Over the past decade, biostatistics has undergone significant advancements, driven by the increasing availability of complex data, the emergence of novel analytical methods, and the growing demand for robust and reproducible results in health sciences. This communication explores key developments in the field and how biostatistics has evolved to meet modern scientific challenges such as, high-dimensionality and data heterogeneity, related to the explosion of big data and the increasing complexity of biomedical research.

Abstract: TBA

A perspective will be offered on the profession of the biometrician, the biostatistician, and more generally the applied statistical scientist, in a continually and rapidly changing environment. The specifics of working in a multi-disciplinary environment will be discussed, referring to collaboration with agronomists, biologists, epidemiologists, medical professionals, etc. At the same time, interactions with other semi- or fully quantitative fields will be touched upon, such as computational biologists, computer scientists, engineers, etc. The current-day (r)evolution towards data science will be placed against a historical timeline of our field, which saw, over a relatively brief period of just one century, the coming of epidemiology and observational studies, (statistical) genetics, bioinformatics, the omics, big data, data science, data analytics, atrificial intelligence, etc. Historical notes related to the international evolution of our field, with particular emphasis on the hispanic world, will be offered.

Joint modeling of longitudinal and survival (JM-LS) data allow the inclusion of longitudinal information in survival models, as well as the addition of missing data processes in longitudinal studies. These models are very attractive from a methodological point of view and very valuable in biomedical studies. They were also a point of union between a group of researchers from VABAR (València Bayesian Research Group) and GRBIO, who started a joint task of studying these models: we learned a lot and had more fun. We present a Bayesian JM-LS which accounts for longitudinal continuous information in the unit interval and ordinal longitudinal covariates to learn about competing risk models and discuss their application to a Heart Failure (HF) study where patients underwent cardiac resynchronization therapy.

Sports Analytics have rapidly grown, offering new avenues for research and applications in performance improvement, injury prevention, and game strategy. This talk explores the evolution and current impact of Sports Analytics while emphasizing the untapped potential of interdisciplinary networks among statisticians and researchers. Such collaborations offer unique opportunities to tackle complex challenges, innovate methodologies, and uncover new insights. Through practical examples, we will illustrate how data-driven approaches are transforming sports science and highlight the future possibilities this growing field holds for research and application.

Integrating high-dimensional multi-omics data is essential for understanding the different layers of biological control. Single-omics methods offer useful insights but often miss the complex relationships between genes, proteins, and metabolites. In this talk, I will present GAUDI (Group Aggregation via UMAP Data Integration), a non-linear, unsupervised method that uses independent UMAP embeddings to analyze multiple data types together. GAUDI reveals relationships across omics layers better than several current methods. It not only clusters samples by their multi-omics profiles but also identifies key features contributing to each cluster, providing clear and interpretable visualizations. I will discuss how GAUDI enables researchers to identify meaningful patterns and potential biomarkers across diverse omics types.

In ecology and environmental sciences, combining diverse datasets has become an essential tool for managing the increasing complexity and volume of ecological data. However, as data complexity and volume grow, the computational demands of previously proposed models for data integration escalate, creating significant challenges for practical implementation. This study introduces a sequential consensus Bayesian inference procedure designed to offer the flexibility of integrated models while significantly reducing computational costs. The method is based on sequentially updating some model parameters and hyperparameters, and combining information about random effects after the sequential procedure is complete. The implementation of the approach is provided through two different algorithms. The strengths, limitations, and practical use of the method are explained and discussed throughout the methodology and examples. Finally, we demonstrate the method's performance using three different examples—one simulated and two with real ecological data—highlighting its strengths and limitations in practical ecological and environmental applications.

Personalized medicine, a promising branch of modern healthcare, has been made possible by the rapid development of next-generation sequencing (NGS), which has revolutionized genetic diagnostics and provided unprecedented opportunities for tailored treatments. However, the clinical utility of NGS remains constrained by the challenge of interpreting the impact of the genetic variants it uncovers. A significant portion of these variants remains classified as Variants of Uncertain Significance (VUS), undermining their clinical utility and creating anxiety for patients and their families. This situation has driven the development of computational pathogenicity predictors, machine learning tools trained to produce binary classifications—benign or pathogenic—of variants. While these methods have been integrated into clinical workflows, their accuracy and interpretability still fall short of meeting the stringent requirements of medical applications. In this context, recent years have witnessed a paradigm shift toward continuous prediction models, which aim to provide more precise quantitative assessments of variant impacts on protein function. These approaches leverage a combination of technologies that include data from deep mutational scanning experiments and machine learning techniques. By moving beyond binary labels, continuous predictors hold the promise of elucidating critical aspects of variant effects, such as disease severity and therapeutic response, thereby enhancing their relevance for clinical decision-making in precision medicine. This talk will explore the current state of methodologies to estimate the impact of protein variants, focusing on an original approach developed in our group to address the problem of using a small amount of protein-specific datasets to generate predictions for any protein, combining regression models and an ensemble-based approach. I will discuss, among other things, the results obtained both in rigorous validation experiments as well as in our participation in the CAGI5 and CAGI6 challenges, comparing our performance with that of other methods in the field.

