The classification performance of an inference model trained in a supervised manner depends substantially on the size and quality of the labeled training data. The characteristics of the underlying data distribution significantly impact the generalization ability of a trained model, particularly in cases where some class overlap can be observed. In such cases, training a single model on the entirety of the labeled data can result in an increase in the complexity of the resulting decision boundary, leading to over-fitting and consequently to some poor generalization performance. In the current work, a cluster-based sample weighting approach is proposed in order to improve the generalization ability of a classification model while dealing with such complex data distributions. The approach consists of first performing a clustering of the training data and subsequently optimizing cluster-specific classification models, using a weighted loss based on the samples-to-cluster-center distances. An unseen sample is first assigned a cluster and subsequently classified based on the model specific to its assigned cluster. The proposed approach was evaluated on three different pain recognition datasets, and the performed evaluation showed that the approach is not only able to attain state-of-the-art classification performances but also systematically outperforms its single model counterpart.

Model transformation languages are domain-specific languages, which are designed to comfortably define transformations. With the increasing use of transformations in various domains, the complexity and size of input models are also increasing. However, developers often lack suitable models for performance testing. We have therefore conducted experiments in which we predict the performance of model transformations based on characteristics of input models using machine learning approaches. In particular, we focused on how to predict the performance of transformations that also transform attributes whose values can have arbitrary size. This dataset contains our raw and processed input data, the scripts necessary to repeat our experiments, and the results we obtained. Our input data consists of the time measurements for six different transformations defined in the Atlas Transformation Language (ATL), as well as the collected characteristics of the real-world input models we used. In this data set, we provide the script that implements our experiments. We predict the execution time of ATL transformations using the machine learning approaches linear regression, random forests and support vector regression using a radial basis function kernel. We also investigate different sets of characteristics of input models as input for the machine learning approaches. These are described in detail in the provided documentation.pdf. The results of the experiments are provided as raw data in individual cvs files. Furthermore, we provide our Eclipse plugin, which collects the characteristics for a set of given models. A detailed documentation is available in documentaion.pdf.

The successful integration of neural networks in a clinical setting is still uncommon despite major successes achieved by artificial intelligence in other domains. This is mainly due to the black box characteristic of most optimized models and the undetermined generalization ability of the trained architectures. The current work tackles both issues in the radiology domain by focusing on developing an effective and interpretable cardiomegaly detection architecture based on segmentation models. The architecture consists of two distinct neural networks performing the segmentation of both cardiac and thoracic areas of a radiograph. The respective segmentation outputs are subsequently used to estimate the cardiothoracic ratio, and the corresponding radiograph is classified as a case of cardiomegaly based on a given threshold. Due to the scarcity of pixel-level labeled chest radiographs, both segmentation models are optimized in a semi-supervised manner. This results in a significant reduction in the costs of manual annotation. The resulting segmentation outputs significantly improve the interpretability of the architecture’s final classification results. The generalization ability of the architecture is assessed in a cross-domain setting. The assessment shows the effectiveness of the semi-supervised optimization of the segmentation models and the robustness of the ensuing classification architecture.

In recent years, several deep learning approaches have been successfully applied in the field of medical image analysis. More specifically, different deep neural network architectures have been proposed and assessed for the detection of various pathologies based on chest X-ray images. While the performed assessments have shown very promising results, most of them consist in training and evaluating the performance of the proposed approaches on a single data set. However, the generalization of such models is quite limited in a cross-domain setting, since a significant performance degradation can be observed when these models are evaluated on data sets stemming from different medical centers or recorded under different protocols. The performance degradation is mostly caused by the domain shift between the training set and the evaluation set. To alleviate this problem, different unsupervised domain adaptation approaches are proposed and evaluated in the current work, for the detection of cardiomegaly based on chest X-ray images, in a cross-domain setting. The proposed approaches generate domain invariant feature representations by adapting the parameters of a model optimized on a large set of labeled samples, to a set of unlabeled images stemming from a different data set. The performed evaluation points to the effectiveness of the proposed approaches, since the adapted models outperform optimized models which are directly applied to the evaluation sets without any form of domain adaptation.

Onkologische Therapien sind von Nebenwirkungen begleitet, die in unterschiedlichem Ausmaß die Lebensqualität einschränken und Einfluss auf die zeit- und dosisgerechte Applikation der Systemtherapie haben können. Therapieassoziierte Nebenwirkungen frühzeitig zu erkennen und zu behandeln ist daher ein entscheidender Faktor für Therapiecompliance und Lebensqualität der Patient*innen. Aktuell werden Nebenwirkungen oder Unverträglichkeiten in der Regel bei den Arztbesuchen, meist im Abstand von 14-28 Tagen erfasst. Eine Untersuchung des MSKCC zeigte, dass ein häufigeres Monitoring von Nebenwirkungen durch Patient*innen selbst die gesundheitsbezogene Lebensqualität (health related quality of life, HRQL) von Patient*innen während einer Chemotherapie verbessern kann. [1].

