Open Access

Process mining often yields highly complex “Spaghetti Models”, making them difficult to interpret and impeding informed decision-making. Therefore, researchers have explored clustering of event logs to simplify process models and reduce their complexity. However, the unsupervised nature of clustering can introduce an interpretation gap, necessitating manual effort to identify differences and similarities across the resulting process models. To address these issues, we propose an explainable clustering approach that identifies key subprocesses and applies eXplainable Artificial Intelligence (XAI) techniques to clarify the rationale behind model partitioning. Moreover, we integrate a Large Language Model (LLM) into the process discovery procedure to generate natural language descriptions and compare the discovered process models, enhancing user understanding, engagement, and making complex technical details more accessible. A case study demonstrates that our method operates effectively across various LLMs, preserving vital contextual information while simplifying the process discovery workflow. Our findings reveal that larger models generally ensure completeness, whereas smaller ones offer more efficiency at the expense of explanation quality, highlighting the importance of a balanced LLM choice for practical applications.

Smart sensor systems are a key factor to ensure sustainable compute by enabling machine learning algorithms to be executed at the data source. This is particularly helpful when working with moving parts or in remote areas, where no tethered deployment is possible. However, including computations directly at the measurement device places an increased load on the power budget. Therefore, we introduce the Hierarchical Machine Learning framework “HiMLEdge” which enables highly specialized models that are tuned using an energy-aware multi-criteria optimization. We evaluate our framework with prognostic health management in a three-part feasibility study: First, we apply an exhaustive search to find hierarchical taxonomies, which we benchmark against hand-tuned flat classifiers. This test shows a decrease in power consumption of up to 47.63% for the hierarchical approach. Second, the search strategy is improved with Reinforcement Learning. As a novel contribution, we include real measurements in the reward function, instead of using a surrogate metric. This inclusion leads to a different optimal policy in comparison to the literature, which shows the error that may be introduced by an approximation. Third, we conduct tests on the system level, including communication and system-off power draw. In this scenario, the optimized hierarchical model can perform four times as many readings per hour as a flat classifier while achieving the same five years of battery life with similar accuracy. In turn, this also means that the battery life can be increased by the same amount if the readings per hour are kept constant.

The rise of machine-learning applications in domains with critical end-user impact has led to a growing concern about the fairness of learned models, with the goal of avoiding biases that negatively impact specific demographic groups. Most existing bias-mitigation strategies adapt the importance of data instances during pre-processing. Since fairness is a contextual concept, we advocate for an interactive machine-learning approach that enables users to provide iterative feedback for model adaptation. Specifically, we propose to adapt the explanatory interactive machine-learning approach Caipi for fair machine learning. FairCaipi incorporates human feedback in the loop on predictions and explanations to improve the fairness of the model. Experimental results demonstrate that FairCaipi outperforms a state-of-the-art pre-processing bias mitigation strategy in terms of the fairness and the predictive performance of the resulting machine-learning model. We show that FairCaipi can both uncover and reduce bias in machine-learning models and allows us to detect human bias.

