step_measure_peaks_deconvolve() creates a specification of a recipe step
that resolves overlapping peaks using curve fitting. This step requires
peaks to have been detected first using step_measure_peaks_detect().
Usage
step_measure_peaks_deconvolve(
recipe,
model = "gaussian",
optimizer = "auto",
max_iter = 500L,
tol = 1e-06,
n_starts = 5L,
constrain_positions = TRUE,
quality_threshold = 0.8,
store_components = FALSE,
smart_init = TRUE,
peaks_col = ".peaks",
measures_col = ".measures",
role = NA,
trained = FALSE,
skip = FALSE,
id = recipes::rand_id("measure_peaks_deconvolve")
)Arguments
- recipe
A recipe object.
- model
Peak model to use. Either a character string naming a registered model (
"gaussian","emg","bigaussian","lorentzian"), or apeak_modelobject created directly. Usepeak_models()to see all registered models. Default is"gaussian".- optimizer
Optimization method:
"auto"(default),"lbfgsb","multistart", or"nelder_mead". Auto-selection chooses based on problem complexity and signal-to-noise ratio.- max_iter
Maximum iterations for optimization. Default is 500. Tune this with
peak_fit_iterations().- tol
Convergence tolerance. Default is 1e-6.
- n_starts
Number of random starts for
optimizer = "multistart". Default is 5.- constrain_positions
Logical. If
TRUE, enforce that peak centers maintain their relative ordering. Default isTRUE.- quality_threshold
Minimum R-squared to accept fit. Fits below this threshold trigger a warning. Default is 0.8. Tune this with
peak_fit_r_squared().- store_components
Logical. If
TRUE, also store sampled fitted peak curves for diagnostics. Fitted models and parameters are always retained for subsequent component integration. Default isFALSE.- smart_init
Logical. If
TRUE, use smart initialization based on peak properties. Default isTRUE.- peaks_col
Name of the peaks column. Default is ".peaks".
- measures_col
Name of the measures column. Default is ".measures".
- role
Not used.
- trained
Logical indicating if the step has been trained.
- skip
Logical. Should the step be skipped when baking?
- id
Unique step identifier.
Value
An updated recipe with the new step added. The .peaks column
will be updated with deconvolved peak parameters, fitted areas, and
quality metrics.
Details
Peak deconvolution fits mathematical models to overlapping peaks to estimate their individual contributions. The result depends on the selected peak model, starting values and fit constraints.
Peak Models:
Built-in models (use peak_models() to see all):
"gaussian": Symmetric Gaussian (3 params: height, center, width)"emg": Exponentially Modified Gaussian (4 params, handles tailing)"bigaussian": Bi-Gaussian (4 params, flexible asymmetry)"lorentzian": Lorentzian/Cauchy peak (3 params, heavier tails)
Technique packs may register additional models.
Optimizers:
"auto": Selects based on problem complexity and SNR"lbfgsb": L-BFGS-B (fast, local optimization)"multistart": Multiple L-BFGS-B runs from perturbed starting values"nelder_mead": Derivative-free Nelder-Mead simplex
Reported Peak Properties:
For EMG fits, location and height report the apex of the fitted
component within the measured axis range. They use the continuous fitted
curve, including on irregular axes, and differ from the precursor Gaussian
parameters returned by optimize_deconvolution().
Quality Assessment:
Each fit is assessed for quality. The .peaks tibble gains columns:
fit_r_squared: R-squared of the overall fitfit_quality: Quality grade (A/B/C/D/F)purity: How much of signal at peak max comes from this peak
max_iter and quality_threshold are tunable. The R-squared range is
bounded by 0 and 1; choose iteration ranges that fit the model and runtime
budget for your data.
Before finalization, tidy() retains unresolved tuning calls in list columns.
The columns fit_model and fit_params retain each fitted component. The
detected_left_base and detected_right_base columns retain original detected
minima. The area_basis column identifies fitted-component areas.
left_base and right_base define a finite integration interval within the
measured axis. The interval retains detected minima where possible and
expands as needed to contain the reported fitted apex. Fitted areas integrate
each component over this interval without adding a physical baseline.
Examples
library(recipes)
# Create synthetic data with overlapping peaks
set.seed(42)
x <- seq(0, 20, by = 0.1)
y <- 1.5 * exp(-0.5 * ((x - 8) / 1)^2) +
0.8 * exp(-0.5 * ((x - 12) / 1.5)^2) +
rnorm(length(x), sd = 0.02)
df <- data.frame(id = "sample1", location = x, value = y)
# \donttest{
# Deconvolve overlapping peaks
rec <- recipe(~., data = df) |>
update_role(id, new_role = "id") |>
step_measure_input_long(value, location = vars(location)) |>
step_measure_peaks_detect(min_height = 0.5, min_prominence = 0.3) |>
step_measure_peaks_deconvolve(model = "gaussian") |>
prep()
result <- bake(rec, new_data = NULL)
# Check fitted peaks
result$.peaks[[1]]
#> # A tibble: 2 × 17
#> peak_id location height left_base right_base area prominence_left_base
#> <int> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
#> 1 1 8.00 1.50 1.7 10.2 3.72 1.7
#> 2 2 12.0 0.792 10.2 20 2.64 10.2
#> # ℹ 10 more variables: prominence_right_base <dbl>, detected_left_base <dbl>,
#> # detected_right_base <dbl>, fit_model <list>, fit_params <list>,
#> # area_basis <chr>, fit_r_squared <dbl>, fit_quality <chr>, purity <dbl>,
#> # overlap_degree <dbl>
# }