P281424/AMR
1
0
Fork 0
mirror of https://github.com/msberends/AMR.git synced 2025-07-01 19:18:35 +02:00
Code Releases Activity
AMR/vignettes/AMR_with_tidymodels.Rmd

524 lines
17 KiB
Plaintext
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

---
title: "AMR with tidymodels"
output:
rmarkdown::html_vignette:
toc: true
toc_depth: 3
vignette: >
%\VignetteIndexEntry{AMR with tidymodels}
%\VignetteEncoding{UTF-8}
%\VignetteEngine{knitr::rmarkdown}
editor_options:
chunk_output_type: console
---
```{r setup, include = FALSE, results = 'markup'}
knitr::opts_chunk$set(
warning = FALSE,
collapse = TRUE,
comment = "#>",
fig.width = 7.5,
fig.height = 5
)
```
> This page was entirely written by our [AMR for R Assistant](https://chat.amr-for-r.org), a ChatGPT manually-trained model able to answer any question about the `AMR` package.
Antimicrobial resistance (AMR) is a global health crisis, and understanding resistance patterns is crucial for managing effective treatments. The `AMR` R package provides robust tools for analysing AMR data, including convenient antimicrobial selector functions like `aminoglycosides()` and `betalactams()`.
In this post, we will explore how to use the `tidymodels` framework to predict resistance patterns in the `example_isolates` dataset in two examples.
This post contains the following examples:
1. Using Antimicrobial Selectors
2. Predicting ESBL Presence Using Raw MICs
3. Predicting AMR Over Time
## Example 1: Using Antimicrobial Selectors
By leveraging the power of `tidymodels` and the `AMR` package, we’ll build a reproducible machine learning workflow to predict the Gramstain of the microorganism to two important antibiotic classes: aminoglycosides and beta-lactams.
### **Objective**
Our goal is to build a predictive model using the `tidymodels` framework to determine the Gramstain of the microorganism based on microbial data. We will:
1. Preprocess data using the selector functions `aminoglycosides()` and `betalactams()`.
2. Define a logistic regression model for prediction.
3. Use a structured `tidymodels` workflow to preprocess, train, and evaluate the model.
### **Data Preparation**
We begin by loading the required libraries and preparing the `example_isolates` dataset from the `AMR` package.
```{r lib packages, message = FALSE, warning = FALSE, results = 'asis'}
# Load required libraries
library(AMR) # For AMR data analysis
library(tidymodels) # For machine learning workflows, and data manipulation (dplyr, tidyr, ...)
```
Prepare the data:
```{r}
# Your data could look like this:
example_isolates
# Select relevant columns for prediction
data <- example_isolates %>%
# select AB results dynamically
select(mo, aminoglycosides(), betalactams()) %>%
# replace NAs with NI (not-interpretable)
mutate(across(where(is.sir),
~replace_na(.x, "NI")),
# make factors of SIR columns
across(where(is.sir),
as.integer),
# get Gramstain of microorganisms
mo = as.factor(mo_gramstain(mo))) %>%
# drop NAs - the ones without a Gramstain (fungi, etc.)
drop_na()
```
**Explanation:**
- `aminoglycosides()` and `betalactams()` dynamically select columns for antimicrobials in these classes.
- `drop_na()` ensures the model receives complete cases for training.
### **Defining the Workflow**
We now define the `tidymodels` workflow, which consists of three steps: preprocessing, model specification, and fitting.
#### 1. Preprocessing with a Recipe
We create a recipe to preprocess the data for modelling.
```{r}
# Define the recipe for data preprocessing
resistance_recipe <- recipe(mo ~ ., data = data) %>%
step_corr(c(aminoglycosides(), betalactams()), threshold = 0.9)
resistance_recipe
```
For a recipe that includes at least one preprocessing operation, like we have with `step_corr()`, the necessary parameters can be estimated from a training set using `prep()`:
```{r}
prep(resistance_recipe)
```
**Explanation:**
- `recipe(mo ~ ., data = data)` will take the `mo` column as outcome and all other columns as predictors.
- `step_corr()` removes predictors (i.e., antibiotic columns) that have a higher correlation than 90%.
Notice how the recipe contains just the antimicrobial selector functions - no need to define the columns specifically. In the preparation (retrieved with `prep()`) we can see that the columns or variables `r paste0("'", suppressMessages(prep(resistance_recipe))$steps[[1]]$removals, "'", collapse = " and ")` were removed as they correlate too much with existing, other variables.
