Source-linked AI summary
An infectious disease model on empirical networks of human contact: bridging the gap between dynamic network data and contact matrices
Anna Machens, Francesco Gesualdo, Caterina Rizzo, Alberto E Tozzi, Alain Barrat, Ciro Cattuto
TL;DR
The paper asks how to summarize detailed contact data for epidemic models without losing information needed to estimate outcomes and group-level risks. It compares an SEIR model on several representations of hospital-ward contacts and finds that a distribution-based contact matrix approximates key results from high-resolution data better than a mean-only matrix. The study also identifies limits tied to its small, single-time-point dataset and short latent periods.
Problem
For structured populations, the appropriate balance between compact contact summaries and retained detail is unclear, although realistic group-level risk estimates matter for evaluating targeted interventions.
Source: Background · pp. 3–4Method
The study compares SEIR simulations on multiple representations of one week of high-resolution hospital contact data, including a contact matrix of negative binomial duration distributions.
Source: Empirical contact patterns · p. 6, Discussion · p. 12Results
DYN, HET, and CMD give similar role-class attack-rate patterns, whereas HOM and CM overestimate attack rates across classes and mis-rank their relative risks.
Source: Results · p. 11Takeaways & Limitations
In this case study, a distribution-based contact matrix provides a compact representation that estimates relative risk across role classes while retaining little time-varying network information.
Source: Conclusions · p. 14The study uses a relatively small population observed at one time point and considers diseases with short latent periods.
Source: Conclusions · p. 14
Abstract
from arXiv · showhide
The integration of empirical data in computational frameworks to model the spread of infectious diseases poses challenges that are becoming pressing with the increasing availability of high-resolution information on human mobility and contacts. This deluge of data has the potential to revolutionize the computational efforts aimed at simulating scenarios and designing containment strategies. However, the integration of detailed data sources yields models that are less transparent and general. Hence, given a specific disease model, it is crucial to assess which representations of the raw data strike the best balance between simplicity and detail. We consider high-resolution data on the face-to-face interactions of individuals in a hospital ward, obtained by using wearable proximity sensors. We simulate the spread of a disease in this community by using an SEIR model on top of different mathematical representations of the contact patterns. We show that a contact matrix that only contains average contact durations fails to reproduce the size of the epidemic obtained with the high-resolution contact data and also to identify the most at-risk classes. We introduce a contact matrix of probability distributions that takes into account the heterogeneity of contact durations between (and within) classes of individuals, and we show that this representation yields a good approximation of the epidemic spreading properties obtained by using the high-resolution data. Our results mark a step towards the definition of synopses of high-resolution dynamic contact networks, providing a compact representation of contact patterns that can correctly inform computational models designed to discover risk groups and evaluate containment policies. We show that this novel kind of representation can preserve in simulation quantitative features of the epidemics that are crucial for their study and management.
Background
The paper asks how much detail contact-pattern representations need to retain to model epidemics accurately while remaining compact. It addresses this question for structured populations, where group-level risk estimates inform targeted interventions.
Background
- Contact-matrix summaries can miss important fluctuations in contact numbers and durations, while the appropriate level of detail for a given modeling problem remains unclear.Coarse representations are analytically tractable, but realistic simulations are needed to estimate epidemic outcomes and compare targeted interventions.Source: Background · p. 3
- Prior work considered contact-network summaries in unstructured populations, but structured-population data summarization had not been investigated in this way.The study is motivated by applications such as prioritizing groups for vaccination or prophylaxis.Source: Background · p. 4
- The paper tests whether different summaries of contact patterns reproduce epidemic outcomes and identify at-risk groups, balancing compactness against retained detail.The motivation is to support interventions targeting classes of people, such as vaccination, rather than requiring individual-level risk estimates.Source: Background · p. 4
- The study applies an SEIR model to high-resolution face-to-face contact data from a hospital ward, comparing multiple representations of those contacts.The data record the time-ordered contacts between subject pairs, and the population is organized by hospital role.Source: Background · p. 5
Data
The empirical dataset records close-proximity contacts among 119 people in a pediatric hospital ward over one week. Participants are divided into five role classes, and wearable badges measured face-to-face proximity.
