Panagiotis Papapetrou

Lena Mondrejevski, Ioanna Miliou, Annaclaudia Montanino et al.

Although Machine Learning (ML) can be seen as a promising tool to improve clinical decision-making for supporting the improvement of medication plans, clinical procedures, diagnoses, or medication prescriptions, it remains limited by access to healthcare data. Healthcare data is sensitive, requiring strict privacy practices, and typically stored in data silos, making traditional machine learning challenging. Federated learning can counteract those limitations by training machine learning models over data silos while keeping the sensitive data localized. This study proposes a federated learning workflow for ICU mortality prediction. Hereby, the applicability of federated learning as an alternative to centralized machine learning and local machine learning is investigated by introducing federated learning to the binary classification problem of predicting ICU mortality. We extract multivariate time series data from the MIMIC-III database (lab values and vital signs), and benchmark the predictive performance of four deep sequential classifiers (FRNN, LSTM, GRU, and 1DCNN) varying the patient history window lengths (8h, 16h, 24h, 48h) and the number of FL clients (2, 4, 8). The experiments demonstrate that both centralized machine learning and federated learning are comparable in terms of AUPRC and F1-score. Furthermore, the federated approach shows superior performance over local machine learning. Thus, the federated approach can be seen as a valid and privacy-preserving alternative to centralized machine learning for classifying ICU mortality when sharing sensitive patient data between hospitals is not possible.

Zhendong Wang, Ioanna Miliou, Isak Samsten et al.

Among recent developments in time series forecasting methods, deep forecasting models have gained popularity as they can utilize hidden feature patterns in time series to improve forecasting performance. Nevertheless, the majority of current deep forecasting models are opaque, hence making it challenging to interpret the results. While counterfactual explanations have been extensively employed as a post-hoc approach for explaining classification models, their application to forecasting models still remains underexplored. In this paper, we formulate the novel problem of counterfactual generation for time series forecasting, and propose an algorithm, called ForecastCF, that solves the problem by applying gradient-based perturbations to the original time series. ForecastCF guides the perturbations by applying constraints to the forecasted values to obtain desired prediction outcomes. We experimentally evaluate ForecastCF using four state-of-the-art deep model architectures and compare to two baselines. Our results show that ForecastCF outperforms the baseline in terms of counterfactual validity and data manifold closeness. Overall, our findings suggest that ForecastCF can generate meaningful and relevant counterfactual explanations for various forecasting tasks.

Pranav Raikote, Korbinian Randl, Ioanna Miliou et al.

Scaling educational assessment with large language models requires not just accuracy, but the ability to recognize when predictions are trustworthy. Instruction-tuned models tend to be overconfident, and their reliability deteriorates as curricula evolve, making fully autonomous deployment unsafe in high-stakes settings. We introduce CHiL(L)Grader, the first automated grading framework that incorporates calibrated confidence estimation into a human-in-the-loop workflow. Using post-hoc temperature scaling, confidence-based selective prediction, and continual learning, CHiL(L)Grader automates only high-confidence predictions while routing uncertain cases to human graders, and adapts to evolving rubrics and unseen questions. Across three short-answer grading datasets, CHiL(L)Grader automatically scores 35-65% of responses at expert-level quality (QWK >= 0.80). A QWK gap of 0.347 between accepted and rejected predictions confirms the effectiveness of the confidence-based routing. Each correction cycle strengthens the model's grading capability as it learns from teacher feedback. These results show that uncertainty quantification is key for reliable AI-assisted grading.

Maria Movin, Claudia Hauff, Aron Henriksson et al.

