Calibration of grey-box simulation models is a constrained optimization problem in which model evaluations are expensive, the parameter space can be high-dimensional, and the search must respect plausibility constraints. Although the simulation code is fully available to the analyst, the joint effect of multiple parameters remains difficult to predict analytically. Classical optimizers such as Nelder–Mead (NM) are simple to deploy but sample-inefficient, particularly under constraints. Modern Bayesian Optimization methods achieve competitive solutions with far fewer evaluations but require non-trivial modeling machinery for constraint handling. We introduce an agentic calibration method in which a large language model acts as the optimizer, with constraints incorporated as a plain-language section of the system prompt. We evaluate the agentic method, NM, and Bayesian Optimization (BO) on an anal cancer simulation model under both unconstrained and clinically constrained calibration. Under unconstrained calibration, the agentic method achieves substantially lower best error than BO and NM, while requiring fewer model evaluations. Under constrained calibration, the agentic method reaches comparable error levels and both outperform NM. These results are obtained at the cost of increased inference time per iteration. Agentic calibration achieves competitive performance with substantially fewer model evaluations, and constraint handling is essentially free at the modeller-facing interface through simple textual specifications rather than additional modelling machinery. The main trade-off lies in increased per-iteration inference cost, making the approach particularly suitable when simulation time dominates. Beyond performance, the per-iteration rationale makes the search auditable and explainable, so its decisions can be scrutinised and justified to third parties.

Quantum Federated Learning (QFL) is an emerging framework that integrates Federated Learning (FL) with Quantum Computing (QC), enabling collaborative training of quantum models across distributed and privacy-preserving environments. While prior work has primarily focused on classical machine learning tasks using Variational Quantum Circuits (VQCs), the application of QFL to inherently quantum problems remains largely unexplored. In this work, we investigate QFL in the context of one of the most fundamental tasks in quantum information theory, namely, Quantum State Discrimination (QSD). We design and evaluate a set of toy yet realistic scenarios involving small, imbalanced, and noisy quantum datasets, reflecting practical constraints common in quantum sensing and metrology. Our findings demonstrate that QFL can successfully overcome local dataset biases and quantum noise, achieving near-optimal performance comparable to analytical results. All code and experiments are made publicly available at https://github.com/roccobb/QFL4QSD to support reproducibility and encourage further research.

Cheese-making is a complex process involving numerous stages, with multiple factors contributing and complex interactions occurring among the physicochemical elements involved. Understanding the process and optimizing its stages has attracted the attention of numerous investigations. In recent years, Machine Learning (ML) has established itself as one of the most advanced tools for data analysis and modeling thanks to its ability to capture complex and non-linear patterns. In the area of food science and engineering, these algorithms have started to be used as an alternative to more traditional statistical and mathematical prediction models. This paper explores the main research on ML applied to the study of cheese, from its production stages (i.e., fermentation or coagulation process) to the final product (i.e., detection of adulterations or food fraud). Particularly, we review 42 papers published between January 2014 and January 2025, with the aim of identifying common approaches. First, we present an explanation of the main concepts required to bring these approaches closer to researchers who are not experienced in applying ML. Then, we analyze the selected publications to detail the tasks of interest and the algorithms proposed to solve them. Finally, we detect gaps and opportunities to incorporate ML into future cheese research.

Social media platforms offer an invaluable wealth of data to understand what is taking place in our society. However, social media data hides demographic biases related to characteristics such as gender or age. Therefore, considering social media data as representative of the population can lead to fallacious interpretations. For instance, in France in 2021, women represent 51.6% of the population1, whereas on Twitter they represent only 33.5% of French users2. With such a significant difference between social network user demographics and the actual population, detecting the gender or age before delving into a deeper analysis of social phenomena becomes a priority. In this paper, we tackle the gender detection problem on Twitter. We introduce miniAM2, which is an assemblage model of an enriched distillation with weak-supervised learning. Our contributions are threefold: (i) a novel multilingual model that outperforms existing models in both accuracy and speed, allowing for real-time gender detection and organization status on Twitter based on their name, screen_name, and description, making it lighter and faster than state-of-the-art; (ii) an innovative assemblage multi-language strategy that enriches a distillation process with weak-supervised learning using minimal annotated data, and (iii) a unique method to adapt the model to similar languages without requiring annotated data in the target language, which provides significant advancements in handling resource-poor languages in gender detection tasks. We provide our model on demand so social scientists can use it for their analysis.

