Poster (Other) FZJ-2026-00929

http://join2-wiki.gsi.de/foswiki/pub/Main/Artwork/join2_logo100x88.png
Improving predictive ability for epidemic situations by developing optimal variable sets for epidemic panels using Machine Learning methods

 ;  ;  ;  ;  ;  ;  ;  ;  ;

2025

Helmholtz AI Conference 2025, HAICON 2025, KarlsruheKarlsruhe, Germany, 3 Jun 2025 - 5 Jun 20252025-06-032025-06-05 ()

Abstract: Population-based studies provide a valuable means of gathering data from individuals or households, enabling researchers to capture crucial information on disease prevalence, risk factors, behavioral patterns, and other relevant aspects within the population [Vaughan et al., 2021]. These studies are instrumental in estimating disease frequency and identifying individuals infected with specific pathogens, including those who may not have sought medical care or were asymptomatic. Furthermore, population-based research data plays a crucial role in identifying risk factors and understanding disease transmission determinants, encompassing individual characteristics, socio-demographic factors, behavioral patterns, and environmental exposures contributing to disease spread.One such study, MuSPAD, is a population-based sequential multilocal seroprevalence study conducted in two cross-sectional parts during 2020 to 2021, involving more than 32,000 participants across eight regions in Germany [Gornyk et al., 2021]. This study collected blood samples and survey questionnaires to gather comprehensive data. In our work, we first use the epidemiological knowledge to narrow down the number of variables available in the MuSPAD data. We then aggregate individual-level data to represent the local population and employ Machine Learning techniques to predict local COVID-19 incidences. Our objective is to identify best incidence predictors, facilitating the rapid implementation of measurements and surveys of these variables during future pandemics.We employ a number of methods, gradually increasing the complexity of the predictive models. We start with those without temporal awareness, namely LASSO and MLP, and move towards the ones leveraging the time-flow, regularised ARIMA and LSTM. We assess the interpretability of the models using the techniques LIME and SHAP. We demonstrate that, despite a moderate number of participants per time point in our dataset, it is feasible to predict local COVID-19 incidence while keeping the overfitting at an acceptable level.Differential privacy (DP) was applied to the MLP model to preserve data integrity and privacy concerns. DP ensures that the inclusion or exclusion of any single data point in a dataset does not significantly alter the results of an analysis by introducing carefully calibrated noise. The stability of the MLP at different stages of DP was investigated by comparing their feature importance values.


Contributing Institute(s):
  1. Jülich Supercomputing Center (JSC)
Research Program(s):
  1. 5112 - Cross-Domain Algorithms, Tools, Methods Labs (ATMLs) and Research Groups (POF4-511) (POF4-511)
  2. Helmholtz AI Consultant Team FB Information (E54.303.11) (E54.303.11)

Appears in the scientific report 2025
Click to display QR Code for this record

The record appears in these collections:
Document types > Presentations > Poster
Workflow collections > Public records
Institute Collections > JSC
Publications database

 Record created 2026-01-22, last modified 2026-02-20



Rate this document:

Rate this document:
1
2
3
 
(Not yet reviewed)