Modelling Trajectories of Antimicrobial Resistance in Human Populations

This study develops a Bayesian semi-mechanistic framework that explicitly accounts for data source heterogeneity and uses a logistic growth model to accurately estimate and predict antimicrobial resistance trajectories for *Acinetobacter baumannii* across 32 Asian countries, demonstrating superior performance over existing ensemble methods.

The Gist

The Big Picture: The "Superbug" Weather Forecast

Imagine Antimicrobial Resistance (AMR) as a storm. Bacteria are learning to survive the "rain" of antibiotics, making infections harder to treat. Scientists need to know where this storm is heading (which countries, which years) so they can prepare.

However, there's a problem: The data we have is messy. It's like trying to predict the weather for 32 different countries using reports from 18 different sources, where some sources use thermometers, others use rain gauges, and some only report when it's raining heavily. Some countries have daily reports; others haven't sent a report in years.

This paper is about building a better weather forecasting machine to predict how bad the "superbug" storm will get in Asia, specifically for a bacteria called Acinetobacter baumannii.

---

The Old Way vs. The New Way

The Old Approach (The "Black Box"):
Previous methods were like using a generic computer algorithm that just looked at the numbers and guessed the trend. It was flexible, but it didn't understand why the bacteria were getting stronger. It was like a weather app that says, "It's going to rain," without knowing if it's because of a cold front or a hurricane. Sometimes, these guesses didn't make biological sense.

The New Approach (The "Semi-Mechanistic" Model):
The authors built a new model that acts like a smart, biological detective.

1. The Logic: They used a "Logistic Growth" model. Think of this like filling a bathtub. At first, the water (resistant bacteria) rises quickly. But eventually, the tub gets full, and the water level stabilizes. The model understands this natural limit.
2. The "Source" Problem: They realized that data from a big hospital in a city might look very different from data in a small rural clinic, even in the same country. It's like comparing a photo taken with a professional camera to one taken with a blurry phone. The new model adds a "lens correction" (called source-level scaling) to adjust for these differences so the data can be compared fairly.
3. The "Neighbor" Effect: If Country A has no data, the model looks at its neighbors (Country B and C). If they are similar, the model "borrows" their information to make a smart guess for Country A. This is like asking your neighbors what the weather is like if your own window is covered in fog.

How They Tested It: The "Ensemble" Strategy

Instead of betting on just one model, the authors built six different versions of their detective, each with slightly different rules:

• Some looked only at the country.
• Some looked at the country + the data source quality.
• Some looked at the country + the source + outside factors (like temperature or how much medicine people buy).
• Some looked at the neighbors (spatial effects).

Then, they used a technique called Bayesian Stacking. Imagine a panel of judges. Instead of picking just one winner, they give each judge a vote based on how good they were at guessing the past. The final prediction is a weighted average of all the judges. This "Team of Experts" approach usually gives the most accurate forecast.

What They Found

1. Context Matters: Models that accounted for the differences between data sources (the "lens correction") were almost always better than those that didn't. You can't compare apples to oranges without peeling them first.
2. The Team Wins: The "Stacked Ensemble" (the team of judges) was the most accurate at predicting the future, especially for the most recent years where data was tricky.
3. Filling the Gaps: For countries with very little data, the model successfully used information from neighboring countries to create plausible estimates. It didn't just say "we don't know"; it said, "Based on your neighbors, here is a likely scenario."
4. The Drivers: They found that factors like antibiotic consumption (how much medicine is used) and temperature actually drive these resistance trends. The model can quantify how much a change in these factors might change the resistance levels.

The Takeaway

This paper isn't just about math; it's about better decision-making.

By creating a model that respects the messy reality of real-world data (different labs, different countries, missing reports) and combines it with biological logic (how bacteria actually grow), the authors have given public health officials a powerful tool.

The Analogy:
If the old way was like trying to guess the future of a game by looking at a blurry, torn-up scorecard, this new method is like having a high-definition replay system that corrects for camera angles, fills in missing frames using the players' movements, and consults a team of statisticians to give you the most accurate prediction of the final score.

This helps governments decide where to send resources, how to regulate antibiotic use, and how to stop the "superbug" storm before it hits.

Technical Summary

1. Problem Statement

Antimicrobial resistance (AMR) poses a critical global health threat, yet understanding its temporal trends is hindered by sparse, heterogeneous, and patchy surveillance data, particularly in low- and middle-income countries (LMICs). Existing approaches to modeling country-level resistance trajectories face two primary limitations:

• Phenomenological models: While flexible (e.g., Gaussian processes with stacked ensembles), they often lack biological plausibility, struggle to represent non-equilibrium dynamics, and cannot easily explore counterfactual scenarios or long-term forecasts.
• Mechanistic models: While biologically grounded, they often require complex compartmental structures that are difficult to fit to global, multi-source data with varying sampling protocols.
• Data Heterogeneity: Resistance data originates from diverse sources (clinical settings, national surveillance, systematic reviews) with distinct sampling protocols, patient populations, and reporting standards. Previous ensemble methods often failed to explicitly account for these source-level heterogeneities.

2. Methodology

The authors propose a Bayesian semi-mechanistic framework that bridges the gap between purely statistical and fully mechanistic models.