The multinomial logistic regression (MLR) model is widely used in statistics and machine learning. On the one hand, stochastic gradient descent (SGD) is the most common approach for determining the parameters of a such model in big data scenarios, due to its simplicity and low computational complexity property. Furthermore, SGD has proven convergence under reasonable conditions. However, SGD has slow sub-linear rates of convergence and it often reduces convergence speed due to the plateau phenomenon. On the other hand, stochastic natural gradient descent (SNGD), proposed by Amari, is a manifold optimization method shown to be Fisher efficient when it converges, but its convergence properties remain unproven and it is often computationally prohibitive for models with a large number of parameters. Here, we propose dual stochastic natural gradient descent (DSNGD), a stochastic optimization method for MLR based on manifold optimization concepts. In the discrete scenario, DSNGD (i) has linear per-iteration computational complexity in the number of parameters, and (ii) is proven to converge. To achieve (i) we leverage the dual flatness of the family of joint distributions for MLR to simplify computations. To ensure (ii) DSNGD builds on the foundational ideas of convergent stochastic natural gradient descent (CSNGD), a variant of SNGD with guaranteed convergence, using an independent sequence to construct a bounded approximation of the natural gradient. By generalizing a result from Sunehag et al., we prove that DSNGD converges in the discrete case and maintains linear computational complexity per iteration. Beyond its convergence property and linear computational complexity, DSNGD empirically demonstrates fast convergence comparable to SNGD, improves upon SGD performance, and exhibits stability where SNGD does not.

The multinomial logistic regression (MLR) model is widely used in statistics and machine learning. Stochastic gradient descent (SGD) is the most common approach for determining the parameters of a MLR model in big data scenarios. However, SGD has slow sub-linear rates of convergence. A way to improve these rates of convergence is to use manifold optimization. Along this line, stochastic natural gradient descent (SNGD), proposed by Amari, was proven to be Fisher efficient when it converged. However, SNGD is not guaranteed to converge and it is computationally too expensive for MLR models with a large number of parameters. Here, we propose a stochastic optimization method for MLR based on manifold optimization concepts which (i) has per-iteration computational complexity is linear in the number of parameters and (ii) can be proven to converge. To achieve (i) we establish that the family of joint distributions for MLR is a dually flat manifold and we use that to speed up calculations. Sánchez-López and Cerquides have recently introduced convergent stochastic natural gradient descent (CSNGD), a variant of SNGD whose convergence is guaranteed. To obtain (ii) our algorithm uses the fundamental idea from CSNGD, thus relying on an independent sequence to build a bounded approximation of the natural gradient. We call the resulting algorithm dual stochastic natural gradient descent (DNSGD). By generalizing a result from Sunehag et al., we prove that DSNGD converges. Furthermore, we prove that the computational complexity of DSNGD iterations are linear on the number of variables of the model.

Federated learning (FL) is a recent technique that emerged to handle the vast amount of training data needed in machine learning algorithms while fulfilling data owners’ privacy challenges in such scenarios. Simultaneously, the field of quantum computing (QC), using quantum properties such as entanglement and superposition to perform computation, has experienced exponential growth, theoretically proving to be more efficient in specific machine learning tasks and creating the discipline known as quantum machine learning (QML). Thus, an emerging body of knowledge has started studying the combination of these two research agendas, giving rise to the field of quantum federated learning (QFL). In this review, we systematically classify the existing literature through a novel taxonomy, identify current trends and challenges, and highlight research gaps and future directions to support the continued development of this emerging field.

Cumulus expansion is an important indicator of oocyte maturation and has been suggested to be indicative of greater oocyte developmental capacity. Although multiple methods have been described to assess cumulus expansion, none of them is considered a gold standard. Additionally, these methods are subjective and time-consuming. In this manuscript, the reliability of three cumulus expansion measurement methods was assessed, and a deep learning model was created to automatically perform the measurement. Cumulus expansion of 232 cumulus-oocyte complexes was evaluated by three independent observers using three methods: (1) measurement of the cumulus area, (2) measurement of three distances between the zona pellucida and outer cumulus, and (3) scoring cumulus expansion on a 5-point Likert scale. The reliability of the methods was calculated in terms of intraclass-correlation coefficients (ICC) for both inter- and intra-observer agreements. The area method resulted in the best overall inter-observer agreement with an ICC of 0.89 versus 0.54 and 0.30 for the 3-distance and scoring methods, respectively. Therefore, the area method served as the base to create a deep learning model, AI-xpansion, which reaches a human-level performance in terms of average rank, bias and variance. To evaluate the accuracy of the methods, the results of cumulus expansion calculations were linked to embryonic development. Cumulus expansion had increased significantly in oocytes that achieved successful embryo development when measured by AI-xpansion, the area- or 3-distance method, while this was not the case for the scoring method. Measuring the area is the most reliable method to manually evaluate cumulus expansion, whilst deep learning automatically performs the calculation with human-level precision and high accuracy and could therefore be a valuable prospective tool for embryologists.