LLM-driven GUI agents are increasingly used in production systems to automate workflows and simulate users for evaluation and optimization. Yet most GUI-agent evaluations emphasize task success and provide limited evidence on whether agents interact in human-like ways. We present a trace-level evaluation framework that compares human and agent behavior across (i) task outcome and effort, (ii) query formulation, and (iii) navigation across interface states. We instantiate the framework in a controlled study in a production audio-streaming search application, where 39 participants and a state-of-the-art GUI agent perform ten multi-hop search tasks. The agent achieves task success comparable to participants and generates broadly aligned queries, but follows systematically different navigation strategies: participants exhibit content-centric, exploratory behavior, while the agent is more search-centric and low-branching. These results show that outcome and query alignment do not imply behavioral alignment, motivating trace-level diagnostics when deploying GUI agents as proxies for users in production search systems.

Franco Rugolon, Korbinian Randl, Braslav Jovanovic et al.

Multimodal Machine Learning offers a holistic view of a patient's status, integrating structured and unstructured data from electronic health records (EHR). We propose a framework to predict metastasis risk one month prior to diagnosis, using six months of clinical history from EHR data. Data from four cancer cohorts collected at Karolinska University Hospital (Stockholm, Sweden) were analyzed: breast (n = 743), colon (n = 387), lung (n = 870), and prostate (n = 1890). The dataset included demographics, comorbidities, laboratory results, medications, and clinical text. We compared traditional and deep learning classifiers across single modalities and multimodal combinations, using various fusion strategies and a Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) 2a design, with an 80-20 development-validation split to ensure a rigorous, repeatable evaluation. Performance was evaluated using AUROC, AUPRC, F1 score, sensitivity, and specificity. We then employed a multimodal adaptation of SHAP to analyze the classifiers' reasoning. Intermediate fusion achieved the highest F1 scores on breast (0.845), colon (0.786), and prostate cancer (0.845), demonstrating strong predictive performance. For lung cancer, the intermediate fusion achieved an F1 score of 0.819, while the text-only model achieved the highest, with an F1 score of 0.829. Deep learning classifiers consistently outperformed traditional models. Colon cancer, the smallest cohort, had the lowest performance, highlighting the importance of sufficient training data. SHAP analysis showed that the relative importance of modalities varied across cancer types. Fusion strategies offer distinct strengths and weaknesses. Intermediate fusion consistently delivered the best results, but strategy choices should align with data characteristics and organizational needs.

Tim Kreuzer, Jelena Zdravkovic, Panagiotis Papapetrou

Time series forecasting has seen considerable improvement during the last years, with transformer models and large language models driving advancements of the state of the art. Modern forecasting models are generally opaque and do not provide explanations for their forecasts, while well-known post-hoc explainability methods like LIME are not suitable for the forecasting context. We propose PAX-TS, a model-agnostic post-hoc algorithm to explain time series forecasting models and their forecasts. Our method is based on localized input perturbations and results in multi-granular explanations. Further, it is able to characterize cross-channel correlations for multivariate time series forecasts. We clearly outline the algorithmic procedure behind PAX-TS, demonstrate it on a benchmark with 7 algorithms and 10 diverse datasets, compare it with two other state-of-the-art explanation algorithms, and present the different explanation types of the method. We found that the explanations of high-performing and low-performing algorithms differ on the same datasets, highlighting that the explanations of PAX-TS effectively capture a model's behavior. Based on time step correlation matrices resulting from the benchmark, we identify 6 classes of patterns that repeatedly occur across different datasets and algorithms. We found that the patterns are indicators of performance, with noticeable differences in forecasting error between the classes. Lastly, we outline a multivariate example where PAX-TS demonstrates how the forecasting model takes cross-channel correlations into account. With PAX-TS, time series forecasting models' mechanisms can be illustrated in different levels of detail, and its explanations can be used to answer practical questions on forecasts.

Rami Mochaourab, Sugandh Sinha, Stanley Greenstein et al.

