Introduction: Early recognition of diseases in pets is essential, yet owners often face challenges in interpreting clinical symptoms. Digital symptom checkers offer a promising approach to encode veterinary knowledge, but their reliability and diagnostic accuracy remain largely unvalidated. This study addresses this gap through a method validation of a expert-knowledge-based veterinary symptom checker using synthetically generated test cases, enabling systematic exploration of the symptom-disease space in the absence of clinical data.

Methods: System performance was quantified using simulated user-checker dialogs across ≈550 diseases for dogs and cats, respectively. Robustness and efficiency were evaluated through three research questions: convergence probability, convergence speed, and structural factors influencing convergence.

Results: The system achieved full convergence under ideal conditions (100%), with rapid convergence (mean rank of one after ≈20 questions) and short response times (0.213-0.258 msec per disease). Under probabilistic user-answering strategies, performance decreased slightly but remained robust, with non-converging cases rare and correct diagnoses typically among top-ranked results (ranks 1-6 for dogs; 1-4 for cats). Structural analysis identified the number and uniqueness of symptoms as key predictors of diagnostic difficulty, with significant variation across anatomical regions.

Discussion: Findings confirm the system's internal consistency, robustness, and computational efficiency, establishing a validated foundation for evidence-based veterinary diagnostic support. Future work will include clinical and user studies to confirm performance under authentic conditions and address current limitations of synthetic data.

This publication presents an established engineering-driven framework for pipelining large language model (LLM) agents for automatic code translation. The framework is applied in a long-term ongoing industrial project with Danfoss Power Solutions aimed at translating five million lines of Delphi to C#. To manage this scale, the source codebase is divided into translatable chunks, which are processed independently through an agentic LLM pipeline, with subsequently the reassembly of the translated code chunks in the target language. Due to the inherent stochasticity of LLMs, the framework incorporates processing and validation functions to ensure consistency and correctness. As an extended version of our short paper (Bruneel et al., 2025), this paper provides a significantly more comprehensive architectural blueprint. Specifically, we present an indepth background & related work section, formalise the chunking and reassembly strategies, detail the internal mechanics of agent stages, and introduce sequential agent orchestration alongside the dynamic LLM pipeline architecture. Furthermore, we report early empirical observations regarding the pipeline's computational overhead, validation dynamics, and translation quality, noting a practical impact of an initial 10x acceleration over estimated manual translation efforts. Ultimately, we share these expanded insights to inform future development and application in similar large-scale translation tasks.

Accurate parcel geometry is important for many applications in agriculture and forestry. The assessments of agricultural parcel geometries are, e.g., crucial for calculating financial subsidies. Today, authorities use aerial images to check parcels for homogeneous cultivation and incorrect parcel geometries (homogeneity-check). Due to the large number of parcels and limited resources of authority offices, there is a need for an automated method of risk-based selection of a manageable subset of parcels for manual review. To this end, this study evaluated four machine learning approaches; Image Classification, Semantic Segmentation, Edge Detection, and Clustering, with respect to (i) their sensitivity to spatial resolution of aerial images, (ii) the impact of the balancing method used in training, and (iii) their overall performance with respect to preselecting parcels for homogeneity-check. In addition, we introduced a modified version of the commonly used Precision@K metric. The proposed Precision@K% replaces the absolute number of top-ranked items with a relative proportion, facilitating a more scalable evaluation of risk-based selection strategies. The results show that all approaches performed similarly in terms of the area under the receiver operating characteristic curve (AUROC) and the Precision@2% metric; Image Classification: AUROC = 77.4%, Precision@2% = 80%, Semantic Segmentation: AUROC = 83.0%, Precision@2% = 75%, Edge Detection: AUROC = 79.2%, Precision@2% = 85%, and Clustering: AUROC = 84.4%, Precision@2% = 100%. Notably, the Clustering approach consistently maintained a stable Precision@2% of 100%.

