Type: Package
Title: Clinical Reference Interval Estimation with Reference Interval Network (RINet)
Version: 0.1.0
Description: Predicts statistics of a reference distribution from a mixture of raw clinical measurements (healthy and pathological). Uses pretrained CNN models to estimate the mean, standard deviation, and reference fraction from 1D or 2D sample data. Methods are described in LeBien, Velev, and Roche-Lima (2026) "RINet: synthetic data training for indirect estimation of clinical reference distributions" <doi:10.1016/j.jbi.2026.104980>.
License: MIT + file LICENSE
Encoding: UTF-8
Depends: R (≥ 3.5.0)
Imports: reticulate
SystemRequirements: Python (>= 3.8), TensorFlow (>= 2.16), Keras (>= 3.0), scikit-learn
RoxygenNote: 7.3.3
NeedsCompilation: no
Packaged: 2026-01-26 07:48:36 UTC; jack
Author: Jack LeBien [aut, cre]
Maintainer: Jack LeBien <jackgl4124@gmail.com>
Repository: CRAN
Date/Publication: 2026-01-29 21:50:02 UTC

RINet: Predict Clinical Reference Intervals from Mixture Distributions

Description

Predict the statistics of an underlying reference distribution from a mixture distribution of raw clinical measurements (healthy and pathological patients). Uses pre-trained convolutional neural networks to estimate means, standard deviations, correlations, and reference fractions from 1D or 2D sample data.

Details

The main functions in this package are:

Models are automatically loaded on first use and cached for efficiency. Model files should be placed in inst/models/ with names:

Author(s)

Jack

Maintainer: Jack <your.email@example.com>

Convert correlation to covariance matrix

Description

Convert correlation to covariance matrix

Usage

.correlation_to_covariance(corr_matrix, std_vector)

Extract histogram features from standardized data

Description

Extract histogram features from standardized data

Usage

.extract_features(
  data,
  ndim,
  feature_grid_range = c(-4, 4),
  feature_grid_nbins = 100
)

Internal function to load model and scaler

Description

Internal function to load model and scaler

Usage

.load_model(ndim)

Internal function to load scaler

Description

Internal function to load scaler

Usage

.load_scaler(ndim)

Predict statistics of the underlying reference distribution from mixture distributions using RINet

Description

Automatically detects whether input data is 1D or 2D and calls the appropriate prediction function. This is the main user-facing function. It estimates the statistics of a "healthy" reference population from a mixture of healthy and pathological measurements.

Usage

predict_rinet(
  data,
  feature_grid_range = c(-4, 4),
  feature_grid_nbins = 100,
  verbose = 0,
  log_scale = TRUE,
  percentiles = c(0.025, 0.975),
  n_bootstrap = 0,
  confidence_level = 0.95
)

Arguments

data

A numeric vector, matrix, or list. For 1D: vector or matrix with 1 column. For 2D: matrix with 2 columns. Can also be a list of such objects.

feature_grid_range

Numeric vector of length 2 specifying the range for histogram binning. Default is c(-4, 4).

feature_grid_nbins

Integer specifying the number of histogram bins. Default is 100.

verbose

Integer controlling verbosity (0 = silent). Default is 0.

log_scale

Logical indicating whether to log-transform the data before prediction. If TRUE (default), returns log-scale statistics and calculates reference intervals in the original scale. Default is TRUE.

percentiles

Numeric vector of length 2 specifying the lower and upper percentiles for the reference interval. Default is c(0.025, 0.975).

n_bootstrap

Integer specifying the number of bootstrap resamples for confidence intervals. Default is 0 (no bootstrap). When > 0, confidence intervals are computed for all predicted statistics using batch inference.

confidence_level

Numeric specifying the confidence level for bootstrap intervals. Default is 0.95.

Value

A list of predictions. Each element contains:

mean

Predicted mean(s) (log-scale if log_scale=TRUE)

std

Predicted standard deviation(s) (log-scale if log_scale=TRUE)

covariance

Predicted covariance matrix

correlation

Predicted correlation (NA for 1D)

reference_fraction

Predicted reference component fraction

reference_interval

Reference interval in original scale (if log_scale=TRUE)

log_scale

Logical indicating whether log-scaling was used

bootstrap_ci

List of bootstrap confidence intervals (if n_bootstrap > 0)

Examples

## Not run: 
  # 1D sample (using positive data for log-scale)
  sample_1d <- exp(rnorm(1000, mean = 2, sd = 0.5))
  result <- predict_rinet(sample_1d)

  # 2D sample (using positive data for log-scale)
  sample_2d <- exp(matrix(rnorm(2000, mean = 2, sd = 0.5), ncol = 2))
  result <- predict_rinet(sample_2d)

  # Multiple samples (automatically detected)
  samples <- list(exp(rnorm(1000, mean = 2, sd = 0.5)),
                  exp(rnorm(1000, mean = 2, sd = 0.5)))
  results <- predict_rinet(samples)

## End(Not run)

Predict statistics of the underlying reference distribution from 1D mixture distributions using RINet

Description

Takes one or more 1D samples and predicts the underlying reference population statistics (mean, std, reference fraction) from a mixture of healthy and pathological measurements.