We consider counterfactual explanations for private support vector machines (SVM), where the privacy mechanism that publicly releases the classifier guarantees differential privacy. While privacy preservation is essential when dealing with sensitive data, there is a consequent degradation in the classification accuracy due to the introduced perturbations in the classifier weights. For such classifiers, counterfactual explanations need to be robust against the uncertainties in the SVM weights in order to ensure, with high confidence, that the classification of the data instance to be explained is different than its explanation. We model the uncertainties in the SVM weights through a random vector, and formulate the explanation problem as an optimization problem with probabilistic constraint. Subsequently, we characterize the problem's deterministic equivalent and study its solution. For linear SVMs, the problem is a convex second-order cone program. For non-linear SVMs, the problem is non-convex. Thus, we propose a sub-optimal solution that is based on the bisection method. The results show that, contrary to non-robust explanations, the quality of explanations from the robust solution degrades with increasing privacy in order to guarantee a prespecified confidence level for correct classifications.

John Pavlopoulos, Panagiotis Papapetrou

A language model can be used to predict the next word during authoring, to correct spelling or to accelerate writing (e.g., in sms or emails). Language models, however, have only been applied in a very small scale to assist physicians during authoring (e.g., discharge summaries or radiology reports). But along with the assistance to the physician, computer-based systems which expedite the patient's exit also assist in decreasing the hospital infections. We employed statistical and neural language modeling to predict the next word of a clinical text and assess all the models in terms of accuracy and keystroke discount in two datasets with radiology reports. We show that a neural language model can achieve as high as 51.3% accuracy in radiology reports (one out of two words predicted correctly). We also show that even when the models are employed only for frequent words, the physician can save valuable time.

Vasiliki Kougia, John Pavlopoulos, Panagiotis Papapetrou et al.

This paper introduces RTEx, a novel methodology for a) ranking radiography exams based on their probability to contain an abnormality, b) generating abnormality tags for abnormal exams, and c) providing a diagnostic explanation in natural language for each abnormal exam. The task of ranking radiography exams is an important first step for practitioners who want to identify and prioritize those radiography exams that are more likely to contain abnormalities, for example, to avoid mistakes due to tiredness or to manage heavy workload (e.g., during a pandemic). We used two publicly available datasets to assess our methodology and demonstrate that for the task of ranking it outperforms its competitors in terms of NDCG@k. For each abnormal radiography exam RTEx generates a set of abnormality tags alongside an explanatory diagnostic text to explain the tags and guide the medical expert. Our tagging component outperforms two strong competitor methods in terms of F1. Moreover, the diagnostic captioning component of RTEx, which exploits the already extracted tags to constrain the captioning process, outperforms all competitors with respect to clinical precision and recall.

Maria Bampa, Panagiotis Papapetrou

We study the problem of detecting adverse drug events in electronic healthcare records. The challenge in this work is to aggregate heterogeneous data types involving diagnosis codes, drug codes, as well as lab measurements. An earlier framework proposed for the same problem demonstrated promising predictive performance for the random forest classifier by using only lab measurements as data features. We extend this framework, by additionally including diagnosis and drug prescription codes, concurrently. In addition, we employ a recursive feature selection mechanism on top, that extracts the top-k most important features. Our experimental evaluation on five medical datasets of adverse drug events and six different classifiers, suggests that the integration of these additional features provides substantial and statistically significant improvements in terms of AUC, while employing medically relevant features.

Amin Azari, Panagiotis Papapetrou, Stojan Denic et al.

Prediction of user traffic in cellular networks has attracted profound attention for improving resource utilization. In this paper, we study the problem of network traffic traffic prediction and classification by employing standard machine learning and statistical learning time series prediction methods, including long short-term memory (LSTM) and autoregressive integrated moving average (ARIMA), respectively. We present an extensive experimental evaluation of the designed tools over a real network traffic dataset. Within this analysis, we explore the impact of different parameters to the effectiveness of the predictions. We further extend our analysis to the problem of network traffic classification and prediction of traffic bursts. The results, on the one hand, demonstrate superior performance of LSTM over ARIMA in general, especially when the length of the training time series is high enough, and it is augmented by a wisely-selected set of features. On the other hand, the results shed light on the circumstances in which, ARIMA performs close to the optimal with lower complexity.