Data
- The dataset covers 119 people in a general pediatric ward over one week, from 9–16 November 2009.The ward had 44 beds in 22 rooms and was physically separated from other hospital wards and facilities.Source: Data · p. 5
- Wearable badges measured close-proximity interactions through bidirectional low-power radio communication.The devices assess face-to-face proximity between individuals.Source: Data · p. 6
- Participants comprised 37 patients, 20 physicians, 21 nurses, 10 ward assistants, and 31 caregivers.Badges missing for more than 25% of their assigned time were excluded; the dataset included more than 90% of on-duty healthcare workers.Source: Data · p. 6
Empirical contact patterns
The paper compares contact representations that progressively aggregate the hospital ward’s time-resolved interaction data. Its proposed contact matrix of distributions retains role-pair variation in contact durations while discarding individual network structure.
Empirical contact patterns
- DYN is the reference representation: it retains individual contact events and their start and end times at 20-second resolution.For simulations longer than the one-week observation period, the recorded history is repeated.Source: Empirical contact patterns · p. 6
- HET and HOM preserve which individuals met, but HET retains heterogeneous total contact durations whereas HOM assigns every observed link the same average duration.Both static representations discard contact timing and ordering.Source: Empirical contact patterns · p. 7
- CM assigns each individual pair a weight determined by their role classes, using the average contact time for that class pair.The matrix’s rows and columns correspond to role classes; its complete-graph representation does not preserve empirical network topology.Source: Empirical contact patterns · p. 7
- The proposed CMD represents each role pair with a fitted negative binomial distribution and samples edge weights from that distribution.Zero weights represent missing links, and CMD preserves the average number of nonzero links between each pair of classes without retaining the empirical network’s detailed structure.Source: Empirical contact patterns · p. 8
- CMB is an intermediate matrix representation that preserves class-pair link density and average edge weights, but not edge-weight variance.It assigns the class-pair average weight with probability equal to observed link density, and otherwise assigns zero.Source: Empirical contact patterns · p. 8
Infectious disease model
The study simulates an SEIR epidemic on each contact representation, starting with one randomly chosen infectious person. Two scenarios vary the latent and infectious periods, and the comparison keeps microscopic epidemiological parameters fixed.
Infectious disease model
- In the SEIR model, susceptible people may become exposed through contact with infectious people, exposed people become infectious at rate σ, and infectious people recover at rate ν.Recovered individuals acquire permanent immunity; births, deaths, and introductions are excluded.Source: Infectious disease model · p. 8
- Simulations begin with a susceptible population and one infectious individual chosen at random.Source: Infectious disease model · p. 9
- The study uses two scenarios: 1/σ=1 day and 1/ν=2 days, or 1/σ=0.5 day and 1/ν=1 day.The second scenario halves both periods to probe their interplay with the faster timescale of human contacts.Source: Infectious disease model · p. 9
- For each scenario, the study runs 16,000 epidemics for DYN, HOM, and HET, and 16,000 epidemics across sampled CMD and CMB networks.The CMD and CMB runs use 1,600 network samples, each with 10 epidemic realizations.Source: Infectious disease model · p. 9
- The main comparison holds microscopic epidemiological parameters fixed across contact representations; an alternative equal-R0 comparison is reported to yield similar results.Source: Infectious disease model · p. 9
Analysis
The analysis compares epidemic outcomes across contact-pattern representations, using simulations on the detailed DYN representation as the reference for faithfulness. It focuses on outbreak size, seed-role effects, and role-specific attack rates, with additional indicators reported in supplementary analyses.