Usage

predict_rinet_1d(
  data,
  feature_grid_range = c(-4, 4),
  feature_grid_nbins = 100,
  verbose = 0,
  log_scale = TRUE,
  percentiles = c(0.025, 0.975),
  n_bootstrap = 0,
  confidence_level = 0.95
)

Arguments

data

A numeric vector, matrix, or list of vectors. Each sample should contain observations from a 1D mixture distribution.

feature_grid_range

Numeric vector of length 2 specifying the range for histogram binning. Default is c(-4, 4).

feature_grid_nbins

Integer specifying the number of histogram bins. Default is 100.

verbose

Integer controlling verbosity (0 = silent). Default is 0.

log_scale

Logical indicating whether to log-transform the data before prediction. If TRUE (default), returns log-scale statistics and calculates reference intervals in the original scale. Default is TRUE.

percentiles

Numeric vector of length 2 specifying the lower and upper percentiles for the reference interval. Default is c(0.025, 0.975).

n_bootstrap

Integer specifying the number of bootstrap resamples for confidence intervals. Default is 0 (no bootstrap). When > 0, confidence intervals are computed for all predicted statistics.

confidence_level

Numeric specifying the confidence level for bootstrap intervals. Default is 0.95.

Value

A list of predictions. Each element contains:

mean

Predicted mean (scalar, log-scale if log_scale=TRUE)

std

Predicted standard deviation (scalar, log-scale if log_scale=TRUE)

covariance

Covariance matrix (1x1 matrix)

correlation

Always NA for 1D

reference_fraction

Predicted reference component fraction

reference_interval

Reference interval in original scale (if log_scale=TRUE)

log_scale

Logical indicating whether log-scaling was used

bootstrap_ci

List of bootstrap confidence intervals (if n_bootstrap > 0): mean_ci, std_ci, reference_fraction_ci, reference_interval_lower_ci, reference_interval_upper_ci

Examples

## Not run: 
  # Single sample (using positive data for log-scale)
  sample1 <- exp(rnorm(1000, mean = 2, sd = 0.3))
  result <- predict_rinet_1d(sample1)
  print(result[[1]]$mean)

  # Multiple samples
  samples <- list(exp(rnorm(1000, 2, 0.3)), exp(rnorm(1000, 1.5, 0.4)))
  results <- predict_rinet_1d(samples)

## End(Not run)

Predict statistics of the underlying reference distribution from 2D mixture distributions using RINet

Description

Takes one or more 2D samples and predicts the underlying reference population statistics (means, stds, correlation, reference fraction) from a mixture of healthy and pathological measurements.

Usage

predict_rinet_2d(
  data,
  feature_grid_range = c(-4, 4),
  feature_grid_nbins = 100,
  verbose = 0,
  log_scale = TRUE,
  percentiles = c(0.025, 0.975),
  n_bootstrap = 0,
  confidence_level = 0.95
)

Arguments

data

A matrix or list of matrices. Each sample should be a matrix with 2 columns representing observations from a 2D mixture distribution.

feature_grid_range

Numeric vector of length 2 specifying the range for histogram binning. Default is c(-4, 4).

feature_grid_nbins

Integer specifying the number of histogram bins. Default is 100.

verbose

Integer controlling verbosity (0 = silent). Default is 0.

log_scale

Logical indicating whether to log-transform the data before prediction. If TRUE (default), returns log-scale statistics and calculates reference intervals in the original scale. Default is TRUE.

percentiles

Numeric vector of length 2 specifying the lower and upper percentiles for the reference interval. Default is c(0.025, 0.975).

n_bootstrap

Integer specifying the number of bootstrap resamples for confidence intervals. Default is 0 (no bootstrap). When > 0, confidence intervals are computed for all predicted statistics.

confidence_level

Numeric specifying the confidence level for bootstrap intervals. Default is 0.95.

Value

A list of predictions. Each element contains:

mean

Predicted means (vector of length 2, log-scale if log_scale=TRUE)

std

Predicted standard deviations (vector of length 2, log-scale if log_scale=TRUE)

covariance

Predicted covariance matrix (2x2 matrix)

correlation

Predicted correlation coefficient (scalar)

reference_fraction

Predicted reference component fraction

reference_interval

Reference region ellipse vertices (100x2 matrix) in original scale (if log_scale=TRUE)

log_scale

Logical indicating whether log-scaling was used

bootstrap_ci

List of bootstrap confidence intervals (if n_bootstrap > 0): mean_ci (2x2 matrix), std_ci (2x2 matrix), correlation_ci, reference_fraction_ci

Examples

## Not run: 
  # Single 2D sample (using positive data for log-scale)
  sample1 <- exp(matrix(rnorm(2000, mean = 2, sd = 0.3), ncol = 2))
  result <- predict_rinet_2d(sample1)
  print(result[[1]]$mean)
  print(result[[1]]$covariance)

  # Multiple samples
  samples <- list(exp(matrix(rnorm(2000, mean = 2, sd = 0.3), ncol = 2)),
                  exp(matrix(rnorm(2000, mean = 2, sd = 0.3), ncol = 2)))
  results <- predict_rinet_2d(samples)

## End(Not run)

mirror server hosted at Truenetwork, Russian Federation.