The overall classification performance as well as generalization ability of a traditional information fusion architecture (built upon so called handcrafted features) is limited by its reliance on specific expert knowledge in the underlying domain of application. The integration of both feature engineering and fusion parameters’ optimization in a single optimization process using deep neural networks has shown in several domains of application (e.g. computer vision) its potential to significantly improve not just the inference performance of a classification system, but also its ability to generalize and adapt to unseen but related domains. This is done by enabling the designed system to autonomously detect, extract and combine relevant information directly from the raw signals accordingly to the classification task at hand. The following work focuses specifically on pain recognition based on bio-physiological modalities and consists of a summary of recently proposed deep fusion approaches for the aggregation of information stemming from a diverse set of physiological signals in order to perform an accurate classification of several levels of artificially induced pain intensities.

Model transformation languages are special-purpose languages, which are designed to define transformations as comfortably as possible, i.e., often in a declarative way. With the increasing use of transformations in various domains, the complexity and size of input models are also increasing. However, developers often lack suitable models for performance testing. We have therefore conducted experiments in which we predict the performance of model transformations based on characteristics of input models using machine learning approaches. This dataset contains our raw and processed input data, the scripts necessary to repeat our experiments, and the results we obtained. Our input data consists of the time measurements for six different transformations defined in the Atlas Transformation Language (ATL), as well as the collected characteristics of the real-world input models that were transformed. We provide the script that implements our experiments. We predict the execution time of ATL transformations using the machine learning approaches linear regression, random forests and support vector regression using a radial basis function kernel. We also investigate different sets of characteristics of input models as input for the machine learning approaches. These are described in detail in the provided documentation.pdf. The results of the experiments are provided as raw data in individual cvs files. Additionally, we calculated the mean absolute percentage error in % and the 95th percentile of the absolute percentage error in % for each experiment and provide these results. Furthermore, we provide our Eclipse plugin, which collects the characteristics for a set of given models, the Java projects used to measure the execution time of the transformations, and other supporting scripts, e.g. for the analysis of the results. A short introduction with a quick start guide can be found in README.md and a detailed documentation in documentaion.pdf.

Predicting the execution time of model transformations can help to understand how a transformation reacts to a given input model without creating and transforming the respective model. In our previous data set (https://doi.org/10.5281/zenodo.8385957), we have documented our experiments in which we predict the performance of ATL transformations using predictive models obtained from training linear regression, random forest and support vector regression. As input for the prediction, our approach uses a characterization of the input model. In these experiments, we only used data from real models. However, a common problem is that transformation developers do not have enough models available to use such a prediction approach. Therefore, in a new variant of our experiments, we investigated whether the three considered machine learning approaches can predict the performance of transformations if we use data from generated models for training. We also investigated whether it is possible to achieve good predictions with smaller training data. The dataset provided here offers the corresponding raw data, scripts, and results. A detailed documentation is available in documentaion.pdf.

In general, classification tasks can differ significantly in their task complexity. For instance, image-based differentiation between vehicles and pedestrians is most likely expected to be less complex than CT-scan-based differentiation between several lung diseases. Intuitively, based on a human point of view, one can identify some classification tasks as more complex than other classification tasks. Moreover, based on expert knowledge and/or task-specific meta information, one could attempt to estimate the complexity ranks of specific classification tasks. In this work, based on the publicly available BioVid Heat Pain Database (BVDB), we experimentally confirm the intuitive assumption that the task of automated pain intensity recognition (PIR) is very challenging. Inspired by the field of chaos theory, we show that the BVDB-specific PIR task can not only be seen as highly complex, but is even identified as a classification task of chaotic nature. To this end, we apply Hao’s working definition for chaotic systems and provide an experiment-based chaos check method. To validate our approach, as a non-complex counterpart, we include a task of handwritten numerals distinction. Our study provides two main contributions, i.e.: i) an enhanced understanding for the still present and – more importantly – substantial gap between the ground truth and the predictions reported by different research groups in combination with automated PIR tasks; and ii) an approach for a numerical complexity check based on chaos theory. Different research directions are discussed for future work. Note that improving PIR accuracy performance is not part of the study objective.

