Table 3 Key methodological instantiations for the metamodel
From: Computational whole-body-exposome models for global precision brain health
A. Data heterogeneity | ||||||
|---|---|---|---|---|---|---|
Technique (example) | Required data | Complexity | Interpretability | Computational cost | Multimodal integration | Robustness |
Large-Scale deep learning (accelerated brain age in large cohorts)10 | Large annotated and balanced datasets required | High complexity | Limited and indirect | High computational demands for model training | Integration possible, but not directly built in the model | High, when trained with sufficient data and/or data augmentation |
Normative modeling (population reference distributions)98 | Requires large samples with good reference data coverage across diversity axes | Variable, depends on the specific model | Supports subject-level interpretation | Computationally expensive for high-dimensional data | Variable, depends on the specific model | Inherently models uncertainty and robust to heterogeneities |
Federated learning (distributed training to keep privacy)282 | Distributed data requirements | Variable, depends on the specific model | Variable, depends on the specific model | Convergence issues with non-independent and identically distributed data, limited by the least resourced sites | Variable, depends on the specific model | Enhances representation of diverse populations, mitigates site-specific biases, evolves as new data is available |
B. Multimodal integration (brain + extracerebral metrics) | ||||||
|---|---|---|---|---|---|---|
Technique (example) | Required data | Complexity | Interpretability | Computational cost | Multimodal integration | Robustness |
Whole-brain biophysical modeling (perturbations and extracerebral modulation)283 | Low, allows single-subject modeling | Variable, but limited by high computational cost | Both mechanistic and causal interpretability | High computational demands for model training | Can integrate systemic and environmental modulators via priors or order parameters | Sensitive to assumptions about coupling between neural regions |
Dynamic causal modeling (DCM, neuropharmacological dynamics)284 | Low, allows single-subject modeling | Variable, but limited by high computational cost | Causal interpretability of directed brain connectivity, rigorous Bayesian model comparison, physiologically meaningful parameters | High computational demands for parameter optimization | Supports multimodal extensions | Requires strong a priori model specification, sensitive to model specification details |
Active inference (Generative brain-body models)285 | Low, allows single-subject modeling | Variable, but limited by high computational cost | Limited empirical application, abstract constructs require biological mapping. | High computational demands for parameter optimization | Unifies perception, action, and regulation | Explicit uncertainty modeling but difficult model specification |
Deep multimodal learning (joint representation models)286 | Requires large multimodal datasets | High complexity, can approximate arbitrary non-linear decision functions | Limited and indirect | High computational demands for model training | Captures non-linear cross-modal interactions | Prone to modality dominance, limited native uncertainty estimation |
C. Individual level trajectories | ||||||
|---|---|---|---|---|---|---|
Technique (example) | Required data | Complexity | Interpretability | Computational cost | Multimodal integration | Robustness |
Bayesian sequential Inference (state-space models)287 | High for high-dimensional or poorly sampled systems | High model complexity | Interpretability in terms of personalized trajectories | High computational demands for high-dimensional systems | Integration possible, but not directly built in the model | Adaptive updating with new observations, explicit uncertainty quantification, handles irregular sampling and noise |
Markov chain models (e.g., Markov & hidden Markov models for spatiotemporal brain dynamics)288 | Large amounts of data needed for rare transitions | Low complexity, memoryless assumption can oversimplify dynamics, discretization may lose information | Simple interpretability, analytically tractable, clear clinical stage modeling | Low computational demands | Integration possible, but not directly built in the model | Sensitive to model initialization |
