Bi-Hemispheric Discrepancy Model (BiHDM)

• BiHDM is a computational model that quantifies asymmetries between paired hemispheric systems across diverse scientific fields.
• It employs methods such as Siamese networks, coupled MLPs, and discrepancy embeddings to analyze bilateral data for robust anomaly detection.
• The model enhances precision in applications like neuroimaging, robotics, and geophysics by unifying asymmetry metrics and correcting observational biases.

The Bi-Hemispheric Discrepancy Model (BiHDM) refers to a class of computational and mathematical models that explicitly represent, learn, or explain functional, structural, or dynamical asymmetries between two hemispheres in bilaterally organized systems. BiHDM frameworks appear across diverse domains—including neuroscience, geophysics, signal processing, robotics, and artificial intelligence—where left/right (or north/south, east/west) differences encode essential system-level properties. These models typically exploit symmetric or homologous structure, use paired representations or inter-hemispheric coupling mechanisms, and quantify deviation, anomaly, or specialization associated with hemispheric discrepancies. The following sections survey principal mathematical formulations, methodological approaches, empirical results, and their scientific and technological significance.

1. Core Principles and Mathematical Formulations

At the core of BiHDM is the explicit modeling of two parallel streams—one per hemisphere—and their comparative analysis through discrepancy metrics, coupling parameters, or embedded representations. The model is instantiated as:

• Difference or discrepancy embedding: For homologous brain regions, BiHDM computes an asymmetry embedding via a shared encoder and a subtraction operation between left and right features, projecting the result into an embedding space where deviation from a learned "normal" center quantifies abnormality (Deangeli et al., 2023).
• Blockwise coupling matrices: Large-scale dynamical models (e.g., Wilson–Cowan neural mass, FitzHugh–Nagumo oscillators) employ intra-hemispheric and inter-hemispheric coupling terms, parameterized by distinct gains or control parameters $G_1$ (intra) and $G_2$ (inter), enabling selective adjustment and probing of hemispheric integration (Plüss et al., 28 Jun 2025, Ramlow et al., 2019).
• Two-channel correlation coefficients: In binaural signal processing, BiHDM is formalized via complex-valued correlation coefficients between analytic signals in each hemisphere, encoding both mean and fluctuation variance of interaural differences (Encke et al., 2021).
• Sensitivity-corrected indices: In geomagnetic studies, BiHDM corrects for hemispheric bias in global indices by modeling location- and context-dependent sensitivity, yielding homogeneous indices that reflect true underlying physical variations (Lockwood et al., 2018).

The table below summarizes selected mathematical forms:

Domain	BiHDM Core Operation	Primary Discrepancy Metric
Neuroimaging	$z = g_\phi(f_{\theta}(x_L) - f_{\theta}(x_R))$	$D(x) = \|z(x) - w\|_2$
Neural Dynamics	$$\mathcal{I}_k = G_1(\mathrm{intra}) + G_2(\mathrm{inter})$$	SC–FC correlation gain tuning
Binaural Codes	$$\gamma = \langle l_a^* r_a \rangle / \sqrt{\langle |l_a|^2 \rangle \langle |r_a|^2 \rangle}$$	$|\gamma|$, $\arg(\gamma)$
Geomagnetism	$S(\theta, \phi)$ sensitivity scaling	Cor($aa_{HN}, aa_{HS}$), equinoctial recovery

2. Model Architectures and Methodological Strategies

BiHDM architectures are structured to preserve and explicitly utilize bilateral input representations, enforcing either shared encodings (to restrict comparison to homologous properties) or specialized loss functions (to promote complementary feature development):

• Siamese Networks: Used for anatomical anomaly detection, these networks have weight-shared encoders for symmetric structures, merging features via subtraction and embedding into an asymmetry space optimized with one-class objectives (Deangeli et al., 2023).
• Hemispheric RNN Traversals and Discrepancy Subnets: For high-dimensional spatial signals such as EEG, directional RNNs process left and right hemispheres separately (along e.g. horizontal and vertical axes), followed by pairwise discrepancy calculation and high-level fusion, further combined with adversarial domain discrimination for domain-invariant emotion recognition (Li et al., 2019).
• Bilateral CNNs with Specialization: In deep vision models, left and right hemispheric networks are specialized for local and global features via distinct loss functions, with late fusion by an attention-like linear classifier (Rajagopalan et al., 2022).
• Coupled MLPs for Motor Control: In robotics, left/right MLPs are trained with different coordination vs. stability objectives, and their outputs merged with a tunable weighting, sometimes augmented by limited cross-talk mimicking biological corpus callosum function (Rinaldo et al., 2024).

Parameter estimation approaches range from grid search in low-dimensional coupling space (e.g., $G_1,G_2$ in neural mass models) to deep unsupervised learning on control samples for hypersphere embedding (neuroimaging) or supervised end-to-end training with task-specific loss mixtures (AI/robotics).