Amin Azari, Panagiotis Papapetrou, Stojan Denic et al.

Traffic prediction plays a vital role in efficient planning and usage of network resources in wireless networks. While traffic prediction in wired networks is an established field, there is a lack of research on the analysis of traffic in cellular networks, especially in a content-blind manner at the user level. Here, we shed light into this problem by designing traffic prediction tools that employ either statistical, rule-based, or deep machine learning methods. First, we present an extensive experimental evaluation of the designed tools over a real traffic dataset. Within this analysis, the impact of different parameters, such as length of prediction, feature set used in analyses, and granularity of data, on accuracy of prediction are investigated. Second, regarding the coupling observed between behavior of traffic and its generating application, we extend our analysis to the blind classification of applications generating the traffic based on the statistics of traffic arrival/departure. The results demonstrate presence of a threshold number of previous observations, beyond which, deep machine learning can outperform linear statistical learning, and before which, statistical learning outperforms deep learning approaches. Further analysis of this threshold value represents a strong coupling between this threshold, the length of future prediction, and the feature set in use. Finally, through a case study, we present how the experienced delay could be decreased by traffic arrival prediction.

Isak Karlsson, Jonathan Rebane, Panagiotis Papapetrou et al.

Time series classification has received great attention over the past decade with a wide range of methods focusing on predictive performance by exploiting various types of temporal features. Nonetheless, little emphasis has been placed on interpretability and explainability. In this paper, we formulate the novel problem of explainable time series tweaking, where, given a time series and an opaque classifier that provides a particular classification decision for the time series, we want to find the minimum number of changes to be performed to the given time series so that the classifier changes its decision to another class. We show that the problem is NP-hard, and focus on two instantiations of the problem, which we refer to as reversible and irreversible time series tweaking. The classifier under investigation is the random shapelet forest classifier. Moreover, we propose two algorithmic solutions for the two problems along with simple optimizations, as well as a baseline solution using the nearest neighbor classifier. An extensive experimental evaluation on a variety of real datasets demonstrates the usefulness and effectiveness of our problem formulation and solutions.

Irvin Homem, Panagiotis Papapetrou, Spyridon Dosis

DNS tunneling techniques are often used for malicious purposes but network security mechanisms have struggled to detect these. Network forensic analysis has thus been used but has proved slow and effort intensive as Network Forensics Analysis Tools struggle to deal with undocumented or new network tunneling techniques. In this paper we present a method to aid forensic analysis through automating the inference of protocols tunneled within DNS tunneling techniques. We analyze the internal packet structure of DNS tunneling techniques and characterize the information entropy of different network protocols and their DNS tunneled equivalents. From this, we present our protocol prediction method that uses entropy distribution averaging. Finally we apply our method on a dataset to measure its performance and show that it has a prediction accuracy of 75%. Our method also preserves privacy as it does not parse the actual tunneled content, rather it only calculates the information entropy.

Andreas Henelius, Kai Puolamäki, Henrik Boström et al.

Clustering is a widely used unsupervised learning method for finding structure in the data. However, the resulting clusters are typically presented without any guarantees on their robustness; slightly changing the used data sample or re-running a clustering algorithm involving some stochastic component may lead to completely different clusters. There is, hence, a need for techniques that can quantify the instability of the generated clusters. In this study, we propose a technique for quantifying the instability of a clustering solution and for finding robust clusters, termed core clusters, which correspond to clusters where the co-occurrence probability of each data item within a cluster is at least $1 - α$. We demonstrate how solving the core clustering problem is linked to finding the largest maximal cliques in a graph. We show that the method can be used with both clustering and classification algorithms. The proposed method is tested on both simulated and real datasets. The results show that the obtained clusters indeed meet the guarantees on robustness.