Articulated haulers are heavy construction machines used for earthmoving tasks. To optimize factors such as weight, cost, CO₂ footprint, and productivity, maintaining the optimal chassis weight is critical. To support this, a digital twin approach is developed. In this study, two digital twin concepts for articulated haulers are evaluated and compared based on their prediction accuracy for component life. Both concepts leverage vehicle simulations, but one introduces segmented simulation data to increase flexibility and improve data utilization. Cross-validation is employed to assess the core idea of the digital twins. Results show that the best-performing concept achieves a mean cross-validation accuracy of 99%, demonstrating strong potential for future design optimization.

The purpose of this study was to compare manual assessment of corneal nerve fiber length (CNFL) and dendritic cell (DC) density with an automated assessment method utilizing deep learning segmentation to perform rule-based density estimation. Corneal images were acquired using in vivo confocal microscopy (IVCM) from 100 participants with persistent ocular symptoms after mild COVID-19 (Group 1) and 30 controls without symptoms (Group 2). In total, 1,300 IVCM images were selected and manually annotated for CNFL, and 1,300 for DCs (with dendrites and without dendrites), using FIJI tools. The between-method difference in mean CNFL density was 0.2  (95% CI: [0.09, 0.23]) for Group 1 and −0.2  (95% CI: [−0.34, −0.10]) for Group 2. For Group 1, the mean difference for DCs with dendrites was −1.1  (95% CI: [−1.78, −0.39]), and for DCs without dendrites it was −3.1  (95% CI: [−5.1, −1.0]). For Group 2, the mean difference for DCs with dendrites was −1.0  (95% CI: [−1.79, −0.27]), and for DCs without dendrites it was 0.3  (95% CI: [−1.93, 2.60]). Both manual and automated methods showed significant between-group differences for CNFL (p=0.012 and p=0.034, respectively) and DC densities (p=0.005 and p=0.010). The automated approach performed comparably to manual assessment, supporting its potential for reliable, scalable analysis of CNFL and DC in IVCM images.

Maintaining mobility is vital for older adults. However, standardized functional tests often overlook crucial qualitative aspects, and expert assessments (EA) are costly and lack standardization. This project aims to develop an AI-based movement screening tool (SAT-Movement Analysis) utilizing the low-cost Skeleton Avatar Technique (SAT) and standardized Observational Movement Analysis (OMA) to detect deviations in daily movement. The initial phase automated expert assessments to establish a reliable foundation for machine learning. Five participants (ages 35–57) performed Sit-To-Stand, Stand-To-Sit, and One-Leg Stance, assessed by three physiotherapists using a modified IRAF protocol. Results demonstrated correspondence between automatically aggregated expert scores and consensus scores across all aggregation levels (Pearson’s r = 0.90–0.97, ICC = 0.91–0.98, = 0.78–1.00). These findings motivate continued development of an AI-trained screening tool providing accurate movement quality feedback based on 2D smartphone video, supporting early detection and personalized intervention.

Many real-world modeling problems are hindered by limited data availability. In such cases, transfer learning leverages related source domains to improve predictions in a target domain of interest. We extend the classical gradient tree boosting paradigm to a regression transfer algorithm by modeling the weak learner as a sum of two regression trees. The trees are fitted on source data and target data, respectively, and jointly optimized for the target data. We derive optimal coefficients for the model update under the least-squares, the least-absolute-deviation, and the Huber loss functions. We benchmark our approach against boosting-based regression transfer methods in twelve transfer scenarios. The results indicate that our approach constitutes a competitive alternative within the realm of boosting-based regression transfer. Moreover, we provide a theoretical justification as well as empirical validation that our approach is robust under larger domain shifts.

