P281424/AMR
1
0
Fork 0
mirror of https://github.com/msberends/AMR.git synced 2026-06-24 14:16:19 +02:00
Code Releases Activity

(v3.0.1.9059) Fix WISCA in vignette

Browse Source
This commit is contained in:
dr. M.S. (Matthijs) Berends
2026-06-23 14:38:59 +02:00
parent 3f9f931777
commit 9898b5df4b
41 changed files with 1310 additions and 757 deletions
Download Patch File Download Diff File Expand all files Collapse all files

125
vignettes/WISCA.Rmd
View File

@@ -22,75 +22,58 @@ knitr::opts_chunk$set(
)
```
> This explainer was largely 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.
## Why WISCA?
## Introduction
When a clinician starts empirical antimicrobial therapy, the causative pathogen is unknown. The question they need answered is not *"what proportion of* E. coli *is susceptible to ciprofloxacin?"* but rather *"what is the probability that this regimen will adequately cover whatever pathogen turns out to be causing my patient's infection?"*
Clinical guidelines for empirical antimicrobial therapy require *probabilistic reasoning*: what is the chance that a regimen will cover the likely infecting organisms, before culture results are available?
The traditional cumulative antibiogram, as standardised by CLSI M39, cannot answer that question. It presents susceptibility percentages per species per antibiotic, but:
This is the purpose of **WISCA**, or **Weighted-Incidence Syndromic Combination Antibiogram**.
- **It fragments information by organism.** The clinician must mentally combine susceptibility rates across multiple species, weighting by how often each species causes the syndrome, a calculation nobody does at the bedside.
- **It ignores pathogen incidence.** A species that causes 2% of infections is given the same visual weight as one that causes 60%.
- **It does not evaluate combination regimens.** Much empirical therapy consists of two or more agents, but the traditional antibiogram only shows monotherapy per organism.
- **It provides no measure of uncertainty.** A reported "90% susceptible" based on 50 isolates has a 95% confidence interval of roughly 78-97% (Clopper-Pearson), yet the antibiogram presents it as a point estimate without context.
WISCA is a Bayesian approach that integrates:
**WISCA** (Weighted-Incidence Syndromic Combination Antibiogram) resolves all four limitations. It estimates the probability that a regimen will provide adequate empirical coverage for a given infection syndrome, weighted by local pathogen incidence, with full uncertainty quantification via Bayesian inference.
- **Pathogen prevalence** (how often each species causes the syndrome),
- **Regimen susceptibility** (how often a regimen works *if* the pathogen is known),
The concept was introduced by Hebert *et al.* (2012), who demonstrated that traditional antibiogram susceptibility rates could be misleading: ciprofloxacin appeared 84% effective against *E. coli* in the traditional antibiogram, but WISCA revealed only 62% coverage for UTI and 37% for abdominal infections, because enterococci (intrinsically resistant) and other species contribute substantially to these syndromes. Randhawa *et al.* (2014) showed that WISCA-guided regimen selection could improve time-to-adequate-coverage on the ICU by over 40%. Bielicki *et al.* (2016) introduced the Bayesian framework now used in this package, enabling credible intervals and multi-centre pooling. Cook *et al.* (2022) applied it globally across 52 hospitals in 23 countries.
to estimate the **overall empirical coverage** of antimicrobial regimens, with quantified uncertainty.
This vignette explains how WISCA works, why it is useful, and how to apply it using the `AMR` package.
## Why traditional antibiograms fall short
A standard antibiogram gives you:
```
Species → Antibiotic → Susceptibility %
```
But clinicians don’t know the species *a priori*. They need to choose a regimen that covers the **likely pathogens**, without knowing which one is present.
Traditional antibiograms calculate the susceptibility % as just the number of resistant isolates divided by the total number of tested isolates. Therefore, traditional antibiograms:
- Fragment information by organism,
- Do not weight by real-world prevalence,
- Do not account for combination therapy or sample size,
- Do not provide uncertainty.
## The idea of WISCA
## The idea
WISCA asks:
> "What is the **probability** that this regimen **will cover** the pathogen, given the syndrome?"
This means combining two things:
This means combining two quantities:
- **Incidence** of each pathogen in the syndrome,
- **Pathogen incidence** in the syndrome (how often each species causes it),
- **Susceptibility** of each pathogen to the regimen.
We can write this as:
$$\text{Coverage} = \sum_i (\text{Incidence}_i \times \text{Susceptibility}_i)$$
For example, suppose:
For example, suppose in your hospital:
- *E. coli* causes 60% of cases, and 90% of *E. coli* are susceptible to a drug.
- *Klebsiella* causes 40% of cases, and 70% of *Klebsiella* are susceptible.
