Tracing Pathogen Through Airport Wastewater Sampling

Abstract

Airports are critical entry points for infectious diseases, and wastewater sampling offers a scalable way to monitor pathogens carried by arriving travellers. At Amsterdam Airport Schiphol, sampling currently takes place only at the wastewater treatment plant, which pools all arriving flights and cannot attribute a detected pathogen to a specific flight or origin. While the literature has shown that aircraft and terminal sampling can provide origin specificity, no study has formalised the selection of sampling points within a single airport. This thesis addresses that gap through the research question: How can the wastewater sampling network at Amsterdam Airport Schiphol be designed and optimised to maximise the detection probability of infectious diseases while supporting attribution to flights and staying within operational and resource constraints?

The problem is formalised as a generic mixed-integer non-linear programme that selects among four node types (aircraft tanks, lavatory trucks, concourses, and the treatment plant) to maximise attribution-weighted detection capacity across all arriving flights subject to a budget constraint. A probability chain connects the epidemiological situation at each flight's origin to a detection event, combining per-node probabilities through a conditional noisy-OR formulation. The model is implemented in Python with Gurobi and applied to Schiphol using data from a representative busy day under three disease scenarios, with an extensive sensitivity and robustness analysis.

The central finding is that there is no single optimal network. The configuration that maximises detection capacity depends on how disease risk is distributed across arriving flights. For the broadly distributed SARS-CoV-2 base case, maximum coverage through concourse and lavatory truck sampling is optimal: detection capacity reaches 40.07 at budget 15,000 using all seven concourses and all 140 truck nodes, with no aircraft nodes selected. Spending the first 15,000 cost units on trucks and concourses gains 38.63 in detection capacity compared to only sampling the treatment plant, while adding aircraft nodes between budget 25,000 and 75,000 gains only 0.74. For the rare and geographically concentrated Hantavirus scenario, with a single at-risk flight from Argentina, the strategy differs. A single aircraft tank node targeting that flight achieves 5.08 x 10^-6 at budget 10,000, a 33% improvement over the combined truck, concourse, and treatment plant configuration. The same model, applied to the same airport on the same day, recommends opposite strategies depending only on the disease.

A comparison against a greedy benefit-cost heuristic shows that exact optimisation adds value precisely where it matters most: for SARS-CoV-2 the two approaches produce identical configurations through saturation, but exact optimisation outperforms greedy by 33% for Hantavirus and 7% for a concentrated outbreak. The optimal configuration is also robust across most sources of uncertainty, remaining invariant to a x0.1 to x10 scaling of infection probability and stable across six peak-season days. The main limitation is that the model evaluates a static 24-hour window and therefore optimises detection probability rather than detection timing. Extending it with a temporal layer, forward-looking prevalence estimation, a finer haul classification, and a multi-pathogen formulation are the most valuable directions for further research.

The findings translate into concrete guidance for surveillance practice. Moving beyond the current treatment-plant-only setup does not require large investment: adding all seven concourse nodes raises detection capacity from 1.44 to 9.46, a 6.6x improvement, at a cost of only 71 relative units and with no airside access. For routine surveillance of a broadly distributed disease, lavatory truck sampling should form the backbone of the network, and aircraft sampling is not cost-justified. Aircraft sampling should instead be held ready as a targeted capability for rare or geographically concentrated threats. The optimal network is not a fixed installation but a configuration that should be switched by scenario, so the operational value lies in retaining the flexibility to reconfigure it. Because the configuration is sensitive only to the cost ratios between node types, establishing the true relative cost of node types at a given airport is the parameter most worth validating before deployment.