Cancer progression and evolutionary accumulation models have been developed to discover dependencies in the irreversible acquisition of binary traits (e.g., mutations) from cross-sectional data. They have been used in computational oncology and virology but also in widely different problems such as malaria progression. Some of these methods have been applied to data with phylogenetic and longitudinal dependencies in questions including tool acquisition in animals and antimicrobial resistance in tuberculosis. Because of their interest, new methods continue to be developed. These tools have been used to make predictions about future and unobserved states of the system, identify different routes of, and dependencies in, feature acquisition in subsets of the data, and improve patient stratification and survival prediction based on the evolutionary trajectories and denoising of the data. The rich variety of available models increases their utility as markedly different dependency structures can be compared on the same data. These methods also hold promise to help identify therapeutic targets and improve evolutionary-based treatment approaches. I will first give an overview of the available methods. Then, using fitness landscapes, and discussing the conflation of lines of descent, path of the maximum, and mutational profiles, I will focus on how and why inferences might not be about the processes we intend, in particular under bulk sequencing. I will comment on major research opportunities, including translational uses, identifying dependencies that derive from frequency-dependent selection, and the relationship of these methods with phylogenetic comparative methods.

Overall mortality trends are the summation of cause-specific mortality experiences. Consequently modelling and forecasting changes in cause of death patterns allows us to recognize the drivers of all-cause mortality and identify emerging health challenges. When dealing with cause-specific mortality, we need to ensure that cause specific deaths must sum to the total number of deaths. We propose a simple and fast method to obtain coherent cause-specific mortality trajectories based on Lagrange multipliers and penalized splines. We apply the method proposed to fit and forecast mortality of males in the USA for the five leading causes of death.

Nowadays we are increasingly able to collect more complex data, and many challenges in data analysis stem from that complexity. The progression from a single numerical value as the unit of study, to a multivariate vector, then to a functional curve, or even to a skeletal shape representation, illustrates this evolution. In other words, there has been a shift from using large sample sizes in low-dimensional spaces to using relatively small ones in high-dimensional spaces. The perspective offered by functional data analysis (FDA) often provides a framework that allows the analysis of curves, images, or functions in high dimensions overcoming the problem of high dimensionality. For this reason, FDA has started to appear in the computational and bioinformatics literature over the last years. On the other hand, fuzzy classification assigns a degree of membership to each unit, often used in disease diagnosis to classify patients based on medical data and with artificial intelligence techniques to address uncertainty in diagnosis. Using the COVID-19 Raman spectroscopy data set we show the usefulness of combining functional data analysis and the distance-based fuzzy classifier FC-DF highlighting their strengths and limitations.

In rare diseases, single-arm, non-randomised, open-label trials are frequently conducted, mainly due to ethical reasons or the study being unfeasible as patients reject to participate. However, there are some inherent limitations in this type of designs, for example, time-to-event endpoints and patient reported outcomes are not interpretable without a control arm in the study. There are other circumstances, where a randomised control trial is doable but the number of subjects in the control arm are insufficient. The use of external data (clinical trial data or real-world data) appears as a way to overcome these limitations and improve the efficiency of clinical trials. A critical step in bringing external data is to ensure that the external data is comparable to the study population in terms of study entry criteria, in particular to measured baseline prognostic/ confounding variables. Ideally, both external data and study population should be exchangeable with each other. There are several frequentist methodologies to adjust for differences in baseline prognostic/ confounding factors, such as, the propensity scores (Rosenbaum and Rubin, 1983) based on matching, stratification, inverse probability of treatment weights, or covariate adjustment on propensity score methods. These methods balance the prognostic factors, then the comparison of outcomes between the treatment groups yields an unbiased treatment effect estimate, as long as all the confounding variables are included in the propensity score model. Also, Bayesian methods have been developed to borrow information from external data by creating an informative prior distribution. The prior can be derived based on different approaches such as the meta analytic predictive method. It is important to note that the type I error may be inflated by incorporating external data as a nonrandomised comparison may introduce bias due to unmeasured confounding covariates. Therefore, simulations should be carried out to evaluate the operating characteristics when including external data. Regulatory agencies have not ignored this situation and have taken some initiatives and released corresponding guidance with recommendations when designing externally controlled clinical trials. However, the use of external controls is not mature enough yet and interactions with regulatory agencies are advisable at the time of the study design.