- *E. coli* causes 60% of UTIs, and 90% of *E. coli* are susceptible to a drug.
- *Klebsiella* causes 40% of UTIs, and 70% of *Klebsiella* are susceptible.
Then:
$$\text{Coverage} = (0.6 \times 0.9) + (0.4 \times 0.7) = 0.82$$
But in real data, incidence and susceptibility are **estimated from samples**, so they carry uncertainty. WISCA models this **probabilistically**, using conjugate Bayesian distributions.
That 82% is a far more clinically meaningful number than the species-level "90% of *E. coli*" and "70% of *Klebsiella*" reported separately in a traditional antibiogram, because it directly answers the question the clinician actually faces.
## The Bayesian engine behind WISCA
But in real data, both incidence and susceptibility are **estimated from finite samples**, so they carry uncertainty. A sample of 50 isolates is not a census. WISCA models this uncertainty **probabilistically**, using conjugate Bayesian distributions.
## The Bayesian engine
### Pathogen incidence
Let:
- $K$ be the number of pathogens,
- $\boldsymbol{\alpha} = (1, 1, \ldots, 1)$ be a $\text{Dirichlet}$ prior (uniform),
- $\boldsymbol{n} = (n_1, \ldots, n_K)$ be the observed counts per species.
- $\boldsymbol{\alpha} = (1, 1, \ldots, 1)$ be a $\text{Dirichlet}$ prior (uniform, non-informative),
- $\boldsymbol{n} = (n_1, \ldots, n_K)$ be the observed isolate counts per species.
Then the posterior incidence is:
@@ -100,15 +83,17 @@ To simulate from this, we use:
$$x_i \sim \text{Gamma}(\alpha_i + n_i,\ 1), \quad p_i = \frac{x_i}{\sum_{j=1}^{K} x_j}$$
The Dirichlet is the conjugate prior for multinomial data. With the non-informative prior $\text{Dirichlet}(1, 1, \ldots, 1)$, the posterior is dominated by the data once sample sizes are reasonable. With small samples, the posterior is appropriately more diffuse, reflecting genuine uncertainty, and the resulting credible intervals will be wider.
### Susceptibility
Each pathogen--regimen pair has a prior and data:
Each pathogen-regimen pair has a prior and observed data:
- Default prior: $\text{Beta}(0.5, 0.5)$ (Jeffreys prior)
- Intrinsically resistant pairs: $\text{Beta}(1, 9999)$, forcing near-zero susceptibility regardless of observed data (based on EUCAST Expected Resistant Phenotypes)
- Data: $S$ susceptible out of $N$ tested
The $S$ category could also include values SDD (susceptible, dose-dependent) and I (intermediate \[CLSI\], or susceptible, increased exposure \[EUCAST\]).
The $S$ category could also include values SDD (susceptible, dose-dependent) and I (intermediate [CLSI], or susceptible, increased exposure [EUCAST]).
Then the posterior is:
@@ -129,9 +114,25 @@ Repeat this simulation (e.g., 1000 times) and summarise:
- **Mean** = expected coverage
- **Quantiles** = credible interval (95% by default)
Because each simulation draws from the full posterior, the resulting distribution of coverage estimates naturally captures the joint uncertainty in both pathogen incidence and susceptibility. The credible interval tells you how confident you can be in the coverage estimate, something a traditional antibiogram never provides.
## When to use WISCA vs. traditional antibiograms
| Goal | Recommended approach |
|------|---------------------|
| Guide empirical therapy decisions | **WISCA** |
| Compare regimens for a syndrome | **WISCA** |
| Evaluate combination regimens | **WISCA** |
| Antimicrobial stewardship (A-team) | **WISCA** |
| Track resistance trends per species | Traditional / Combination |
| AMR surveillance reporting | Traditional / Syndromic |
| Understand species-level epidemiology | Traditional |
In short: if the end goal involves a *patient* who does not yet have a culture result, WISCA is the appropriate tool. If the end goal is *surveillance* of resistance at the species level, the traditional antibiogram remains fit for purpose.
