Epidemiology Demo

DeepNLME for Population-Level Viral Dynamics

Niklas Korsbo

2025-10-12

Setup

Epidemiological Modeling Challenge

The Problem: Age-Stratified Seroprevalence

The Challenge

  • Multi-country seroprevalence data across age groups
  • Unknown force of infection λ(age) patterns
  • Heterogeneous data quality: Some countries well-sampled, others sparse
  • Need to share information between heterogeneous countries
  • Discover age-specific transmission patterns

DeepNLME approach: Let data discover the functional form!

DeepNLME Epidemiological Model

epi_model = @model begin
  @param begin
    N ∈ RealDomain(; lower=0, init=1000)
    λ ∈ MLPDomain(2, 5, 5, (1, softplus); reg=L2(1e-2))
  end
  @random η ~ Normal(0, 0.1)
  @init S = 1
  @dynamics begin
    S' = - λ(t/100, η)[1] * S
  end
  @derived PosFrac ~ @. Beta(abs(N*(1-0.99S)), abs(N*(0.99S)))
end

SIR model, except that we don’t need the I or R here.

Key features:

  • Neural network λ discovers age-specific force of infection
  • Individual random effects η allow country-specific scaling
  • Shared function across populations with individual tuning
  • Mechanistic structure preserves epidemiological interpretation

Results

Model Predictions

Excellent fit across all countries, including sparsely sampled test data!

Comparison with Truth

Neural network accurately recovers the true underlying force of infection!

Discovered Force of Infection Pattern

Complex multi-modal pattern discovered automatically from data!

Information Sharing Across Populations

Dense data informs sparse data through shared neural network function!

Key Insights

What DeepNLME Achieved

  • Data-driven discovery of force of infection λ(age)
  • Automatic pattern recognition without pre-specifying functional forms
  • Information sharing
  • Individual adaptation / posterior predictions
  • Robust extrapolation to poorly sampled regions/countries