In emergency situations, social media platforms produce a vast amount of real-time data that holds immense value, particularly in the first 72 h following a disaster event. Despite previous efforts, efficiently determining the geographical location of images related to a new disaster remains an unresolved operational challenge. Currently, the state-of-the-art approach for dealing with these first response mapping is first filtering and then submitting the images to be geolocated to a volunteer crowd, assigning the images randomly to the volunteers. In this work, we extend our previous paper (Ballester et al., 2023) to explore the potential of artificial intelligence (AI) in aiding emergency responders and disaster relief organizations in geolocating social media images from a zone recently hit by a disaster. Our contributions include building two different models in which we try to (i) be able to learn volunteers’ error profiles and (ii) intelligently assign tasks to those volunteers who exhibit higher proficiency. Moreover, we present methods that outperform random allocation of tasks, analyze the effect on the models’ performance when varying numerous parameters, and show that for a given set of tasks and volunteers, we are able to process them with a significantly lower annotation budget, that is, we are able to make fewer volunteer solicitations without losing any quality on the final consensus.

A critical aspect of simulation models used in cost-effectiveness analysis lies in accurately representing the natural history of diseases, requiring parameters such as probabilities and disease burden rates. While most of these parameters can be sourced from scientific literature, they often require calibration to align with the model’s expected outcomes. Traditional optimization methods can be time-consuming and computationally expensive, as they often rely on simplistic heuristics that may not ensure feasible solutions. In this study, we explore using Bayesian optimization to enhance the calibration process by leveraging domain-specific knowledge and exploiting structural properties within the solution space. Specifically, we investigate the impact of additive kernel decomposition and a stepwise approach, which capitalizes on the sequential block structure inherent in simulation models. This approach breaks down large optimization problems into smaller ones without compromising solution quality. In some instances, parameters obtained using this methodology may exhibit less error than those derived from naive calibration techniques. We compare this approach with two state-of-the-art high-dimensional Bayesian Optimization techniques: SAASBO and BAxUS. Our findings demonstrate that Bayesian optimization significantly enhances the calibration process, resulting in faster convergence and improved solutions, particularly for larger simulation models. This improvement is most pronounced when combined with a stepwise calibration methodology.

Optimizing drug dosages is essential for effective treatment. Clinical protocols may not suit all types of patients evenly, due to many drug trials not being designed to account for all comorbities or clinically relevant outcomes. Methodologies to optimize drug policies with observational data exist, but struggle due to limited data completeness in clinical settings. Computational methods can help overcome these challenges by leveraging field knowledge. This paper proposes an Individualized Doser (IDoser), a core dosing model that links drug dose to relevant covariates via a set of coefficients and includes a loss function to code needed assumptions and requirements. Coordinate descent is used to obtain a fitted model with minimal loss. The loss function also measures performance when validating the model with unseen data. We validated the proposed approach using the case of follicle-stimulating hormone (FSH) dosing for controlled ovarian stimulation (COS). When compared to clinical practice, IDoser achieved a net improvement of up to 31.97% in the validation cases. We present a simple but effective method to bridge the gap between current clinical dosing policies and gold policies based on the true underlying and often unknown dose-response functions.

Background Adequately selecting the initial follicle-stimulating hormone (FSH) dose during controlled ovarian stimulation (COS) is key for success in assisted reproduction. The objective of COS is to obtain an optimal number of oocytes to increase the chances of achieving a pregnancy, while avoiding complications for the patient. Current clinical protocols do achieve good results for the majority of patients, but further refinements in individualized FSH dosing may reduce the risk of poor ovarian response while also limiting the risk of ovarian hyperstimulation syndrome (OHSS) risk. Models to select the first FSH dose in COS have been presented in literature with promising results. However, most have only been developed and tested in normo-ovulatory women under the age of 40 years. Methods This is a randomized, controlled, multicenter, single blinded, clinical trial. This study will be performed in 236 first cycle in vitro fertilization (IVF) and/or ICSI (intracytoplasmic sperm injection) patients, randomized 1:1 in two arms. In the intervention arm, the dose of FSH will be assigned by a machine learning (ML) model called IDoser, while in the control arm, the dose will be determined by the clinician following standard practice. Stratified block randomization will be carried out depending on the patient being classified as expected low responder, high responder, or normo-responder. Patients will complete their participation in the trial once the first embryo transfer result is known. The primary outcome of the study is the number of metaphase II (MII) oocytes retrieved at ovarian pick up (OPU) and the hypothesis of non-inferiority of the intervention arm compared to the control. Secondary outcomes include the number of cycle cancelations (due to low response or no retrieval of mature oocytes), risk of ovarian hyperstimulation syndrome (OHSS), and clinical pregnancy and live birth rates per first transfer. Discussion To our knowledge, this is the first randomized trial to test clinical performance of an all-patient inclusive model to select the first dose of FSH for COS. Prospective trials for machine learning (ML) models in healthcare are scarce but necessary for clinical application. Trial registration ClinicalTrials.gov, NCT05948293 . Registered on 14 July 2023.