## Practical use in the `AMR` package
### Prepare data and simulate synthetic syndrome
### Prepare data
```{r}
library(AMR)
@@ -140,11 +141,11 @@ data <- example_isolates
# Structure of our data
data
# Add a fake syndrome column
data$syndrome <- ifelse(data$mo %like% "coli", "UTI", "No UTI")
# Add a synthetic syndrome column for demonstration
data$syndrome <- ifelse(data$mo %like% "coli", "UTI", "Non-UTI")
```
### Basic WISCA antibiogram
### Basic WISCA
```{r}
wisca(data,
@@ -154,6 +155,8 @@ wisca(data,
### Use combination regimens
Combination regimens are specified with a `+` separator. WISCA evaluates whether *at least one* agent in the combination covers the pathogen:
```{r}
wisca(data,
antimicrobials = c("AMC", "AMC + CIP", "AMC + GEN")
@@ -162,6 +165,8 @@ wisca(data,
### Stratify by syndrome
Use `syndromic_group` to produce separate WISCA estimates per clinical stratum. You can pass a column name or any expression:
```{r}
wisca(data,
antimicrobials = c("AMC", "AMC + CIP", "AMC + GEN"),
@@ -179,6 +184,12 @@ wisca(data,
)
```
### Interpreting the output
Each row shows the estimated empirical coverage for a regimen, with a 95% credible interval. When comparing regimens:
- **Overlapping credible intervals** mean there is no statistically significant difference in coverage. If a narrower-spectrum regimen overlaps with a broader one, the narrower-spectrum option can be preferred on stewardship grounds.
- **Non-overlapping credible intervals** indicate a clinically meaningful difference in coverage.
## Sensible defaults, which can be customised
@@ -186,19 +197,26 @@ wisca(data,
- `conf_interval = 0.95`: coverage interval width
- `combine_SI = TRUE`: count "I" and "SDD" as susceptible
## Practical considerations
- **First isolates only**: always deduplicate using `first_isolate()` before running WISCA. Repeat isolates introduce bias.
- **Pathogen selection**: consider filtering with `top_n_microorganisms()`. Including rare contaminants (e.g. CoNS without clinical context) can distort estimates and may artificially lower coverage (Cook *et al.*, 2022).
- **Sample size**: coverage estimates become reliable with approximately 100+ isolates. For smaller datasets, consider pooling data from multiple sites, but only after verifying that pathogen distributions are sufficiently similar (Bielicki *et al.*, 2016).
- **Culture request bias**: WISCA is only as good as the data it is based on. If cultures are selectively requested (e.g. only after treatment failure), the dataset will be biased towards resistant isolates. A robust culture policy is essential for reliable estimates.
## Limitations
- It assumes your data are representative
- No adjustment for patient-level covariates, although these could be passed onto the `syndromic_group` argument
- WISCA does not model resistance over time, you might want to use `tidymodels` for that, for which we [wrote a basic introduction](https://amr-for-r.org/articles/AMR_with_tidymodels.html)
- It assumes your data are representative of the patient population you are treating
- No direct adjustment for patient-level covariates, although these can be passed onto the `syndromic_group` argument for stratification
- WISCA does not model resistance trends over time; for that, you might want to use `tidymodels`, for which we [wrote a basic introduction](https://amr-for-r.org/articles/AMR_with_tidymodels.html)
## Summary
WISCA enables:
- Empirical regimen comparison,
- Syndrome-specific coverage estimation,
- Fully probabilistic interpretation.
- **Empirical regimen comparison**, answering the clinician's actual question
- **Syndrome-specific coverage estimation**, stratifiable by any clinical variable
- **Fully probabilistic interpretation**, with credible intervals that honestly communicate uncertainty
It is available in the `AMR` package via either:
@@ -208,6 +226,9 @@ wisca(...)
antibiogram(..., wisca = TRUE)
```
## Reference
## References
Bielicki, JA, et al. (2016). *Selecting appropriate empirical antibiotic regimens for paediatric bloodstream infections: application of a Bayesian decision model to local and pooled antimicrobial resistance surveillance data.* **J Antimicrob Chemother**. 71(3):794-802. https://doi.org/10.1093/jac/dkv397
1. Hebert C, Ridgway J, Vekhter B, Brown EC, Weber SG, Robicsek A. Demonstration of the weighted-incidence syndromic combination antibiogram: an empiric prescribing decision aid. *Infect Control Hosp Epidemiol.* 2012;33(4):381-388. https://doi.org/10.1086/664768
2. Randhawa V, Sarwar S, Walker S, Elligsen M, Palmay L, Daneman N. Weighted-incidence syndromic combination antibiograms to guide empiric treatment of critical care infections: a retrospective cohort study. *Crit Care.* 2014;18(3):R112. https://doi.org/10.1186/cc13901
3. Bielicki JA, Sharland M, Johnson AP, Henderson KL, Cromwell DA. Selecting appropriate empirical antibiotic regimens for paediatric bloodstream infections: application of a Bayesian decision model to local and pooled antimicrobial resistance surveillance data. *J Antimicrob Chemother.* 2016;71(3):794-802. https://doi.org/10.1093/jac/dkv397
4. Cook A, Sharland M, Yau Y, Bielicki J. Improving empiric antibiotic prescribing in pediatric bloodstream infections: a potential application of weighted-incidence syndromic combination antibiograms (WISCA). *Expert Rev Anti Infect Ther.* 2022;20(3):445-456. https://doi.org/10.1080/14787210.2021.1967145