The World Health Organization defines health as a complete physical, men- tal, and social well-being and not merely the absence of disease or infirmity. In this sense, patient-reported outcomes (PRO) are becoming primary outcome measurements in observational and experimental studies, as they capture evidence of patients’ status that is difficult to evaluate physically, such as pain, quality of life or, satisfaction with care. PRO are usually obtained using item-based questionnaires, assigning scores to each item response and summing the scores across a group of items to create overall scores, usually called dimensions, which decompose the health aspect they are evaluating. The binomial distribution is the most common candidate when modeling discrete and bounded outcomes, such as PRO dimensions. However, the fact that questionnaire items are answered by the same individuals sets up a correlation structure in the ordinal responses that constitute the final score, which increases the variability beyond the mean-variance structure of the binomial distribution, a property called overdispersion. In fact, PRO scores tend to have skewed distributions, often showing U, J or J-inverse shapes. In this talk, we are going to present the main contributions of our research group in the field of PRO modeling, from the proposal of an optimal probability distribution to a joint model for the analysis of longitudinal PRO and survival data. Additionally, we will present the most clinically significant results obtained from applying the developed models to a health-related quality of life study in patients with Chronic Obstructive Pulmonary Disease (COPD).

Capture-recapture methods are commonly used in ecology to estimate animal population sizes and species richness. These methods have become popular, not only in ecology but also in social and medical sciences, to estimate the size of elusive populations such as illegal immigrants, illicit drug users, or people having a drinking problem. The talk will address a new non-parametric approach for estimating the population size when we only know how many animals or individuals were observed once, twice, ... , as well as how many animals or individuals were observed r or more times (right censoring pattern). Similar to the Chao estimator, the method provides a lower bound on population size as well as bootstrap confidence intervals. The particular case of censoring at r=2 will be studied in detail, along with several applications in ecological and social sciences.

The receiver operating characteristic (ROC) curve is widely used to assess the accuracy of continuous biomarkers for binary outcomes (e.g., healthy and diseased). However, evaluating the impact of additional patient or environmental information on diagnostic accuracy is also important. Furthermore, studies often focus on prognosis rather than diagnosis, especially in survival analysis, where outcomes evolve over time (e.g., alive and death). To assess the accuracy of continuous prognostic biomarkers for time-varying outcomes, time-dependent extensions of the ROC curve have been proposed. This work introduces a novel penalised-based estimator of the cumulative-dynamic time-dependent ROC curve, which accounts for the potential modifying effects of covariates on biomarker accuracy. Building on previous approaches, we adopt a modelling framework that considers flexible models for the conditional hazard function and the biomarker, allowing for the accommodation of non-proportional hazards and nonlinear effects through penalised splines, thus addressing the limitations of earlier methods. We apply our method to evaluate the ability of the Global Registry of Acute Coronary Events (GRACE) risk score to predict mortality after discharge in patients who have experienced acute coronary syndrome, and how this ability may vary with left ventricular ejection fraction.

Multi-omic experiments offer an unprecedented opportunity to explore gene expression regulation, providing deep insights into the intricate regulatory mechanisms of biological systems. However, the high dimensionality, heterogeneity, and multicollinearity of multi-omic datasets present significant challenges for statistical modeling and variable selection when inferring regulatory networks. Additionally, most existing tools for multi-omic regulatory network inference either fail to accommodate diverse omic modalities or lack the ability to generate and compare phenotype-specific networks. To address these limitations, we developed MORE (Multi-Omics Regulation), a novel methodology that leverages regression-based frameworks and advanced variable selection strategies to construct phenotype-specific regulatory networks across any number or type of omic data. MORE integrates prior regulatory knowledge and offers functionalities for systematic comparison of the resulting networks. We benchmarked MORE against other state-of-the-art tools using simulated datasets and applied it to an ovarian cancer case study. Our results demonstrate the robustness and versatility of MORE in unraveling regulatory mechanisms in complex biological systems, underscoring its potential as a valuable resource for multi-omic data analysis.

In this talk I will present a new general strategy for goodness-of-fit testing with survival data. The setting is that of testing for a parametric family of distribution functions when the data are deteriorated due to random censoring and/or random truncation. A key step is the characterization of the null hypothesis through a moment equation which involves the estimation of the observable distribution under both the null and the alternative. A new omnibus test will be proposed, and its theoretical properties will be presented. Particular applications include, but are not limited to, right-censored data, left-, right- or doubly-truncated data, or interval censored data. Advantages with respect to existing methods will be discussed. The finite sample performance of the test will be investigated through simulations. Illustrative real data analyses will be given. This is joint work with Juan Carlos Escanciano.