3. Applications Across Scientific Domains

BiHDM has been developed and operationalized in multiple research spheres:

• Clinical Neuroimaging: Quantifies deviations in anatomical asymmetry (e.g., hippocampal shapes) to detect pathology such as Alzheimer's disease or hippocampal sclerosis, with disease-agnostic unsupervised training on healthy controls (Deangeli et al., 2023).
• Resting-State Functional Connectivity: Improves the fit between simulated and empirical brain FC by separately parameterizing intra- and inter-hemispheric long-range coupling, providing enhanced modeling of both healthy and pathological brain states (e.g., schizophrenia) (Plüss et al., 28 Jun 2025).
• EEG Emotion Recognition: Explicitly models spatial-temporal asymmetry in brain electrical activity, achieving state-of-the-art accuracy on multiple emotion recognition datasets (Li et al., 2019).
• Signal Processing (Auditory Neuroscience): Explains human binaural unmasking via a two-channel (hemispheric) model correlated with both physiological and behavioral data, outperforming single-channel models in predicting detection thresholds (Encke et al., 2021).
• Geophysics: Attributes seismic travel-time anomalies at Earth's inner core (ATIC) to a physical displacement of the solid core, rather than to hemispheric variation in material properties, thus unifying several asymmetries under the decentered-sphere hypothesis (Vamos et al., 2011).
• Geomagnetic Index Modeling: Deconvolves observational biases between hemispheres, producing a homogeneous index that recovers both secular trends and "equinoctial" UT–seasonal modulation (Lockwood et al., 2018).
• Neural Control for Robotics: Creates bilateral controllers that recapitulate human-like dominant and non-dominant hand specialization, enhancing motor performance in both reach and stability tasks (Rinaldo et al., 2024).

4. Quantitative Performance and Empirical Validation

BiHDM frameworks have consistently produced improvements over baseline or conventional unilateral models. Notable empirical results include:

• Neuroimaging anomaly detection: AUC values up to 1.00 for Alzheimer's disease and hippocampal sclerosis, and robust separation between healthy and pathological asymmetry distributions with $p < 0.01$ (Deangeli et al., 2023).
• Functional connectivity modeling: Pearson correlation increases of $+0.09$ to $+0.13$ for simulated-empirical FC correspondence by introducing hemispheric-specific coupling (Plüss et al., 28 Jun 2025).
• EEG emotion recognition: Consistent state-of-the-art accuracy on SEED/SEED-IV/MPED datasets, with BiHDM outperforming architectures such as domain-adversarial nets by several percentage points and achieving statistical significance on paired $t$-tests ($p < 0.05$) (Li et al., 2019).
• Binaural signal detection: 98% variance explained across eight classic psychoacoustic experiments, with fits exceeding $R^2=0.97$ for individual paradigms and $R^2=0.93$ with a global parameter set (Encke et al., 2021).
• Robotic control: Bilaterally specialized models (BiHDM, CC-S) significantly outperformed single-hemisphere and non-specialized controls in both speed-to-goal and “time in goal” metrics (ANOVA, $p < 0.01$) (Rinaldo et al., 2024).

5. Theoretical and Mechanistic Insights

BiHDM models unify seemingly disparate forms of hemispheric asymmetry across scales and systems:

• Neural Basis of Discrepancy: Both electrophysiological and anatomical evidence in mammals demonstrates that callosal (inter-hemispheric) circuits are not merely inhibitory but act as contrast-enhancing, center–surround filters, promoting integration and sharpening representations (e.g., surround inhibition in motor, sensory, and auditory cortex) (Carson, 2020).
• Dynamical Regimes: By adjusting inter-hemispheric coupling parameters, BiHDM predicts transitions between global synchronization, partial (unihemispheric) synchronization, and asymmetric chimeric states, illuminating mechanisms for phenomena such as unihemispheric sleep or task-dependent vigilance (Ramlow et al., 2019).
• Mechanistic Unification: In seismology, recasting hemispheric asymmetry as a geometric shift rather than compositional variation provides a parsimonious explanation for multiple observed anomalies, implicating mechanical, magnetic, and thermal couplings at planetary scale (Vamos et al., 2011).

6. Scientific and Technological Implications

The explicit treatment of bilateral discrepancies confers several advantages:

• Disease-agnostic Detection: By learning the normal envelope of hemispheric asymmetry, BiHDM detects arbitrary deviations, accommodating normal variation across lifespan and unifying heterogeneous pathologies under a single anomaly score (Deangeli et al., 2023).
• Biologically Inspired AI: Bilateral architectures with enforced functional specialization and controlled inter-stream fusion represent a principled inductive bias for machine learning, outperforming unspecialized bilaterals in hierarchical tasks (Rajagopalan et al., 2022). Loss customization per hemisphere enables modular controllers in robotics that adaptively route function for efficiency or stability (Rinaldo et al., 2024).
• Quantitative Correction of Observational Bias: In physical systems (geomagnetism, seismology), BiHDM yields statistically homogeneous indices and physically principled explanations by compensating for geometry- and context-dependent discrepancies (Lockwood et al., 2018, Vamos et al., 2011).
• Neurorehabilitation and Clinical Practice: Mechanistic reinterpretation of hemispheric imbalance after injury supports targeted, bilateral network-based interventions rather than simplistic suppression of presumed pathological inter-hemispheric inhibition (Carson, 2020).

7. Limitations and Future Directions

Current BiHDM instantiations are limited by certain factors:

• Computational Cost: Bilateral or Siamese architectures effectively double model parameters compared to unilaterals; late fusion only partially mitigates this overhead (Rajagopalan et al., 2022).
• Explicit Discrepancy Utilization: Some models measure but do not yet exploit discrepancy magnitude (e.g., for adaptive gating) at multiple scales or depths (Rajagopalan et al., 2022).
• Interconnectivity Constraints: Many implementations fuse features only at the penultimate layer; incorporation of inter-hemispheric skip connections or dynamic cross-talk may further enhance expressivity (Rinaldo et al., 2024).
• Generalization Beyond Bilateral Systems: While the core BiHDM concept is bilateral, its extension to multi-compartment or hierarchically structured systems remains to be fully explored.

A plausible implication is that ongoing work will focus on deeper integration of dynamic discrepancy metrics, learned task-adaptive fusion, biologically grounded connectivity patterns, and domain transfer to non-neural and multilateral systems. Empirical evaluation across broader task suites and integration with neurophysiological data and physical measurements will further refine and expand the applicability of BiHDM frameworks.