Methods for Dynamic and Spatio-temporal Data

With technological advancements, the quantity and quality of animal movement data has increased greatly. Currently, there is no existing movement model that can be used to describe full year of migratory species data that leverages both individual movement data and species distribution data. Herein we propose a full-year stochastic differential equation model for jointly modeling both individual movement data and species distribution data. We show that this joint model, under certain assumptions, results in efficient computation of the spatio-temporal dynamics of the entire population, and thus provides straightforward inference on the species distribution data. We illustrate this model with 215 bird-years of golden eagle movement in western North America and data from eBird for the species distribution. 

Standard CNNs inherently build hierarchical representations, yet their fixed receptive fields and pooling operations can limit effective multi-resolution trend capture, often missing large scale trends in favor of overfitting and finding false signal from local features. We propose an agent-driven, progressive training method that overcomes these limitations by dynamically adapting the model to explicitly integrate multi-resolution inputs. Starting with a low-resolution base model that captures coarse global features, an adaptive agent monitors performance metrics to decide when to incorporate higher-resolution inputs and add additional layers. Applied to Sea Surface Temperature data for El Niño prediction, our method effectively balances global context with local detail, leading to improved predictive performance. We further evaluate our approach on a traditional image detection dataset and compare it with existing multi-resolution techniques in CNNs, such as dilated convolutions, skip connections, and feature pyramid networks. 

Volcanic eruption prediction remains a major challenge due to sparse historical data and complex geological patterns. While most current approaches rely on real-time monitoring without fully using eruption history, survival analysis—a tool for time-to-event prediction—remains underexplored in this domain. To address this gap, we propose an extended Cox proportional hazard model that integrates volcano features, eruption history, and geological information. Key advances include bridging survival analysis with volcanology, dynamic short-term risk adjustment via a self-exciting mechanism, and integration of geological and spatial covariates via dimensionality reduction. We evaluate our model on a global eruption dataset, demonstrating its effectiveness and superiority over baseline methods. 

Neural networks typically assume statistically independent observed responses. However, survival time and failure status of GPUs are known to exhibit dependencies related to location information. We propose a deep learning approach that incorporates random-effect embeddings to model GPU failure outcomes. By assigning each GPU location a learnable embedding with imposed spatial structures, the model captures location-specific dependencies in both survival time and failure type predictions. We distinguish between physical locations (the row, column, slot, node, and cage of the GPU) and logical locations (how GPUs are interconnected through wired connections). By imposing correlation structures based on both physical and logical distances, the embeddings effectively capture strong correlations, particularly among GPUs with few intervening links. Our approach demonstrates improvements over previous deep learning models that did not incorporate spatial structure, and we present comparisons with other machine learning and parametric models. 

Objective: The primary objective of this research is to examine how sleep apnea, measured by the apnea-hypopnea index (AHI) and the percentage of slow wave sleep, varies across different seasons. We hypothesize that sleep parameters can be seasonal.
Methods: In this study, we analyzed diagnostic in-laboratory polysomnography (PSG) data from 27,760 patients from sleep studies. We obtained the apnea-hypopnea index (AHI) for each patient, measured as the number of apnea/hypopnea events per hour during sleep. We also obtained the percentage of sleep time each patient spent in slow wave sleep (SWS) (also known as the deepest non-rapid eye movement sleep). We applied a generalized linear model with harmonic terms to model the seasonality of the data adjusting for covariates age and body mass index. We analyzed data from men and women separately to see if sex differences could be found.
Results: AHI is higher in spring/winter and lower in fall, with men generally exhibiting a higher AHI than women across all seasons. Conversely, the proportion of SWS was greater in the fall and lower in the spring/winter, with women having a higher proportion of SWS than men across all seasons. Significant seasonality of AHI was found for both men (P=0.0001) and women (P=0.0053). Similarly, significant seasonality of SWS was also observed for women (P=0.0099), but interestingly not for men (P=0.13).
Conclusion and Discussion: Clear seasonal effects have been identified for both the AHI and the percentage of SWS outcome variables. The fall appears to be the best season for sleep quality (with the lowest sleep apnea and the highest proportion of SWS), whereas the winter and spring appear to be characterized by more sleep disturbance. The consistency in our findings between men and women strongly suggests that the seasonal effects on sleep are real. We suspect that spring allergy and cold/flu season in the winter may contribute to the seasonal differences. SWS is essential for cells repair and regenerate. Disruption in SWS is linked to an increased risk of metabolic disorders, such as type 2 diabetes. Recently, a few high-profile publications demonstrated that SWS contributes to the homeostasis of various physiological processes, including heart recovery after a myocardial infarction and the brain waste clearance. Our novel findings on the seasonality of sleep characteristics will be informative to clinicians and sleep researchers for assessing seasonal sleep health.
 

Dendrometers are small devices attached to trees which measure stem radius with high precision. These can capture long-term growth, diurnal cycles, and more importantly sustained stem shrinkage due to lack of water. Thanks to a unique network of such dendrometers in Switzerland (TreeNet), which operates continuously, we can model tree water deficit (TWD) in space-time. Crucial for drought monitoring efforts is then the creation of TWD maps at the national scale. However, dendrometers are installed on trees only at a few sites, with a low spatial coverage, which makes spatial prediction a challenging extrapolation task. But at each site a high temporal frequency is available to model TWD as a response to climatic and environmental variables. We leverage this and design a recurrent neural network architecture for joint modeling of multiple TWD series. This allows to not only forecast well in time but also to predict/extrapolate in space.