Analysis
- The main comparisons examine final recovered counts, the effect of the initial seed's role, and attack-rate distributions across role classes.Source: Analysis · p. 9
- The analysis also computes extinction probability and the fraction of runs with a global attack rate below 10%.Source: Analysis · p. 10
- Supplementary analyses report peak time, epidemic duration, and the basic reproduction number.Source: Analysis · p. 10
- Faithfulness is assessed by how closely each representation's simulated epidemic outcomes match those from DYN, which retains precise contact timing and duration.Source: Analysis · p. 10
Results
In hospital-ward simulations, the contact matrix of distributions (CMD) approximates spreading patterns from richer network representations better than the customary average-duration matrix. The results also show that representation choice affects estimated outbreak size and risk rankings, while weekly aggregation performs similarly to the tested daily alternatives.
Results
- CMD fits a negative binomial distribution to cumulative contact durations for each role pair, capturing broad duration heterogeneity and zero-weight links while retaining a class-level matrix form.The empirical dataset records about 16,000 contact events over one week among 119 hospital-ward individuals.Source: Results · p. 10, Discussion · p. 12
- Final-size distributions for DYN, HET, and CMD overlap, whereas HOM and CM overestimate epidemic size and underestimate extinction and small-outbreak probabilities.CMB performs intermediately but remains a poor approximation of HET and CMD.Source: Results · pp. 10–11
- DYN, HET, and CMD identify assistants and nurses as more affected, while HOM and CM overestimate attack rates across classes and misrank role-specific risks.Seed role also influences extinction and small-outbreak probability; assistant-seeded epidemics are most likely to spread.Source: Results · p. 11
- HOM and CM produce higher epidemic peaks than DYN, HET, and CMD, while peak timing varies only slightly across representations.Source: Results · p. 11
- The usual contact matrix is easy to generalize when role classes remain similar, but it discards within-class heterogeneity in contact durations.Source: Discussion · p. 12
- CMD gives good qualitative and quantitative agreement with HET and improves on the customary contact matrix, correctly ranking role classes by infection risk for targeted vaccination policies.Source: Discussion · p. 13
- Daily networks and matrices yielded results close to weekly-aggregated representations, suggesting weekly averages contain enough information for the tested mitigation and containment evaluations.Source: Discussion · pp. 13–14
- Static representations were similar to time-resolved DYN in this study, where spreading and network timescales were well separated; broader timescale comparisons remain of interest.HET had a smaller extinction probability than DYN, and differences between them were larger for faster propagation.Source: Discussion · pp. 13–14
Conclusions
The study finds that average-only contact matrices are too coarse to capture important epidemic features, while a compact matrix of contact-duration distributions can preserve them without retaining the full time-varying network. The evidence is from a specific case study, with broader validation still needed.
Conclusions
- In a case study, a contact matrix of negative binomial distributions modeled important epidemic features, including relative risk across role classes, while retaining a compact representation.The representation summarizes empirical contact data while retaining little information from the time-varying network.Source: Conclusions · p. 14
- Average-only contact matrices are not refined enough, whereas the full time-varying contact network is unnecessarily detailed and complex for most practical cases.The study identifies measuring heterogeneity in contact durations between individuals of given classes as crucial.Source: Conclusions · p. 14
- The study was based on nearly one hundred individuals observed at one point in time and considered diseases with short latent periods.For the dynamic representation, the limited data span required artificially extended activity timelines made by repeating the empirical timeline.Source: Conclusions · p. 14
- Static representations omit behavioral variation across days of the week and differences in the amount of time individuals spend in the ward.The study notes that CMD and HET results might differ more in larger populations because of specific network effects.Source: Conclusions · pp. 14–15
- Further validation calls for data from different structured populations, larger populations, longer time scales, and diseases with longer latent periods.The authors also identify checking the negative-binomial fitting procedure, testing other distribution forms, and evaluating data-driven containment strategies as future work.Source: Conclusions · p. 15