Adjust three SEIR model parameters by ±50% (e.g., R₀, isolation rate, latent period). Reflect on which parameters most significantly changed the infection curve (e.g., timing, peak, or total recovered). What do your findings suggest about how public health interventions affect disease spread? How might AI/ML tools use or misinterpret these dynamics if trained on simplified or idealized data?

Baseline Parameters

The baseline SEIR model uses the following parameters:

– Basic Reproduction Number (R₀): 10

– Latent Exposed Period: 3 days

– Initial Infectious Fraction: 1e-3 (0.1%)

Adjusted Parameters (±50%)

Each of the following parameters was adjusted by ±50% to assess their impact on the epidemic curve:

– R₀: 5 and 15

– Latent Exposed Period: 1.5 days and 4.5 days

– Initial Infectious Fraction: 5e-4 and 1.5e-3

Comparison of Effects on Epidemic Curve

Parameter Change	R₀	Latent Period (days)	Initial I0	Timing of Peak (days)	Peak Infectious (Imax)	Total Recovered (Rfinal)
Baseline	10.0	3.0	10	16	3797.85	9999.55
R₀ ↓ to 5	5.0	3.0	10	Later 24	Lower 2850.94	Smaller 9930.46
R₀ ↑ to 15	15.0	3.0	10	Earlier 13	Higher 4128.99	Larger 10000.00
Latent ↓ to 1.5 d	10.0	1.5	10	Earlier 12	Higher, Sharper 4809.10	Similar 9999.55
Latent ↑ to 4.5 d	10.0	4.5	10	Later 19	Lower, Broader 3176.60	Similar 9999.55
Initial ↓ to 5e-4	10.0	3.0	5	Delayed 16	Similar 3780.95	Similar 9999.55
Initial ↑ to 1.5e-3	10.0	3.0	15	Earlier 15	Similar 3782.00	Similar 9999.55

Observations

– R₀ has the strongest influence on both the peak infectious load and the total number of recovered individuals. A higher R₀ leads to a faster, larger outbreak.

– The latent period significantly affects the timing of the peak and the shape of the infectious curve. A shorter latent period results in a sharper, earlier peak.

– The initial infectious fraction primarily shifts the onset timing of the outbreak. It has minimal effect on the peak magnitude or the final epidemic size.

Based on the exercise, there are a couple aspects provide very useful lessons for planning out a robust strategy for epidemiology:

1. The Critical Role of R₀: Controlling the Final Scale

The R₀ parameter, which represents the average number of new infections caused by one infectious person, was found to be the most significant driver of the Total Recovered (final epidemic size).

• Intervention Implication (Flattening the Curve): Interventions aimed at reducing the transmission rate (ß) or the duration of infectiousness are the most effective strategies.
  • Reducing R₀ below 1 is Everything: A 50% reduction in R₀ (from 10.0 to 5.0) significantly reduced the final epidemic size in the model (from roughly 10,000 to $9,930 in the relevant R₀ = 5 scenarios). A larger reduction, specifically pushing the effective R below 1, is necessary to halt the epidemic entirely.
  • Actionable Measures: This validates high-impact interventions like mask mandates, social distancing, large-scale lockdowns, and vaccination (which reduces ß by making people less susceptible or less infectious). These measures are essential for reducing the total death toll and achieving herd immunity thresholds.

2. Latent Period: The Challenge of Speed and Peak Severity

The Latent Exposed Period (the time before an exposed person becomes infectious) had the most significant impact on the peak magnitude and timing when R₀ was held constant.

• Intervention Implication: The model showed that a 50% shorter latent period (LP -50%) resulted in the single highest and earliest peak of all scenarios (4,809 infections peaking at 12 days).
  • Actionable Measures: This highlights the urgency of managing diseases with short incubation periods. Interventions must be implemented rapidly to get ahead of the curve. This is the rationale behind aggressive contact tracing and testing programs, as the goal is to identify and isolate individuals before they exit their latent period and start spreading the disease, functionally reducing the infectious population faster.

In terms of AI/ML systems trained on simplified or idealized data, here are some beneficial points which can be quite useful:

• Forecasting outbreaks: ML models can learn relationships between parameters (R₀, latent period, infectious duration) and epidemic outcomes (timing, peak, total recovered). This helps in scenario planning and resource allocation.
• Policy simulation: AI can test “what if” interventions (e.g., reducing R₀ via masks or vaccines) and estimate their impact on peak hospital demand.
• Early warning systems: By fitting SEIR-like curves to real-time case data, ML can detect deviations that suggest new variants or intervention effects.

On the other hand, these systems trained on simplified or idealized data might also be easily misinterpret dynamics:

• Overfitting to idealized curves: Epidemics like COVID-19 rarely follow neat SEIR trajectories. Noise, under-reporting, and heterogeneous mixing can make ML models trained on clean data overly confident.
• Ignoring parameter interdependence: In reality, interventions that reduce R₀ also change effective latent/infectious periods (e.g., isolation shortens infectious duration). ML trained on static SEIR assumptions may miss these couplings.
• Bias toward average dynamics: SEIR assumes homogeneous populations. ML trained on such data may fail to capture age, geography, or network effects, leading to poor predictions in subpopulations.

Conclusion:

DOs

Timing control: Rapid testing, surveillance, and isolation extend the effective latent period, buying time for vaccination or treatment rollout.

Peak reduction: Measures that reduce R₀ (vaccination, masks, distancing, ventilation) are the most effective at lowering the peak burden on healthcare systems.

Final size reduction: Only interventions that sustainably reduce R₀ (herd immunity via vaccination, long-term behavior change) shrink the total number of people who ever get infected.

DON’T

AI/ML can amplify insights from SEIR models, but if trained only on idealized data, it risks mistaking “textbook epidemics” for reality. The danger is not just inaccurate forecasts, but misplaced confidence in interventions that don’t behave as neatly in practice.