#### 2. Specifying the Model
We define a logistic regression model since resistance prediction is a binary classification task.
```{r}
# Specify a logistic regression model
logistic_model <- logistic_reg() %>%
set_engine("glm") # Use the Generalised Linear Model engine
logistic_model
```
**Explanation:**
- `logistic_reg()` sets up a logistic regression model.
- `set_engine("glm")` specifies the use of R's built-in GLM engine.
#### 3. Building the Workflow
We bundle the recipe and model together into a `workflow`, which organises the entire modelling process.
```{r}
# Combine the recipe and model into a workflow
resistance_workflow <- workflow() %>%
add_recipe(resistance_recipe) %>% # Add the preprocessing recipe
add_model(logistic_model) # Add the logistic regression model
resistance_workflow
```
### **Training and Evaluating the Model**
To train the model, we split the data into training and testing sets. Then, we fit the workflow on the training set and evaluate its performance.
```{r}
# Split data into training and testing sets
set.seed(123) # For reproducibility
data_split <- initial_split(data, prop = 0.8) # 80% training, 20% testing
training_data <- training(data_split) # Training set
testing_data <- testing(data_split) # Testing set
# Fit the workflow to the training data
fitted_workflow <- resistance_workflow %>%
fit(training_data) # Train the model
```
**Explanation:**
- `initial_split()` splits the data into training and testing sets.
- `fit()` trains the workflow on the training set.
Notice how in `fit()`, the antimicrobial selector functions are internally called again. For training, these functions are called since they are stored in the recipe.
Next, we evaluate the model on the testing data.
```{r}
# Make predictions on the testing set
predictions <- fitted_workflow %>%
predict(testing_data) # Generate predictions
probabilities <- fitted_workflow %>%
predict(testing_data, type = "prob") # Generate probabilities
predictions <- predictions %>%
bind_cols(probabilities) %>%
bind_cols(testing_data) # Combine with true labels
predictions
# Evaluate model performance
metrics <- predictions %>%
metrics(truth = mo, estimate = .pred_class) # Calculate performance metrics
metrics
# To assess some other model properties, you can make our own `metrics()` function
our_metrics <- metric_set(accuracy, kap, ppv, npv) # add Positive Predictive Value and Negative Predictive Value
metrics2 <- predictions %>%
our_metrics(truth = mo, estimate = .pred_class) # run again on our `our_metrics()` function
metrics2
```
**Explanation:**
- `predict()` generates predictions on the testing set.
- `metrics()` computes evaluation metrics like accuracy and kappa.
It appears we can predict the Gram stain with a `r round(metrics$.estimate[1], 3) * 100`% accuracy based on AMR results of only aminoglycosides and beta-lactam antibiotics. The ROC curve looks like this:
```{r}
predictions %>%
roc_curve(mo, `.pred_Gram-negative`) %>%
autoplot()
```
### **Conclusion**
In this post, we demonstrated how to build a machine learning pipeline with the `tidymodels` framework and the `AMR` package. By combining selector functions like `aminoglycosides()` and `betalactams()` with `tidymodels`, we efficiently prepared data, trained a model, and evaluated its performance.
This workflow is extensible to other antimicrobial classes and resistance patterns, empowering users to analyse AMR data systematically and reproducibly.
---
## Example 2: Predicting ESBL Presence Using Raw MICs
In this second example, we demonstrate how to use `<mic>` columns directly in `tidymodels` workflows using AMR-specific recipe steps. This includes a transformation to `log2` scale using `step_mic_log2()`, which prepares MIC values for use in classification models.
This approach and idea formed the basis for the publication [DOI: 10.3389/fmicb.2025.1582703](https://doi.org/10.3389/fmicb.2025.1582703) to model the presence of extended-spectrum beta-lactamases (ESBL).
### **Objective**
Our goal is to:
1. Use raw MIC values to predict whether a bacterial isolate produces ESBL.
2. Apply AMR-aware preprocessing in a `tidymodels` recipe.
3. Train a classification model and evaluate its predictive performance.
### **Data Preparation**
We use the `esbl_isolates` dataset that comes with the AMR package.
