class (SW) scatter components. For a feature j, the Fisher Score Fj is defined as the ratio of these variances:(11)Fj=MSbetweenMSwithin=∑c=1CNc(μj,c−μj)2∑c=1C∑i∈c(zij−μj,c)2 2.5.2. The Projection Matrix (PiS) Let S be the index set of the top k=300 features with the highest Fj scores. We define the Projection Operator ΠS as a diagonal matrix:(12)ΠS=diag(s1,…,sD)∈RD×D,wheresj=I(j∈S) The operator ΠS satisfies ΠS2=ΠS and ΠS⊤=ΠS. Consequently, the transformation x˜=ΠSx represents a true orthogonal projection onto the coordinate subspace spanned by the discriminative basis vectors, nullifying noise dimensions while preserving signal energy [35,36]. The mathematical framework for this selection process, interpreted as a geometric projection, is visualized in Figure 4. 2.6. The Inference Kernel: Stochastic Ensemble Learning The decision boundary is approximated by a Random Forest Ensemble, architected to minimize variance in high-dimensional spaces while maintaining low bias for non-linear physiological boundaries [37]. 2.6.1. Probabilistic Inference (Soft Voting) Unlike binary voting schemes, our system estimates the posterior probability P(y=c∣x). The ensemble constructs a robust estimate by averaging the outputs of B=100 de-correlated trees. The final decision rule y^ is:(13)y^=argmaxc∈{L,M,H}1B∑b=1BPb(c∣x;Θb) where Θb represents the random vector governing the b-th tree (bootstrap sample and feature subset). This averaging mechanism acts as a Stochastic Low-Pass Filter, smoothing out the decision boundaries. 2.6.2. Theoretical Variance Reduction The robustness of the engine is mathematically grounded in the variance decomposition theorem. Let ρ be the correlation between trees and σ2 the variance of a single tree. The ensemble variance satisfies:(14)Var(h¯B)=σ2ρ+1−ρB Implication: By employing Random Feature Subspaces [38], we minimize ρ. As B→∞, the second term vanishes, and the error converges to the irreducible correlation floor, validating the stability of our 50-iteration bootstrap results. The theoretical underpinnings of this learning engine, from the projection logic to the variance reduction principle of the ensemble, are summarized in Figure S3 of the Supplementary Information. To ensure the replicability of the Neuro-Cardiac Symbiotic Engine, a set of implementation standards and hardware assumptions are defined. Although the initial stages of feature extraction and manifold training were executed on the CEDIA High-Performance Computing Cluster (NVIDIA DGX A100), the resulting trained ensemble model remains lightweight, with a memory footprint of approximately 112 MB. As a consequence, real-time inference can be performed on standard consumer-grade CPUs, achieving inference latencies below 200 ms per window. This design choice enables straightforward transferability of the architecture to edge-computing and mobile platforms without reliance on GPU acceleration. Reproducibility across heterogeneous operating systems, including macOS, Linux, and Windows, is supported through a strict data sanitation protocol applied prior to data ingestion. The processing pipeline explicitly removes metadata artifacts commonly generated by file systems (e.g., ._* files) and enforces BIDS-compliant directory structures. These measures are critical to prevent indexing inconsistencies and silent parsing errors when handling large-scale multimodal EEG–ECG datasets distributed across different computational environments. Model stability is further ensured through a fixed and conservative hyperparameter configuration of the Random Forest ensemble, consisting of B=100 trees, Gini impurity as the split criterion, and d features evaluated per split. Sensitivity analyses indicate that classification performance remains robust under minor perturbations of these parameters, provided that the ANOVA-based feature filtration stage preserves at least k=300 features. This constraint is necessary to avoid excessive information loss during the cortical–autonomic feature fusion process. Transferability across acquisition setups additionally assumes minimum signal quality requirements. Specifically, EEG recordings must maintain a sampling rate of at least 250 Hz, and ECG signals must allow reliable R-peak detection. To address temporary degradations in cardiac signal quality, the system incorporates a fallback mechanism: if ECG integrity falls below a 60 s window required for stable RMSSD estimation, the manifold alignment procedure reverts to a cortical-only configuration. This transition entails a temporary reduction in the calibrated stability level but preserves operational continuity. Finally, the stochastic operators governing spectral EEG extraction (ΦPSD), autonomic variability characterization (ΦHRV), and the few-shot calibration logic have been modularized at the architectural level. This modular design facilitates adaptation of the proposed engine to alternative cross-domain scenarios, such as driving simulation environments or cognitive rehabilitation protocols, with minimal structural modification to the underlying processing pipeline. 