A. Core Model Structure

• Underlying Dynamics: The model is based on a logistic growth process derived from a simplified two-compartment Susceptible-Infectious-Susceptible (SIS) ordinary differential equation (ODE). This assumes drug-resistant and drug-sensitive strains are decoupled, and the proportion of resistant colonization ($P_t$) follows a logistic curve.
• Parameterization: The logistic curve is parameterized by:
  • $\beta$: Transmission/growth rate.
  • $\gamma$: Loss/recovery rate.
  • $P_0$: Initial prevalence.
  • The carrying capacity ($K$) is defined as $1 - \gamma/\beta$.
• Observation Model: Observed resistance data ($y$) is modeled as a Binomial distribution, scaled by a source-specific multiplicative factor ($\eta_s$) to account for differences in sampling protocols and detection rates across data sources.

B. Model Variants

Six model variants were developed to test different assumptions about spatial and temporal variation:

1. Model C: Country-level random effects (hierarchical priors) only.
2. Model C+X: Country-level effects + time-varying covariates.
3. Model C+S: Country-level effects + source-level scaling.
4. Model SP: Spatial random effects using Gaussian Process (GP) regression (radial basis kernel) to capture spatial autocorrelation between neighboring countries.
5. Model SP+S: Spatial effects + source-level scaling.
6. Model SP+S+X: Spatial effects + source-level scaling + time-varying covariates.

C. Covariates

Time-dependent, country-level covariates were incorporated to drive parameter variation:

• Mean temperature (Copernicus Climate Change Service).
• Total human antibiotic consumption (Defined Daily Doses per 1000 inhabitants).
• Out-of-pocket healthcare expenditures.

D. Ensemble Approach

To optimize predictive performance, the authors employed Bayesian Model Stacking. Using the loo package in R, they calculated Leave-One-Out Cross-Validation (LOO-CV) weights to combine the predictive distributions of the six variants into a single ensemble model.

E. Implementation

• Framework: Bayesian inference using Stan (Hamiltonian Monte Carlo).
• Data: 333,936 clinical isolates of Acinetobacter baumannii from 32 Asian countries (2000–2022) across 7 antibiotic classes, sourced from 18 different datasets.
• Evaluation: Performance was assessed using LOOIC (for model fit) and Mean Absolute Error (MAE) on a held-out test set (the most recent 4 years of data).

3. Key Contributions

1. Semi-Mechanistic Hybrid: The study introduces a framework that uses a biologically plausible logistic growth solution (derived from SIS dynamics) while modeling the relationship between covariates and model parameters phenomenologically. This allows for biological interpretability without the computational burden of full transmission models.
2. Explicit Source Heterogeneity: Unlike previous ensemble approaches, this framework explicitly models source-level scaling ($\eta_s$), acknowledging that different data sources (e.g., hospital surveillance vs. systematic reviews) observe different proportions of the underlying resistance prevalence.
3. Spatial Imputation: The use of Gaussian Processes allows for the estimation of resistance trajectories in countries with no or sparse data by borrowing strength from geographically neighboring countries.
4. Stacked Ensemble Optimization: Demonstrates that combining multiple model variants via Bayesian stacking yields superior out-of-sample predictive accuracy compared to any single model.

4. Key Results

• Model Performance:
  • Models incorporating source-level scaling consistently outperformed those without it across all seven pathogen-antibiotic combinations (lowest LOOIC values).
  • The Stacked Ensemble model achieved the lowest Mean Absolute Error (MAE) for 4 out of 7 antibiotic classes and was highly competitive for the others.
  • Models with covariates (C+X and SP+S+X) generally improved performance over their non-covariate counterparts, particularly for aminoglycosides, antipseudomonal penicillins, carbapenems, fluoroquinolones, and third-generation cephalosporins.
• Predictive Accuracy:
  • For the held-out test set (2019–2022), the stacked model provided the most accurate forecasts.
  • Spatial models (SP variants) produced narrower credible intervals for data-sparse countries, effectively leveraging spatial correlation.
• Covariate Effects:
  • Antibiotic consumption (DDD/1000) showed a strong positive association with resistance growth rates ($\beta$) across most antibiotic classes.
  • Temperature showed mixed or negative associations in some models, suggesting complex environmental interactions.
• Trajectory Estimates:
  • Most countries exhibited increasing resistance trends, though some (notably Japan) showed declines.
  • Long-term saturation levels (carrying capacity) were estimated to be intermediate for most countries rather than reaching 100% fixation.

5. Significance

• Public Health Surveillance: This methodology provides a robust tool for estimating AMR burdens in regions with poor surveillance infrastructure, filling critical data gaps for global health monitoring (e.g., WHO GLASS).
• Intervention Evaluation: By decoupling intrinsic dynamics from observation noise and incorporating covariates, the framework offers a foundation for evaluating the potential impact of public health interventions (e.g., antibiotic stewardship) through counterfactual simulations.
• Methodological Advancement: The paper demonstrates that semi-mechanistic models combined with Bayesian stacking and explicit source modeling represent a superior approach to AMR trajectory estimation compared to purely phenomenological or purely mechanistic methods.
• Scalability: The framework is designed to be extensible, capable of incorporating more complex mechanistic assumptions (e.g., hospital vs. community compartments) or additional covariates as data availability improves.

In conclusion, the authors present a sophisticated, flexible, and biologically grounded modeling framework that significantly improves the accuracy of AMR trajectory estimation in data-scarce settings, offering a vital tool for global AMR surveillance and policy-making.