```{r}
# Load required libraries
library(AMR)
library(tidymodels)
# View the esbl_isolates data set
esbl_isolates
# Prepare a binary outcome and convert to ordered factor
data <- esbl_isolates %>%
mutate(esbl = factor(esbl, levels = c(FALSE, TRUE), ordered = TRUE))
```
**Explanation:**
- `esbl_isolates`: Contains MIC test results and ESBL status for each isolate.
- `mutate(esbl = ...)`: Converts the target column to an ordered factor for classification.
### **Defining the Workflow**
#### 1. Preprocessing with a Recipe
We use our `step_mic_log2()` function to log2-transform MIC values, ensuring that MICs are numeric and properly scaled. All MIC predictors can easily and agnostically selected using the new `all_mic_predictors()`:
```{r}
# Split into training and testing sets
set.seed(123)
split <- initial_split(data)
training_data <- training(split)
testing_data <- testing(split)
# Define the recipe
mic_recipe <- recipe(esbl ~ ., data = training_data) %>%
remove_role(genus, old_role = "predictor") %>% # Remove non-informative variable
step_mic_log2(all_mic_predictors()) #%>% # Log2 transform all MIC predictors
# prep()
mic_recipe
```
**Explanation:**
- `remove_role()`: Removes irrelevant variables like genus.
- `step_mic_log2()`: Applies `log2(as.numeric(...))` to all MIC predictors in one go.
- `prep()`: Finalises the recipe based on training data.
#### 2. Specifying the Model
We use a simple logistic regression to model ESBL presence, though recent models such as xgboost ([link to `parsnip` manual](https://parsnip.tidymodels.org/reference/details_boost_tree_xgboost.html)) could be much more precise.
```{r}
# Define the model
model <- logistic_reg(mode = "classification") %>%
set_engine("glm")
model
```
**Explanation:**
- `logistic_reg()`: Specifies a binary classification model.
- `set_engine("glm")`: Uses the base R GLM engine.
#### 3. Building the Workflow
```{r}
# Create workflow
workflow_model <- workflow() %>%
add_recipe(mic_recipe) %>%
add_model(model)
workflow_model
```
### **Training and Evaluating the Model**
```{r}
# Fit the model
fitted <- fit(workflow_model, training_data)
# Generate predictions
predictions <- predict(fitted, testing_data) %>%
bind_cols(testing_data)
# Evaluate model performance
our_metrics <- metric_set(accuracy, kap, ppv, npv)
metrics <- our_metrics(predictions, truth = esbl, estimate = .pred_class)
metrics
```
**Explanation:**
- `fit()`: Trains the model on the processed training data.
- `predict()`: Produces predictions for unseen test data.
- `metric_set()`: Allows evaluating multiple classification metrics.
It appears we can predict ESBL gene presence with a positive predictive value (PPV) of `r round(metrics$.estimate[3], 3) * 100`% and a negative predictive value (NPV) of `r round(metrics$.estimate[4], 3) * 100` using a simplistic logistic regression model.
### **Visualising Predictions**
We can visualise predictions by comparing predicted and actual ESBL status.
```{r}
library(ggplot2)
ggplot(predictions, aes(x = esbl, fill = .pred_class)) +
geom_bar(position = "stack") +
labs(title = "Predicted vs Actual ESBL Status",
x = "Actual ESBL",
y = "Count") +
theme_minimal()
```
### **Conclusion**
In this example, we showcased how the new `AMR`-specific recipe steps simplify working with `<mic>` columns in `tidymodels`. The `step_mic_log2()` transformation converts ordered MICs to log2-transformed numerics, improving compatibility with classification models.
This pipeline enables realistic, reproducible, and interpretable modelling of antimicrobial resistance data.
---
## Example 3: Predicting AMR Over Time
In this third example, we aim to predict antimicrobial resistance (AMR) trends over time using `tidymodels`. We will model resistance to three antibiotics (amoxicillin `AMX`, amoxicillin-clavulanic acid `AMC`, and ciprofloxacin `CIP`), based on historical data grouped by year and hospital ward.
### **Objective**
Our goal is to:
1. Prepare the dataset by aggregating resistance data over time.
2. Define a regression model to predict AMR trends.
3. Use `tidymodels` to preprocess, train, and evaluate the model.
### **Data Preparation**
We start by transforming the `example_isolates` dataset into a structured time-series format.