3. Results and Analysis The empirical validation of the Neuro-Cardiac Symbiotic Engine was conducted through a rigorous computational framework applied to the COG-BCI dataset (N=29 subjects). This evaluation protocol was specifically designed to challenge the model’s capacity to generalize across orthogonal cognitive domains—transitioning from the controlled stability of working memory tasks to the stochastic volatility of complex piloting simulations. By integrating high-dimensional cortical dynamics (XEEG) with autonomic regulation (xECG), we aimed to test the hypothesis that multimodal fusion provides a topological anchor against distribution shifts. The analysis presented herein moves beyond standard accuracy metrics to provide a granular dissection of the system’s behavior, examining the statistical convergence of the ensemble, the topological separability of the feature manifold, and the biological plausibility of the learned decision boundaries. To ensure the reproducibility of these findings and define the precise stochastic configuration of the inference engine, the architectural hyperparameters and mathematical definitions utilized during the training phase are formalized in Table 3. The selection of an ensemble size of B=100 estimators was derived from a convergence analysis of the out-of-bag (OOB) error, ensuring minimal variance without incurring unnecessary computational latency for real-time applications [39]. Furthermore, the calibration parameter α=30% was established as the optimal trade-off point between operational feasibility (rapid setup) and sufficient manifold alignment to mitigate the covariate shift between source and target domains [25]. To explicitly characterize the data efficiency of the calibration process, we analyze the decoding performance as a function of the calibration ratio α. As shown in Figure 5, decoding performance increases rapidly as the calibration ratio α increases, exceeding 90% with only 5% target-domain calibration data. This pattern supports the data efficiency of the proposed manifold-alignment strategy, although peak performance is attained at α=30%. Experimental validation was conducted to quantify the extent to which the proposed architecture overcomes the Distribution Shift between the source domain (Memory Task) and the target domain (Piloting). As illustrated in Figure 6, the Zero-Shot baseline yielded a stochastic accuracy of μ≈38.91% (σ=12.4%). This catastrophic failure confirms that cortical patterns (PEEG) do not generalize naively across tasks due to non-stationary signal drift, resulting in a decision boundary that is orthogonal to the operational reality of the target domain [3]. To ensure the scientific reproducibility of the 99.13% accuracy benchmark and to formalize the stochastic architecture of the inference engine, a fixed set of hyperparameters and mathematical definitions was established. Table 3 delineates the specific computational configuration utilized during the training and validation phases, prioritizing the balance between variance reduction via bootstrap aggregating and model interpretability through impurity-based feature selection. In stark contrast, the Neuro-Cardiac Symbiotic Engine achieved a breakthrough global average accuracy of 99.13% with a tightly constrained standard deviation (σ=1.2%), as illustrated in Figure 6. This corresponds to a substantial improvement over the unimodal baseline and is supported by a 50-iteration Monte Carlo bootstrap evaluation that indicates stable convergence rather than isolated peak performance. The mathematical implication is twofold: first, the Few-Shot Calibration successfully aligned the marginal distributions of the feature space; and second, the inclusion of autonomic metrics (vHRV) acted as a “physiological anchor,” stabilizing the high-variance EEG features and preventing the classifier from overfitting to task-irrelevant cortical noise [8]. To ensure holistic reliability beyond simple accuracy, we analyzed the multi-metric profile visualized in the Radar Chart (Figure 7). A common pathology in BCI models is the “Accuracy Paradox,” where high accuracy masks poor performance in minority classes. However, our results demonstrate a symmetrical polygon, with Precision (>0.98), Recall (>0.99), and Specificity (>0.99) remaining consistently high across Low, Medium, and High workload states. This confirms that the system minimizes both Type I (False Alarms) and Type II (Missed Detections) errors, a mandatory requirement for safety-critical deployment. Finally, the Biomarker Stability Analysis (Figure 8) validates the “Variance Reduction Principle” of the Random Forest ensemble. Across 50 independent bootstrap iterations, the feature importance rankings remained invariant. The tightness of the interquartile ranges for the top predictors (e.g., Fp2, RMSSD) confirms that the model converges to a deterministic set of physiological drivers rather than fitting to stochastic artifacts, satisfying the stability conditions required for reproducible and robust neuroergonomics [20]. To substantiate this performance leap with granular statistical evidence, we conducted a comprehensive metric evaluation across the 50 stochastic iterations. As detailed in Table 4, the multimodal architecture exhibits superior convergence not only in global accuracy but across all sensitivity and specificity indices. The dramatic reduction in standard deviation (σ≈1.2% vs. σ≈12.4%) compared to the unimodal baseline confirms that the fusion strategy effectively minimizes the uncertainty inherent in single-trial decoding. To further examine robustness beyond aggregate performance metrics, we analyze subject-level variability and population-level reliability under a strict Leave-One-Subject-Out (LOSO) protocol designed to minimize data leakage and quantify generalization to unseen individuals. The subject-specific decoding accuracy is illustrated in Figure 9, while the population-level distribution and cumulative reliability are detailed in Figure 10 and Figure 11, respectively. The raincloud distribution