The Maximum Boolean Satisfiability Problem (also known as the Max-SAT problem) is the problem of determining the maximum number of disjunctive clauses that can be satisfied (i.e., made true) by an assignment of truth values to the formula’s variables. This is a generalization of the well-known Boolean Satisfiability Problem (also known as the SAT problem), the first problem that was proven to be NP-complete. With the proliferation of quantum computing, a current approach to tackle this optimization problem is Quantum Annealing (QA). In this work, we compare several gadgets that translate 3-SAT problems into Quadratic Unconstrained Binary Optimization (QUBO) problems to be able to solve them in a quantum annealer. We show the performance superiority of the not-yet-considered gadgets in comparison to state-of-the-art approaches when solving random instances in D-Wave’s quantum annealer.

Abstract Cumulus expansion is an important indicator of oocyte maturation, often correlated with greater oocyte developmental capacity. Although multiple methods have been described to assess cumulus expansion, none of them is considered a gold standard. Additionally, these methods are subjective and time-consuming. Here, the reliability of three cumulus expansion measurement methods was evaluated and a deep learning model was created to automatically perform the measurement. Cumulus-oocyte complexes were compared before and after in vitro maturation by three independent observers using three methods: (1) measurement of the cumulus area, (2) measurement of three distances between the zona pellucida and outer cumulus, and (3) scoring cumulus expansion on a 5-point Likert scale. Inter- and intra-observer agreements were calculated using intraclass-correlation coefficients (ICC). The area method resulted in the best overall agreement with an ICC of 0.89 versus 0.54 and 0.30 for the 3-distance and scoring method, respectively. Therefore, the area method served as the base to create the deep learning model, which outperformed two observers while equivalent to the third. Measuring the area is the most reliable method to manually evaluate cumulus expansion, whilst deep learning automatically performs the calculation with human-level accuracy and could therefore be a valuable prospective tool for embryologists.

During disasters, social media can serve as a valuable source of real-time information about the impacts on people and infrastructure. However, due to the lack of geographical information in most social media posts, this information is often underutilized by first responders. Previous research has attempted to estimate the location of individual social media posts using text and image analysis, but limitations still exist in fine-grained disaster area mapping. To address this issue, this paper analyses the performance of combining text from social media post with textual information from the images on improving the geolocation of social media information during a disaster.

Social media have the potential to provide timely information about emergency situations and sudden events. However, finding relevant information among the millions of posts being added every day can be difficult, and in current approaches developing an automatic data analysis project requires time and technical skills. This work presents a new approach for the analysis of social media posts, based on configurable automatic classification combined with Citizen Science methodologies. The process is facilitated by a set of flexible, automatic and open-source data processing tools called the Citizen Science Solution Kit. The kit provides a comprehensive set of tools that can be used and personalized in different situations, particularly during natural emergencies, starting from images and text contained in the posts. The tools can be employed by citizen scientists for filtering, classifying, and geolocating the content with a human-in-the-loop approach to support the data analyst, including feedback and suggestions on how to configure the automated tools, and techniques to gather inputs from citizens. Using flooding scenario as a guiding example, this paper illustrates the structure and functioning of the different tools proposed to support citizens scientists in their projects, and a methodological approach to their use. The process is then validated by discussing three case studies based on the Albania earthquake of 2019, the Covid-19 pandemic, and the Thailand floods of 2021. The results suggest that a flexible approach to tools composition and configuration can support a timely setup of an analysis project by citizen scientists, especially in case of emergencies in unexpected locations.