```{r}
# Load required libraries
library(AMR)
library(tidymodels)
# Transform dataset
data_time <- example_isolates %>%
top_n_microorganisms(n = 10) %>% # Filter on the top #10 species
mutate(year = as.integer(format(date, "%Y")), # Extract year from date
gramstain = mo_gramstain(mo)) %>% # Get taxonomic names
group_by(year, gramstain) %>%
summarise(across(c(AMX, AMC, CIP),
function(x) resistance(x, minimum = 0),
.names = "res_{.col}"),
.groups = "drop") %>%
filter(!is.na(res_AMX) & !is.na(res_AMC) & !is.na(res_CIP)) # Drop missing values
data_time
```
**Explanation:**
- `mo_name(mo)`: Converts microbial codes into proper species names.
- `resistance()`: Converts AMR results into numeric values (proportion of resistant isolates).
- `group_by(year, ward, species)`: Aggregates resistance rates by year and ward.
### **Defining the Workflow**
We now define the modelling workflow, which consists of a preprocessing step, a model specification, and the fitting process.
#### 1. Preprocessing with a Recipe
```{r}
# Define the recipe
resistance_recipe_time <- recipe(res_AMX ~ year + gramstain, data = data_time) %>%
step_dummy(gramstain, one_hot = TRUE) %>% # Convert categorical to numerical
step_normalize(year) %>% # Normalise year for better model performance
step_nzv(all_predictors()) # Remove near-zero variance predictors
resistance_recipe_time
```
**Explanation:**
- `step_dummy()`: Encodes categorical variables (`ward`, `species`) as numerical indicators.
- `step_normalize()`: Normalises the `year` variable.
- `step_nzv()`: Removes near-zero variance predictors.
#### 2. Specifying the Model
We use a linear regression model to predict resistance trends.
```{r}
# Define the linear regression model
lm_model <- linear_reg() %>%
set_engine("lm") # Use linear regression
lm_model
```
**Explanation:**
- `linear_reg()`: Defines a linear regression model.
- `set_engine("lm")`: Uses R’s built-in linear regression engine.
#### 3. Building the Workflow
We combine the preprocessing recipe and model into a workflow.
```{r}
# Create workflow
resistance_workflow_time <- workflow() %>%
add_recipe(resistance_recipe_time) %>%
add_model(lm_model)
resistance_workflow_time
```
### **Training and Evaluating the Model**
We split the data into training and testing sets, fit the model, and evaluate performance.
```{r}
# Split the data
set.seed(123)
data_split_time <- initial_split(data_time, prop = 0.8)
train_time <- training(data_split_time)
test_time <- testing(data_split_time)
# Train the model
fitted_workflow_time <- resistance_workflow_time %>%
fit(train_time)
# Make predictions
predictions_time <- fitted_workflow_time %>%
predict(test_time) %>%
bind_cols(test_time)
# Evaluate model
metrics_time <- predictions_time %>%
metrics(truth = res_AMX, estimate = .pred)
metrics_time
```
**Explanation:**
- `initial_split()`: Splits data into training and testing sets.
- `fit()`: Trains the workflow.
- `predict()`: Generates resistance predictions.
- `metrics()`: Evaluates model performance.
### **Visualising Predictions**
We plot resistance trends over time for amoxicillin.
```{r}
library(ggplot2)
# Plot actual vs predicted resistance over time
ggplot(predictions_time, aes(x = year)) +
geom_point(aes(y = res_AMX, color = "Actual")) +
geom_line(aes(y = .pred, color = "Predicted")) +
labs(title = "Predicted vs Actual AMX Resistance Over Time",
x = "Year",
y = "Resistance Proportion") +
theme_minimal()
```
Additionally, we can visualise resistance trends in `ggplot2` and directly add linear models there:
```{r}
ggplot(data_time, aes(x = year, y = res_AMX, color = gramstain)) +
geom_line() +
labs(title = "AMX Resistance Trends",
x = "Year",
y = "Resistance Proportion") +
# add a linear model directly in ggplot2:
geom_smooth(method = "lm",
formula = y ~ x,
alpha = 0.25) +
theme_minimal()
```
### **Conclusion**
In this example, we demonstrated how to analyze AMR trends over time using `tidymodels`. By aggregating resistance rates by year and hospital ward, we built a predictive model to track changes in resistance to amoxicillin (`AMX`), amoxicillin-clavulanic acid (`AMC`), and ciprofloxacin (`CIP`).
This method can be extended to other antibiotics and resistance patterns, providing valuable insights into AMR dynamics in healthcare settings.
View Git Blame Copy Permalink