explicitly captures the inter-subject variability often encountered in physiological computing. While the majority of participants cluster around a stable neuro-visceral manifold (forming the dense upper region of the distribution), a distinct subset of the population exhibits highly idiosyncratic autonomic and cortical reactions. This biological variance mathematically demonstrates why purely universal, zero-shot models frequently underperform in operational environments. It strongly reinforces the necessity of our proposed few-shot calibration approach, which effectively adapts the decision boundary to each user’s unique baseline state without requiring extensive retraining. Building upon this analysis of individual variability, we quantify the system’s overall operational viability across the entire cohort in Figure 11. A critical contribution of this study is the topological characterization of the functional coupling between cortical dynamics and autonomic regulation. Figure 12 visualizes this “Neuro-Cardiac Connectome,” providing explainable evidence for the model’s decision boundaries. From a graph-theoretical perspective, this visualization depicts a bipartite graph where the autonomic node functions as a central “hub” with high degree centrality, orchestrating information flow to cortical executive regions. This architecture aligns rigorously with the Neurovisceral Integration Model [4], suggesting that executive control centers are directly modulating cardiac output to meet metabolic demands during high-load tasks. Recent literature substantiates this finding, proposing that multimodal EEG-ECG indices are significantly more sensitive to cognitive stress variations than isolated EEG metrics. As demonstrated by Salam et al. (2026) [8], the fusion of these modalities provides a comprehensive view of the autonomic-cortical interplay, reaching an accuracy of up to 94.7%. Statistical Significance and Modality Contribution Analysis To assess whether the observed decoding performance could be attributed to random chance, we conducted a permutation-based statistical significance analysis. An empirical null distribution was constructed by randomly shuffling class labels across multiple iterations while preserving the temporal and structural properties of the signals. The resulting null distribution represents the performance expected under the hypothesis of no meaningful neurophysiological association between inputs and cognitive states. The accuracy achieved by the proposed Neuro-Cardiac Symbiotic Engine lies far outside this distribution, indicating that the observed performance cannot be explained by stochastic fluctuations alone. Figure 13 summarizes the permutation-based significance analysis, quantifying whether the observed performance exceeds chance expectations. Subsequently, Figure 14 presents the modality-ablation results to illustrate how much each physiological modality contributes to the final classifier. Beyond its theoretical and algorithmic contributions, the Neuro-Cardiac Symbiotic Engine carries direct implications for real-world cognitive monitoring systems. As shown in Figure 13, the observed accuracy lies well outside the empirical null distribution, indicating that the model captures a reproducible neurophysiological signal rather than chance structure. In safety-critical domains such as aviation and robotic teleoperation, prior research has shown that EEG-based systems alone remain vulnerable to non-stationarity [16]. As shown in Figure 14, the inclusion of autonomic information materially improves robustness relative to unimodal decoding. This pattern is particularly relevant for longitudinal deployment, where consistency is often more important than marginal gains in peak accuracy. The Feature Importance Topoplot (Figure 15) identifies the Fz (Frontal) and Pz (Parietal) electrodes as the primary cortical hubs for workload decoding. Mathematically, this indicates that the Random Forest algorithm successfully performed a “discriminative spatial filtering,” assigning maximal Gini importance weights to the Frontal-Parietal Attention Network (FPAN) while suppressing unrelated activity [22]. Notably, the model autonomously learned to attenuate visual cortex noise (Occipital), focusing instead on executive function markers. This selective activation pattern is consistent with recent insights from topological EEG analysis, which highlight that combining discriminative feature selection with geometric data representations effectively isolates task-relevant cortical dynamics while mitigating the impact of artifacts [21]. The SHAP summary plot (Figure 16) provides quantitative evidence for the physiological validity of the proposed model through the lens of Cooperative Game Theory. High Heart Rate values (red dots) combined with reduced RMSSD (blue dots) emerge as the strongest positive contributors to the High Stress class, indicating that the model is effectively capturing vagal withdrawal, a well-established marker of increased mental effort [5]. Importantly, the tight clustering of SHAP values associated with autonomic features reflects a low-variance contribution, in contrast to the broader dispersion observed in cortical features. This asymmetric contribution profile supports recent sensor-fusion findings describing a stabilizing mechanism termed metabolic anchoring, whereby autonomic dynamics constrain inference variability under transient cortical noise [40]. To elucidate the intrinsic geometry of the cognitive states, we projected the high-dimensional feature space (R323) onto