Background Finding the correct drug dose for a specific condition is a key step in many treatments, and failing to do so can lead to deleterious consequences to patient health. Clinical protocols are derived from drug development phase prospective trials. While carefully designed, these often do not include all potential patients, comorbidities or clinical outcomes, ultimately leading to sub-optimal dosing policies. Observational datasets provide real-world information that cannot be substituted with data collected in a controlled environment. Several published methodologies have applied observational datasets for the development of clinical protocols, however these are only applicable whenever these datasets are varied and complete. Often, clinical observational datasets do not comply with these requirements. Computational methods can and should exploit field knowledge to address weaknesses associated with clinical observational data. Methods This paper proposes IDoser, a core dosing model that links drug dose to relevant covariates via a set of coefficients, and includes a loss function to codify needed assumptions and requirements. Coordinate descent is used to obtain a fitted model with minimal loss. The loss function is also used to measure performance when validating the model with unseen data. Our proposal is validated using the case of follicle stimulating hormone (FSH) dosing for controlled ovarian stimulation (COS). Results The proposed Individualized Doser (IDoser) achieved significant improvements when loss values were compared to observed clinical practice and a selected literature benchmark and during the validation phase. Conclusions This methodology constitutes a simple but effective method to bridge the gap between current clinical dosing policies and gold policies based on the true underlying and often unknown dose-response functions.

Background Amid the COVID-19 pandemic, there has been a need for rapid social understanding to inform infodemic management and response. Although social media analysis platforms have traditionally been designed for commercial brands for marketing and sales purposes, they have been underused and adapted for a comprehensive understanding of social dynamics in areas such as public health. Traditional systems have challenges for public health use, and new tools and innovative methods are required. The World Health Organization Early Artificial Intelligence–Supported Response with Social Listening (EARS) platform was developed to overcome some of these challenges. Objective This paper describes the development of the EARS platform, including data sourcing, development, and validation of a machine learning categorization approach, as well as the results from the pilot study. Methods Data for EARS are collected daily from web-based conversations in publicly available sources in 9 languages. Public health and social media experts developed a taxonomy to categorize COVID-19 narratives into 5 relevant main categories and 41 subcategories. We developed a semisupervised machine learning algorithm to categorize social media posts into categories and various filters. To validate the results obtained by the machine learning–based approach, we compared it to a search-filter approach, applying Boolean queries with the same amount of information and measured the recall and precision. Hotelling T2 was used to determine the effect of the classification method on the combined variables. Results The EARS platform was developed, validated, and applied to characterize conversations regarding COVID-19 since December 2020. A total of 215,469,045 social posts were collected for processing from December 2020 to February 2022. The machine learning algorithm outperformed the Boolean search filters method for precision and recall in both English and Spanish languages (P<.001). Demographic and other filters provided useful insights on data, and the gender split of users in the platform was largely consistent with population-level data on social media use. Conclusions The EARS platform was developed to address the changing needs of public health analysts during the COVID-19 pandemic. The application of public health taxonomy and artificial intelligence technology to a user-friendly social listening platform, accessible directly by analysts, is a significant step in better enabling understanding of global narratives. The platform was designed for scalability; iterations and new countries and languages have been added. This research has shown that a machine learning approach is more accurate than using only keywords and has the benefit of categorizing and understanding large amounts of digital social data during an infodemic. Further technical developments are needed and planned for continuous improvements, to meet the challenges in the generation of infodemic insights from social media for infodemic managers and public health professionals.

During disasters, social media can serve as a valuable source of real-time information about the impacts on people and infrastructure. However, due to the lack of geographical information in most social media posts, this information is often underutilized by first responders. Previous research has attempted to estimate the location of individual social media posts using text and image analysis, but limitations still exist in fine-grained disaster area mapping. To address this issue, this paper explores the feasibility of automatically extracting textual information from social media images to enhance the creation of a disaster area map. This paper evaluates the effectiveness of the proposed approach and its potential impact on improving the geolocation of social media information during a disaster.

The surge of supervised learning methods for segmentation lately has underscored the critical role of label quality in predicting performance. This issue is prevalent in the domain of medical imaging, where high annotation costs and inter-observer variability pose significant challenges. Acquiring labels commonly involves multiple experts providing their interpretations of the “true” segmentation labels, each influenced by their individual biases. The blind acceptance of these noisy labels as the ground truth restricts the potential effectiveness of segmentation algorithms. Here, we apply coupled convolutional neural network approaches to a small-sized real-world dataset of bovine cumulus oocyte complexes. This is the first time these methods have been applied to a real-world annotation medical dataset, since they were previously tested only on artificially generated labels of medical and non-medical datasets. This dataset is crucial for healthy embryo development. Its application revealed an important challenge: the inability to effectively learn distinct confusion matrices for each expert due to large areas of agreement. In response, we propose a novel method that focuses on areas of high uncertainty. This approach allows us to understand the individual characteristics better, extract their behavior, and use this insight to create a more sophisticated ground truth using maximum likelihood. These findings contribute to the ongoing discussion of leveraging machine learning algorithms for medical image segmentation, particularly in scenarios involving multiple human annotators.