Skin-conductance responses¶

BatemanSCR is the built-in response family for skin-conductance models. The same Response configuration works in R-MSRM, dense/grouped GP MAP and posterior inference, and state-space GP MAP. BachSCR was removed in 0.2.0; see the migration instructions below for historical analyses.

BatemanSCR is available from release 0.2.0. See the installation guide.

Configure a response¶

from multimodalsrm import BatemanSCR, Identity, Response

responses = {
    "reference": Response(Identity(), estimate=False, pooling="shared"),
    "scr": Response(
        BatemanSCR(rise=0.7, decay=3.0, lag=0.0),
        pooling="shared",
        bounds={"rise": (0.3, 1.4), "decay": (1.5, 4.0), "lag": (-1.0, 1.0)},
    ),
}

rise and decay are positive time constants in seconds, and lag shifts the onset. They do not directly specify the time of the peak. Swapping the two time constants leaves the response unchanged; separated bounds identify their labels when both are learned. Equal constants are supported numerically and reduce to a shape-two Erlang response.

The untruncated unit-mass transfer function is 1 / ((1 + s*rise) * (1 + s*decay)). This is the difference-of-exponentials SCR model used by Benedek and Kaernbach (2010). The toolbox uses a support of 90 seconds after lag, with continuous finite L2 normalization. The initial constants are starting values, not validated physiological defaults or a fitted conversion of canonical Bach.

Use with R and GP models¶

Pass responses above to MultimodalSRM(..., responses=responses) for R-MSRM. Cell integration is vectorized and includes complex-step derivatives of the time constants, lag, finite normalizer and moving support endpoints. This requires no optional Bayesian dependencies.

GP models additionally require priors for free response parameters:

import numpy as np
from multimodalsrm.bayesian import BayesianMultimodalSRM, BayesianPriors, Prior

model = BayesianMultimodalSRM(
    features=1,
    responses=responses,
    reference_modality="reference",
    priors=BayesianPriors(
        noise=Prior.lognormal(np.log(0.02), 0.7),  # observation variance
        filters={
            "scr": {
                "rise": Prior.lognormal(np.log(0.7), 0.35),
                "decay": Prior.lognormal(np.log(3.0), 0.35),
                "lag": Prior.normal(0.0, 0.5),
            }
        },
    ),
    length_scale=3.0,
    inference="map",
    linear_algebra="grouped",
    response_quadrature_order=64,
)

Dense/grouped MAP and posterior inference require an explicit response_quadrature_order. Compare orders over the chosen parameter bounds: the analytic covariance_tolerance does not certify quadrature accuracy. Other posterior requirements, including shared response pooling and factor anchors, still apply.

State-space MAP uses two exact exponential response states, plus two Matérn states per factor. It can learn rise, decay and lag. A finite cutoff cannot be represented by this finite-state filter: state-space inference restores the tail beyond 90 seconds and checks a uniform covariance error bound over the full parameter box. Slow decay or broad bounds may fail that check; use dense/grouped integration or explicitly justify a larger tolerance. No tolerance is relaxed automatically. Mixed Bateman/Gaussian models require state_space_gaussian="auto" or "rational". State-space posterior sampling and Bateman spectral inference are unsupported.

The SCR example uses one response configuration for R and GP fitting:

python examples/scr_quickstart.py
JAX_ENABLE_X64=true python examples/scr_quickstart.py --gp grouped
JAX_ENABLE_X64=true python examples/scr_quickstart.py --gp state_space

The observations are synthetic. The example checks fitting, valid prediction and archive replay; it does not establish physiological recovery or uncertainty calibration.

Migrating a Bach analysis¶

Replace Response(BachSCR(...), ...) with an explicit Response(BatemanSCR(...), ...), choose time-constant bounds and GP priors, and refit the model. There is no automatic conversion of Bach parameters, fitted loadings or lag estimates. Their reference shapes differ, so a difference in fitted lag alone cannot establish a change in physiological timing.

Release 0.2.0 removes BachSCR from public imports, R/GP fitting, response approximations and the archive type registry. Loading a GP archive containing a Bach record raises an explicit migration error before reconstructing or fitting a model. Historical R pickle/joblib models also require their original environment. No Bach archive is silently substituted with Bateman.

Release 0.1.0 supports Bach. The last development snapshot before this removal is commit 8eeb84fcc8d2c6a37f36ce748a13e6c3782e0ce8. For exact reproduction, retain the actual source revision, dependencies and preprocessing used for the original fit; an arbitrary older release does not replace that provenance. Warmup checkpoints must resume in their original source/runtime environment. If original observations are unavailable, use that environment for prediction from the Bach fit.

Archives using the remaining supported response families retain their existing identifiers and schema. The generic model default remains Identity when no responses are supplied; the toolbox does not infer which streams contain skin conductance.

Numerical tests cover independent convolution integrals, finite normalization, equal and near-equal poles, cell-integral derivatives, rational transfer functions, likelihood/gradient/prediction agreement, posterior execution and archive replay. Before changing an established scientific analysis, compare held-out prediction, inferred latents and response timing on representative recordings. The initial synthetic comparison supported Bateman as the simpler baseline; neither matching Bach's shape nor a synthetic fit alone establishes empirical superiority.