This research addresses the computational inefficiency in Volvo Construction Equipment's (VCE) co-simulation framework, which integrates the discrete element method (DEM) modeling of bucket-gravel interactions, which runs similar to 100 times slower than real time. We develop a deep learning surrogate to replace the DEM subsystem to predict the forces during wheel loader (WL) bucket-particle interactions. An optimized deep learning (DL) model trained with transfer learning achieves an average element-wise match accuracy [1] of 0.9903 to 0.9910 with mean squared error (MSE) of 3.3869 & times; 10-4 to 4.3427 & times; 10-4 (normalized units), while preserving the relevant sampling frequency content up to 100 Hz. Real-time capability is demonstrated by an inference time of about 6 ms per 100 Hz step, enabling a 32 s short cycle (3200 steps) to be predicted in about 19 s. Model fidelity is verified against DEM using time-series agreement and frequency, including power spectral density (PSD) comparisons, range-pair evaluation, and duty-ratio assessment (predicted duty from DL model divided by DEM duty), with duty ratios generally within the acceptable band of 0.5 to 2.0. The main limitation is that the surrogate is trained and validated on a specific gravel size range (200 to 300 mm, uniform distribution) and a single bucket configuration (5 m3). The DL model is designed for integration into the co-simulation framework, enabling real-time simulation at the DEM subsystem level and accelerating product development. Furthermore, the DL model supports force generation in site simulations and hardware-in-loop (HIL) testing.

Remote sensing techniques are widely used for mapping and monitoring forest attributes, providing valuable information on forest cover, biomass, and overall forest health. In recent years, national airborne laser scanning (ALS) campaigns have been conducted in several countries to map forest resources. When combining ALS with field inventory data, these datasets enable the development of nationwide models for prediction of forest attributes. In this study, we explore the potential of machine learning (ML) to enhance existing modeling approaches for nationwide forest attribute mapping in Sweden. We achieve this by relating ALS data from the most recent ALS campaign of Sweden with field data from the Swedish National Forest Inventory (NFI). By aggregating laser metrics from surveyed areas (NFI plots), as well as over surrounding areas to the plots, we investigate (1) if ML approaches can outperform existing linear regression baseline models and (2) if further enhancements of the predictive capacity can be achieved by including surrounding, spatially correlated ALS data. To this end, we used extreme gradient boosting (XGBoost), as well as a convolutional neural network (CNN), specialized to handle tabular data and spatially correlated data, respectively. The models were evaluated on five forest variables: basal-area weighted mean tree height, basal-area weighted mean stem diameter, basal area, stem volume, and above-ground biomass. All models were evaluated on several nested datasets to assess the robustness, showcasing consistent results across datasets. We achieved significant improvements in prediction accuracy across all investigated forest variables. Furthermore, incorporating surrounding information to the modeling rendered further improvements for diameter, basal area, and biomass predictions. The approaches tested and developed here thus form a promising basis for flexible modeling approaches that can be transferred globally for large-scale forest monitoring and management.

Acanthamoeba are ubiquitous microorganisms found in the air, soil, tap and drinking water, swimming pools, and in both saltwater and freshwater environments [1].They can cause corneal infections that lead to Acanthamoeba keratitis (AK), which is serious medical condition accompanied by intense eye pain, sensitivity to light, and severely impaired vision. AK is commonly associated with contact lens use, particularly in developed countries, where up to 90% of cases are linked to contact lens wear [2]. Timely identification and diagnosis of AK are critical to preventing vision loss. Corneal culture is the gold standard for diagnosing AK, but its low sensitivity and slow turnaround can delay treatment and worsen infection [3]. A complementary method uses non-invasive clinical imaging such as in-vivo confocal microscopy (IVCM) [4,5]. Despite several advantages such as high-resolution imaging, rapid diagnosis, and the ability to monitor disease progression, IVCM has limitations, including the large volume of images generated and the needs for identifying informative/useful images and experienced interpretation. While previous approaches have primarily focused on building models to detect AK in pre- selected images, our decision support system [6], IVCMAssist, automates the entire diagnostic workflow—from processing raw IVCM images and removing artifacts, to sorting images by corneal layer, to identifying informative images, and to detecting AK signs such as cysts in informative images. It enhances the AK diagnosis process by reducing the time spent on image sorting and pre-selection of informative images (non-overlapped images with detected AK cysts), while enabling experienced observers to correct model errors through a human-in-the-loop AI approach.