Multivariate spatio-temporal models with latent dynamics for cardiovascular disease prediction with heterogeneous factors
Description: Out-of-hospital cardiac arrest (OHCA), acute coronary syndrome (ACS), and stroke are the leading causes of out-of-hospital deaths. These are all time-sensitive cardiovascular diseases (CVDs), i.e. prompt access to the patient and early intervention significantly affect immediate outcome, survival, and neurological recovery. The main objective of this research is to develop novel realistic and interpretable statistical models and algorithms that, integrating diverse sources of data, aim to reduce the time to intervention of first responders and ambulance arrival, ultimately leading to a significant mortality reduction and improved neurological outcomes. This wide-ranging goal will be achieved by identifying, thanks to the developed methodologies, high-risk areas and by timely predicting, with accuracy and uncertainty quantification, areas of expected future higher risk. More specifically, we will first estimate and predict the OHCA risk for the municipalities of Ticino by designing a model which accounts for temporal and spatial heterogeneity, space-time interactions, inflation of zero events, demographic features, climate and environmental factors, and with careful prior elicitation. The scalable Bayesian space-time statistical models will be calibrated on different (large and heterogeneous) data sources and will be estimated by the INLA methodology. Furthermore, we suggest the methodological extension of the space-time model to a hidden Markov random field of dynamically varying dimension, with the aim of clustering municipalities over time in a reduced number of groups having comparable cardiovascular risk level. Subsequently, we will extend the model to jointly include other time-sensitive CVDs such as ACS and stroke. Because OHCA may be the presenting disease of ACS or of a massive stroke event, the question arises as of whether the spatio-temporal distributions of ACS and stroke are similar to the distribution of OHCA, and if ACS or stroke events can be used as proxy to forecast OHCA. The three series will be jointly modeled and predicted through a novel Bayesian spatio-temporal multivariate model with shared spatial and temporal patterns and with idiosyncratic patterns that divert from commonalities. An important by-product will therefore be a posterior estimate of the level of spatial and temporal dependence among the three diseases. Model parameters will be optimally estimated to provide cardiovascular risk heat-maps over the Ticino area, updated every week in order to account for the varying weather and pollution conditions and for new cardiovascular events. The heat-maps - in the same spirit as, for instance, crime maps - will be made available to decision makers, such as the local emergency medical systems, with the aim of optimizing the surveillance of hot areas (by police and first responders) and improving ambulance displacement. Moreover results and analyses will be made available to the public with software packages and with an interacting web app for easy (re)use by non-experts. Fast prediction is made feasible by efficient INLA algorithms and by combinations of Markov Chain Monte Carlo techniques. There are novel aspects in our statistical models. First, the multivariate Bayesian spatio-temporal model for the joint analysis of OHCA, ACS and stroke is new and clinically relevant providing posterior estimates of the spatial and temporal dependence among the three diseases. Second, a hidden Markov model approach will include for the first time latent states within a spatio-temporal model, for a dynamic cluster of municipalities with comparable cardiovascular risk level.Third, we combine several sources of heterogeneous information extracted from demographic, socio-economic, meteorological, and environmental data, for a more comprehensive understanding of the driving forces behind OHCA, ACS, and stroke spatio-temporal distributions. This complex system of heterogeneous datasets is a challenging large-scale environment for the implementation of the INLA methodology as currently developed. Finally, our methodology not only may be extended at national level - for which Ticino will provide a first innovative pilot project - but also used for other diseases for which a registry is available (e.g. cancer) as well as for other applications where the time-space evolution of multiple time series is of interest as in finance/economics (to study e.g. defaults of companies).
Amount: CHF 781,600
Funding type: research
Funder scheme: Project funding
Start date: 2021-05-01
End date: 2026-04-30
Investigator: Mira, Antonietta
Institution: Università della Svizzera italiana – USI
Funded outputs count: 11
Funder award ID: 200557
OpenAlex ID: G688092748