lower-dimensional manifolds using Uniform Manifold Approximation and Projection (UMAP). As shown in Figure 17, the multimodal manifold exhibits distinct clusters corresponding to low, medium, and high workload. Although UMAP does not by itself establish strict topological separation, the limited overlap between clusters is consistent with the strong classification performance observed in the supervised analyses [41]. The 3D Physiological State Space reconstruction (Figure 18) reveals a deterministic “Stress Trajectory.” As cognitive demand increases, the subject’s state vector migrates from a high-entropy region (High Variability) to a constrained attractor state characterized by low variability and high heart rate. This trajectory provides a mathematical signature of cognitive overload, independent of individual differences [42]. Finally, to bridge the gap between the formal mathematical architecture defined in the methodology and the observed experimental outcomes, we synthesize the validation of our core theoretical postulates. Table 5 provides a rigorous mapping between the governing theorems of high-dimensional stochastic systems (Variance Reduction, Projection, and Entropy) and their specific empirical manifestations in our results, confirming that the system’s behavior is deterministic and mathematically grounded. To substantiate this performance leap with granular statistical evidence, we conducted a comprehensive metric evaluation across 50 stochastic iterations via Monte Carlo bootstrapping. As detailed in Table 4, the multimodal architecture exhibits superior convergence not only in global accuracy but across all sensitivity and specificity indices. Critically, the analysis of variance reveals a fundamental stabilization of the inference engine: the standard deviation reduced from σ≈12.4% in the unimodal baseline to a negligible σ≈1.2% in the fusion model. This mathematical convergence confirms that the inclusion of the autonomic vector effectively minimizes the aleatoric uncertainty inherent in single-trial decoding. Furthermore, the symmetry between Precision (99.20%) and Recall (99.10%) indicates that the decision boundary is isometric, successfully mitigating the bias often observed in imbalanced physiological datasets without requiring synthetic oversampling. Finally, the rapid convergence time (<120 s on HPC) validates the efficiency of the Affine Manifold Alignment compared to iterative Deep Learning methods, supporting the feasibility of real-time deployment. 4. Discussion The empirical results of this study extend beyond classification accuracy and provide insight into the physiological organization of cognitive effort. By integrating high-dimensional cortical dynamics with autonomic regulation, we evaluate the Neuro-Cardiac Symbiotic Engine as a framework that is both predictively robust (99.13% mean accuracy under the reported evaluation setting) and biologically interpretable. This section deconstructs the mechanisms enabling this performance, contextualizing our findings within the theoretical frameworks of Neurovisceral Integration [4] and recent multimodal fusion paradigms [26]. 4.1. The Vagal Brake: Physiological Validation of the Fusion Hypothesis A pivotal finding of this research is the quantification of the “Vagal Brake” mechanism as a reliable, low-variance proxy for Prefrontal Cortex (PFC) activity. As illustrated in the Violin Plots (Figure 19), the transition from “Low” to “High” workload triggers a statistically significant suppression of RMSSD (p<0.001), accompanied by a compensatory chronotropic elevation in Heart Rate. This inverse trajectory aligns rigorously with Porges’ Polyvagal Theory [5] and recent multimodal validations by Salam et al. (2026) [8], which demonstrate that integrating cardiac markers (such as heart rate and HRV) with cortical features significantly improves the sensitivity of cognitive stress detection. High-stakes piloting necessitates the disinhibition of the sinoatrial node—a process controlled top-down by the PFC—to mobilize metabolic resources. Crucially, the Autonomic Balance Correlation (Figure 20) demonstrates a robust linear coupling (r=0.98) between Global Variability (SDNN) and Vagal Tone (RMSSD). From a signal-processing perspective, this linearity suggests that the autonomic response evolves on a relatively low-dimensional manifold. Unlike the stochastic non-stationarity often observed in EEG due to electrode impedance drifts [44], this cardiac manifold provides a stable physiological ground truth. As noted by Lee et al. (2024) [40], integrating such stable autonomic anchors effectively acts as a regularizer for high-variance cortical features, explaining why our fusion model maintains stability against domain shifts where unimodal models fail. Beyond univariate statistics, the Autonomic Feature Space (Figure 21) visualizes the topological transformation of the subject’s state. The Kernel Density Estimation (KDE) contours reveal a clear phase transition consistent with complex systems theory [45]. This phenomenon, which we term Topological Contraction, signifies a reduction in system complexity and information production rates under stress [11]. To provide a rigorous theoretical basis for this observation, we model the cognitive state transition as a contraction of the phase space volume. This mechanism is governed by Pesin’s Formula, which relates the metric entropy (hμ) to the sum of positive Lyapunov exponents. Under high cognitive load, the uniform dampening of local expansion rates (κ<1) inevitably leads to a reduction in entropy (hstress