Time Series Processor

Processor for irregular time series data with missing values. Supports uniform resampling, two imputation strategies (forward-fill and zero-fill), and automatic z-score normalization using training data statistics.

Details

The processor performs three main steps:

1. Resampling: Converts irregular time series to uniform time grid based on sampling_rate

2. Imputation: Fills missing values using either forward-fill or zero-fill strategy

3. Normalization (default): Applies z-score normalization: (x - mean) / std

Normalization is enabled by default (normalize = TRUE):

• Call fit() on training samples to compute feature-wise mean and standard deviation

• The same statistics are used for all subsequent process() calls (train/val/test)

• This ensures no data leakage between training and validation/test sets

Super classes

RHealth::Processor -> RHealth::FeatureProcessor -> TimeseriesProcessor

Public fields

sampling_rate

A lubridate duration indicating the sampling step size.

impute_strategy

A character string: 'forward_fill' or 'zero'.

normalize

Logical flag indicating whether to apply z-score normalization. Default TRUE.

feature_means

Numeric vector of feature means (computed during fit).

feature_stds

Numeric vector of feature standard deviations (computed during fit).

.size

Number of features (set on first call to process()).

Methods

Public methods

• TimeseriesProcessor$new()

• TimeseriesProcessor$fit()

• TimeseriesProcessor$process()

• TimeseriesProcessor$size()

• TimeseriesProcessor$print()

• TimeseriesProcessor$clone()

Inherited methods

• RHealth::Processor$load()
• RHealth::Processor$save()

---

Method new()

Initialize the processor with a sampling rate, imputation strategy, and normalization option.

Usage

TimeseriesProcessor$new(
  sampling_rate = lubridate::dhours(1),
  impute_strategy = "forward_fill",
  normalize = TRUE
)

Arguments

sampling_rate

A lubridate duration (e.g., lubridate::dhours(1)).

impute_strategy

Either 'forward_fill' or 'zero'.

normalize

Logical, whether to apply z-score normalization. Default TRUE.

---

Method fit()

Fit the processor by computing feature-wise mean and std from training data.

Usage

TimeseriesProcessor$fit(samples, field)

Arguments

samples

A list of named lists representing sample records.

field

A string giving the field name to fit on.

---

Method process()

Process irregular time series into uniformly sampled tensor. Step 1: uniformly sample time points and place values at correct positions. Step 2: impute missing entries using selected strategy. Step 3: (optional) apply z-score normalization using training statistics.

Usage

TimeseriesProcessor$process(value)

Arguments

value

A list: list(timestamps = POSIXct vector, values = matrix).

Returns

A torch tensor of shape [T, F].

---

Method size()

Return the number of features.

Usage

TimeseriesProcessor$size()

Returns

Integer

---

Method print()

Print summary

Usage

TimeseriesProcessor$print(...)

Arguments

...

Ignored.

---

Method clone()

The objects of this class are cloneable with this method.

Usage

TimeseriesProcessor$clone(deep = FALSE)

Arguments

deep

Whether to make a deep clone.

Examples

if (FALSE) { # \dontrun{
library(torch)
library(lubridate)

# Create training samples with timeseries data
train_samples <- list(
  list(
    patient_id = 1,
    labs = list(
      timestamps = as.POSIXct(c("2020-01-01 00:00:00", "2020-01-01 02:00:00"), tz = "UTC"),
      values = matrix(c(100, 50, 150, 60), ncol = 2)
    )
  ),
  list(
    patient_id = 2,
    labs = list(
      timestamps = as.POSIXct(c("2020-01-01 00:00:00", "2020-01-01 01:00:00"), tz = "UTC"),
      values = matrix(c(120, 55, 180, 65), ncol = 2)
    )
  )
)

# Example 1: Default behavior (with normalization)
processor <- TimeseriesProcessor$new(
  sampling_rate = dhours(1),
  impute_strategy = "forward_fill"
  # normalize = TRUE by default
)

# Fit on training data to compute statistics
processor$fit(train_samples, "labs")

# Process samples (applies normalization)
result <- processor$process(train_samples[[1]]$labs)

# Check normalization statistics
print(processor$feature_means)  # Feature means from training data
print(processor$feature_stds)   # Feature standard deviations

# Example 2: Disable normalization if needed
processor_no_norm <- TimeseriesProcessor$new(
  sampling_rate = dhours(1),
  impute_strategy = "forward_fill",
  normalize = FALSE  # Explicitly disable normalization
)
result_no_norm <- processor_no_norm$process(train_samples[[1]]$labs)
} # }