This talk will delve into biostatistical strategies advancing the field of genetic neuropidemiology. Recent advancements have enabled more precise identification of genetic and environmental factors, significantly enhancing brain health, risk stratification, disease prediction, and prevention strategies. Key highlights include applying multivariate models to extensive genomic, environmental, and brain imaging datasets, and the assessment and implementation of statistical tools designed for data integration. These methods further emphasize incorporating diversity and sex-specific mechanisms into study populations, bolstering the applicability and accuracy of our findings. The implications of these advances extend beyond improved diagnostic accuracy, paving the way for potential biological pathways that support personalized medicine, prevention, and targeted therapeutic interventions.

The motivation for my PhD arises from clinical data from the DIVINE project where patients hospitalized due to COVID-19 are followed through several states. One of the aims of this project was to analyze the evolution of those patients and for that a complex multi-state model (MSM) was designed. This MSM allows us to analyze the risk factors for the different events of interest (e.g. non-invasive mechanical ventilation (NIMV), invasive mechanical ventilation (IMV), or death) as well as to predict the course of the disease for new patients, but we realized that we didn't know how to analyze its predictive capacity.

Therefore, the main objective of my PhD is to evaluate the discriminative ability for MSM, and for that, the area under the time-dependent ROC curve (\( AUC(t) \)) can be used. In this work, we focus initially on those patients with severe pneumonia who can transition to two competing events: the need for NIMV or IMV; and we propose an estimator for the global \( AUC(t) \) for a competing risk model.

Under competing risk models, different estimators can be used to estimate the (partial) \( AUC(t) \) of each transition (\( AUC_k(t), k=1,2 \)). In this work, we propose an estimator \( \widehat{AUC}_{CR}(t) \) for the global \( AUC(t) \) (\( AUC_{CR}(t) \)) for a competing risk model as a weighted sum of \( \widehat{AUC}_k(t), k=1,2 \) with each \( AUC_k(t) \) being weighted by the probability of experiencing that event \( k \) before time \( t \). We have proved that \( \widehat{AUC}_{CR}(t) \) is consistent and asymptotically normal.

In the current era, omics technologies such as high-throughput experiments have significantly transformed the fields of biology and medicine. These advances enable the generation of large volumes of biological data, such as gene lists, proteins, and other biological features, under different experimental conditions. Although getting this large amount of information represents a breakthrough, it is crucial to develop appropriate statistical methods to analyze and extract knowledge from these data. In this context, the present study proposes a statistical method based on an equivalence hypothesis test to evaluate biological similarity between feature lists. The central idea is that two or more feature lists can be considered biologically similar if they share a significant proportion of enriched GO terms. First, the choice of the Sorensen index is justified as an appropriate metric for assessing the dissimilarity of joint enrichment between the lists under comparison. Next, the sampling distribution of this measure is studied both theoretically and through approximation using the Bootstrap method, which proves to be particularly effective when the enrichment level is low. Based on these distributions, an equivalence hypothesis test is developed, along with its corresponding irrelevance threshold, which is less arbitrary than the thresholds commonly used in equivalence approaches. Furthermore, the R package goSorensen has been developed, published, and is available on the Bioconductor platform. This informatics tool allows for the efficient application of the proposed methodology. Additionally, a dissimilarity matrix is constructed based on the irrelevance threshold, which defines when two lists are significantly equivalent. This matrix provides an inferential measure of how close or distant the compared lists are from each other. The graphical representation and interpretation of this matrix, such as in an MDS-Biplot, is useful for identifying the GO terms associated with the formation of equivalence between lists. Finally, it is important to note that the proposed methodology has been rigorously evaluated and applied to real gene lists, with an exhaustive comparison of the results obtained against other similar comparison methods.