BACKGROUND Ischemic stroke (IS) and retinal ischemia (IR) share similar vascular risk factors, but differ in their risk for subsequent or recurrent stroke and therapeutic options. This study characterizes the cardiovascular risk profiles and magnitude of atherosclerosis of the carotid artery of patients with central retinal artery occlusion (CRAO) in relation to the presence of the retrobulbar "spot sign" on orbital color-coded sonography (OCCS). METHODS We performed a retrospective analysis on the detailed cardiovascular risk factors and neuroimaging data in patients with IR presenting between 2009 and 2023. Based on OCCS findings, CRAO were further divided into hyperechoic ("spot sign positive", ssCRAO) or hypoechoic CRAO (heCRAO). Statistical analyses were performed with Mann-Whitney-U and χ [2] testing. P-values were considered significant if < 0.05. RESULTS Overall, 112 patients were identified (heCRAO: n = 32; ssCRAO: n = 80). ssCRAO patients were significantly older (median 74 years vs. 66.5 years, Mann-Whitney-U: p-value < 0.001). Overall, 15/103 (14.6%) patients had concurrent acute ischemic stroke- 9 in the ipsilateral internal carotid territory, 2 in other territories and 4 disseminated. Further significant differences were found regarding the echogenicity of atherosclerosis (AS) in the two subgroups with (mainly) echorich AS being more common in the ssCRAO group (p-value < 0.001, n = 108) and the distribution of high-grade vs. low-grade stenoses of the ipsi- and contralateral carotid artery (p-value < 0.05, n = 99). 20 out of 112 patients had atrial fibrillation (aFib) with 17 of these being on ongoing oral anticoagulation. CONCLUSION According to this study, atherosclerosis may be one of the most important risk factors for IR while a specific embolic source could not be demonstrated (i.e. acute plaque rupture). By contrast, current oral anticoagulation for aFib in CRAO patients was high, thus only an incidental finding and may be an incidental finding due to its prevalence in the elderly. Furthermore, we were able to distinguish two subgroups of IR that differ in risk factors and most likely also in etiology, therapy and prognosis. The study underlines the importance of OCCS to detect "spot signs" in IR with indications for both, acute thrombolysis and secondary prevention.

Aims Systematic visualization of predefined anatomical sites is essential for quality assurance and training in esophagogastroduodenoscopy (EGD), but objective, automated assessment of examination completeness in Europe is lacking. We aimed to develop and preliminarily evaluate a deep learning algorithm that classifies still images from the upper gastrointestinal tract (UGIT) into 28 standardized anatomical categories, including 22 gastric segments according to the Systemic Stomach Screening (SSS) protocol by Kenshi Yao. Primary outcomes were macro F1-score and overall accuracy for this 28-class classification task. Methods In this retrospective diagnostic machine learning study, we collected endoscopic still images from two sources: (1) the publicly available GastroHUN dataset (Colombia; Bravo et al., Sci Data 2025) and (2) an internal dataset from a European tertiary referral centre (Augsburg, Germany). All images were obtained during routine EGD using Olympus endoscopes and anonymized prior to analysis. A total of 1,553 patients were included: 385 from GastroHUN (5,202 images) and 1,114 from UKA (2,793 images), yielding 8,095 images. Each image was labelled at image level into one of 28 anatomical classes: oropharynx, esophagus, esophagogastric junction, duodenal bulb, second part of the duodenum, papilla and 22 gastric segments based on the SSS classification using a multi-annotator process. We trained a ConvNeXt-Tiny convolutional neural network for image-level classification using patient-wise 5-fold cross-validation to prevent information leakage between training and test sets. Performance was evaluated as mean macro F1-score and overall accuracy across folds. Results Across 5-fold patient-wise cross-validation, the model achieved a mean macro F1-score of 0.79±0.01 and an overall accuracy of 0.79±0.01 for 28-class anatomical site classification. The network differentiated major UGIT regions (esophagus, stomach, duodenum) and also assigned images to fine-grained gastric segments according to the SSS protocol. Qualitatively, performance was higher in broader anatomical regions and lower in some of the most detailed gastric segments (e.g., lower vs. upper corpus posterior wall), consistent with the increased difficulty of very fine-grained localization. Conclusions This preliminary study demonstrates the feasibility of automated recognition of UGIT anatomical sites using a ConvNeXt-Tiny CNN trained on a combination of public and tertiary hospital data. The model achieved moderate performance across 28 predefined anatomical categories from the oropharynx to the second part of the duodenum, including 22 SSS gastric segments, indicating that fine-grained anatomical localization of the entire upper GI tract is achievable from still images alone. To our knowledge, this is among the first studies to leverage a public endoscopy dataset together with institutional data for upper GI anatomical site classification.