Invited speakers

 Sina Mews, Statistics and Data Analysis Group, University of Bielefeld (Germany)

Title.  Latent Markov Models in Ecology

Abstract. Latent (or hidden) Markov models are powerful tools for analysing time series or other sequential data that depend on underlying but unobserved states. Due to their flexible hierarchical framework, which separates the noisy observation process from the latent Markovian state process, they have gained prominence across numerous empirical disciplines. In ecology, in particular, they have become immensely popular, as ecological data collected over time are often characterised by indirect and incomplete observations linked to latent states – such as animal behaviours that cannot be directly measured. For instance, in movement ecology, observed step lengths and turning angles are driven by an animal’s hidden behavioural modes, such as resting, foraging, and travelling. As such underlying behaviours are of particular interest in animal ecology, I will focus on latent Markov models with discrete (rather than continuous) state spaces in my talk, specifically (discrete- and continuous-time) hidden Markov models and Markov-modulated Poisson processes. In particular, I will present three ecological applications: modelling diel activity patterns of fruit flies, surfacing times of minke whales, and capture-recapture data on bottlenose dolphin movement along the Scottish east coast. While the corresponding model formulations differ – particularly in their treatment of time – they share the same general structure, allowing for the same inferential methods to be applied and providing a unified framework for latent Markov models.

Silvia Pandolfi, Università degli Studi di Perugia (Italy)

Title. Hidden Markov models for longitudinal data: advances to deal with missing data, dropout, and variable selection

Abstract. Hidden Markov (HM) models may be usefully applied for the analysis of longitudinal data, as they deal with time dependence in a flexible way and allow us to perform a dynamic model-based clustering. We illustrate methodological and inferential advances to address the problem of missing responses and dropout, which typically arise in the context of longitudinal data. Additionally, we perform variable selection to reduce model complexity. The proposed approach is based on an HM model for multivariate continuous responses, according to which, given the latent state, the response variables are assumed to follow a multivariate Gaussian distribution with state-specific parameters. The method explicitly accounts for diFerent patterns of missing data, that is, the intermittent pattern, treated under the missing at random assumption, and the monotone pattern also known as informative dropout. To address the latter, we introduce an extra absorbing state in the latent state space. Furthermore, we propose an approach for simultaneously perform variable and model selection to choose the optimal number of informative variables and clusters (or states) by relying on a greedy search algorithm. The objective is that of obtaining a parsimonious model that provides more stable parameter estimates and enhances interpretability, particularly for high-dimensional data. We illustrate the proposed methodology through two applications: one based on historical data about primary biliary cholangitis and another using macroeconomic data referred to socioeconomic indicators.

Vasileios Papageorgiou, Aristotle University of Thessaloniki (Greece)

Title. TBA

 Abstract. The modeling of epidemic dynamics is essential in the field of epidemiology, as new methods offer better estimates of disease spread. Deterministic approaches often fail to fully capture epidemic evolution due to the inherent uncertainty in these processes. A shift to a stochastic perspective is therefore necessary, explored here through three fundamental epidemiological models. A stochastic SIRD (Susceptible (S), Infectious (I), Recovered (R), Deceased (D)) model with imperfect immunity is presented, based on a continuous-time Markov chain. Formulas and algorithms are developed to compute probabilities and moments of time-related quantities, such as disease extinction time, alarm time, and the infection time of a susceptible individual. Sensitivity analysis is then used to examine the behavior of these quantities with respect to the parameters of the stochastic model. A Markovian SIHRD (Susceptible (S), Infectious (I), Hospitalized (H), Recovered (R), Deceased (D)) model is also introduced to analyze hospitalizations, focusing on size-related characteristics like total and maximum hospitalizations, and their joint distribution with infections. These analytical tools are then combined with dynamic parameter estimation via particle filtering, using a SPIR (Susceptible (S), Presymptomatic (P), Infectious (I), Recovered-Deceased (R)) model to capture time-varying epidemic behavior. This includes tracking the number of infections until extinction, the timing of specific death counts, and infections caused by presymptomatic or infectious individuals. Applied to monkeypox (mpox) data from Ghana, the method produces more accurate estimates than fixed-parameter models. Sensitivity analysis further reveals how infectious and presymptomatic individuals influence transmission.

Reza Skandari, Imperial College Business School (UK)

Title. Balancing Complexity and Personalization in Sequential Decision Models for Chronic Disease Management

 Abstract. Markov Decision Processes (MDPs) and their partially observable counterparts (POMDPs) are widely used in healthcare for decision-making across the continuum of care, from screening and diagnosis to treatment and long-term monitoring. These models provide a structured framework for integrating clinical data, patient heterogeneity, and uncertainty into medical decisions. However, healthcare systems face a critical trade-off: while fully personalized policies optimize patient-specific outcomes, their exponential complexity makes them impractical. Conversely, one-size-fits-all approaches, though easier to implement in real-world settings, often lead to inefficiencies and disparities in patient care. This talk explores how machine learning and optimization can help bridge this gap by determining the optimal size and configuration of population stratification. By clustering patients based on key characteristics, we can design equitable and scalable policies that improve health outcomes while remaining computationally and practically feasible. We will discuss recent advancements at the intersection of machine learning, optimization, and healthcare decision science, highlighting key methodological challenges and open questions. To illustrate these concepts, we present a case study on cancer surveillance, demonstrating how stratifying patients based on a small number of meaningful characteristics can significantly enhance equity and patient outcomes compared to a one-size-fits-all policy—without requiring fully personalized approaches. This work underscores the potential of integrating sequential decision-making frameworks with data-driven population stratification to improve healthcare delivery.