Platform trials enhance drug development by offering increased flexibility and efficiency. They evaluate the efficacy of multiple treatment arms, with the added benefit of permitting treatment arms to enter the trial over time and to stop early based on interim data. Efficacy is usually assessed using a shared control arm. For arms entering later, the control data is divided into concurrent and non-concurrent controls (NCC), referring to control patients recruited while the given treatment arm is in the platform and before it enters, respectively. Including NCC can reduce the sample size and increase power, but also lead to bias in the effect estimates, if there are time trends. For platform trials with continuous endpoints without interim analyses, a regression model has been proposed that utilizes NCC and adjusts for time trends by including the factor “period” as a fixed effect. Here, periods are defined as time intervals bounded by any treatment arm entering or leaving the platform. It was shown that this model leads to unbiased effect estimates and asymptotically controls the type I error rate regardless of the time trend pattern, if the time trend affects all arms in the trial equally and is additive on the model scale. However, if interim analyses are included, the definition of the factor periods becomes data dependent and the number of periods to adjust for depends on previous results. Furthermore, due to early stopping the sample sizes in different arms become outcome dependent, and therefore the effect estimates are no longer unbiased. This can affect the adjustment for time trends in the linear model, and the type I error rate might no longer be controlled. In this work, we examine the performance of the currently available model in group-sequential platform trials and show that it leads to a loss of the type I error rate control and bias in the effect estimators. In addition, we describe how the weight of the non-concurrent controls in the treatment effect estimator is stochastically dependent on the outcome in the non-concurrent controls. Moreover, we will investigate adjusted treatment effect estimators that aim to eliminate or reduce the potential bias and resulting type I error rate inflation. Focusing on a simple platform trial with two experimental treatment arms and a continuous endpoint, we will present results from a simulation study, where we evaluate the performance of the considered approaches and compare them to current methods.

Background and objective: From March 2020 to July 2022, 6 waves of the COVID-19 pandemic were registered in Spain. There are several studies comparing different COVID-19 waves but, as far as we know, none of them uses a matching procedure to make patients comparable or accounts for ceiling of care. Our aim is to compare in-hospital mortality across waves in patients with and without ceiling of care at hospital admission.

Methods: Data come from an observational study conducted during four waves of COVID-19 (March 2020-August 2021) in 5 hospitals in Catalonia. Three models were constructed to compare in-hospital mortality by wave: 1) a raw logistic model with only wave as a covariate; 2) a fully clinical adjusted logistic regression model with wave and patient baseline information as covariates and 3) a logistic model with weights obtained from an inverse probability weighting procedure to account for differences in baseline profile between waves. Models were presented stratified by ceiling of care. All analyses were conducted using R software version 4.3.0.

Results: A total of 3982 patients without ceiling of care and 1831 patients with ceiling of care were included. Patients with ceiling of care were, in median, 20 years older than patients without ceiling of care and in-hospital mortality ranged from 5\% to 45\%. The adjusted odds ratio (OR) of in-hospital mortality in the second wave were 0.57 (95\%CI 0.40 to 0.80), in the third 0.56 (95\%CI 0.37 to 0.84) and in the fourth 0.34 (95\%CI 0.21 to 0.56) compared with the first wave in subjects without ceiling of care. The adjusted odds ratio were significantly lower in the fourth (0.38 95\%CI 0.25 to 0.58) wave compared to the first wave in subjects with ceiling of care.

Discussion: The likely impact of the wave on in-hospital mortality differs between patients with and without ceiling of care. In patients without ceiling of care, mortality decreased over time which may be explained by better disease knowledge and management. In ceiling of care, only fourth-wave patients were less likely to die than first-wave patients. In a future infectious disease pandemic, it will be a challenge to improve the management of patients with ceiling of care.

The development of methods to address censored covariates has gained signif- icant attention in recent years. Although the problem itself is not new, its presence in real-world data has often been overlooked. Interval-censored covari- ate data, in particular, is frequently replaced by a single imputed value, which is known to introduce bias and underestimate variance. While recent methods have emerged for handling discrete time-to-event covariates, these approaches are often limited to survival analysis contexts, leaving other applications unad- dressed. In this talk, we shift our focus to analytical chemistry, specifically to data re- lated to the quantification of compounds in mixtures. Compounds are often defined by multiple analytes, each measured via liquid chromatography and subject to analyte-specific detection and quantification limits. This chemical technique results in interval-censored data for the overall quantity of a com- pound. Our motivating example originates from metabolomics, exploring the association between circulating carotenoids—molecules present in the blood- stream—and cardiometabolic health. Advancing research in this area requires fitting generalized linear models for cardiometabolic biomarkers while incorpo- rating interval-censored circulating carotenoid levels as a covariate. Building on this example, we present an extension of the GEL algorithm, which was originally developed for time-to-event interval-censored covariates in linear models. The GEL algorithm is an EM-type method that alternates between estimating the distribution of the censored covariate and maximizing the model’s likelihood function. However, like other recent approaches in the literature, it relies heavily on the assumption that the censored covariate has a discrete support, which limits its applicability. Our extension overcomes this limitation by handling interval-censored covariates nonparametrically and regardless of the distribution’s support, broadening its usability to a wide range of applications.