Mobile applications have increasingly entered the healthcare sector. Besides being daily companions, so-called mHealth applications have the potential to enable individuals to collect data, document issues, and share them with healthcare professionals to better adjust medical treatment, side effects, or quality of life. While patient empowerment should be a paramount goal, the setup of these applications in a reliable and communication-effective way is under discussion. In particular, including mHealth applications in the clinical practice routine is crucial to boost their development. Security concerns are of utmost importance as such applications deal with personal data. Considering the sensitive nature of many of the involved data, a trustworthy transfer protocol to the respective health care providers is essential to convince potential users. On the same grounds, healthcare providers, which represent another major stakeholder, might be skeptical of utilizing mHealth applications. This issue is often not prioritized by app developers, and there is a multitude of apps lacking clear and transparent data transfer concepts with a focus on both security and usability. In the following, we present and discuss two different approaches for managing and reporting sensitive clinical information and their secure inter-sectoral transfer. Both use cases are currently implemented into clinical practice, and their applicability is under constant evaluation. Besides, to empower inter-sectoral communication, both approaches have been developed in close collaboration with healthcare providers to maximize both communication and effectiveness of the mHealth applications. Based on our work, we conclude that while mHealth applications can be important in many aspects of improving health care, there are often significant limitations of mHealth-based communication, which can hamper its integration in clinical settings. To overcome these limitations, we show how to apply and re-elaborate on existing security and communication strategies. Finally, we highlight how these approaches can strengthen both patient and healthcare professionals' empowerment.

The performance of multiple classifier systems can be significantly improved by the use of intelligent classifier combination approaches. In this study, we introduce a novel late fusion architecture, which can be interpreted as a combination of the well-known mixture of experts and stacked generalization methods. Our proposed method aggregates the outputs of classification models and corresponding sample-specific weighting models. A special feature of our proposed architecture is that each weighting model is trained on an individual set of meta labels. Using individual sets of meta labels allows each weighting model to separate regions, on which the predictions of the corresponding classification model can be associated to an estimated confidence value. We test our proposed architecture on a set of publicly available databases, including different benchmark data sets. The experimental evaluation shows the effectiveness and potential of our proposed method. Moreover, we discuss different approaches for further improvement of our proposed architecture.

The occurrence of adverse events frequently accompanies tumor treatments. Side effects should be detected and treated as soon as possible to maintain the best possible treatment outcome. Besides the standard reporting system Common Terminology Criteria for Adverse Events (CTCAE), physicians have recognized the potential of patient-reporting systems. These are based on a more subjective description of current patient reporting symptoms. Patient-reported symptoms are essential to define the impact of a given treatment on the quality of life and the patient's wellbeing. They also act against an underreporting of side effects which are paramount to define the actual value of a treatment for the individual patient. Here, we present a study protocol for a clinical trial that assesses the potential of a smartphone application for CTCAE conform symptom reporting and tracking that is adjusted to the standard clinical reporting system rather than symptom oriented descriptive trial tools. The presented study will be implemented in two parts, both lasting over six months. The first part will assess the feasibility of the application with 30 patients non-randomly divided into three equally-sized age groups (<55years, 55-75years, >75years). In the second part 36 other patients will be randomly assigned to two groups, one reporting using the smartphone and one not. This prospective second part will compare the impact of smartphone reported adverse events regarding applied therapy doses and quality of life to those of patients receiving standard care. We aim for early detection and treatment of adverse events in oncological treatment to improve patients' safety and outcomes. For this purpose, we will capture frequent adverse events of chemotherapies, immunotherapies, or other targeted therapies with our smartphone application. The presented trial is registered at the U.S. National Library of Medicine ClinicalTrials.gov (NCT04493450) on July 30, 2020.

The subjective nature of pain makes it a very challenging phenomenon to assess. Most of the current pain assessment approaches rely on an individual's ability to recognise and report an observed pain episode. However, pain perception and expression are affected by numerous factors ranging from personality traits to physical and psychological health state. Hence, several approaches have been proposed for the automatic recognition of pain intensity, based on measurable physiological and audiovisual parameters. In the current paper, an assessment of several fusion architectures for the development of a multi-modal pain intensity classification system is performed. The contribution of the presented work is two-fold: (1) 3 distinctive modalities consisting of audio, video and physiological channels are assessed and combined for the classification of several levels of pain elicitation. (2) An extensive assessment of several fusion strategies is carried out in order to design a classification architecture that improves the performance of the pain recognition system. The assessment is based on the SenseEmotion Database and experimental validation demonstrates the relevance of the multi-modal classification approach, which achieves classification rates of respectively 83.39%, 59.53% and 43.89% in a 2-class, 3-class and 4-class pain intensity classification task.

Traditional pain assessment approaches ranging from self-reporting methods, to observational scales, rely on the ability of an individual to accurately assess and successfully report observed or experienced pain episodes. Automatic pain assessment tools are therefore more than desirable in cases where this specific ability is negatively affected by various psycho-physiological dispositions, as well as distinct physical traits such as in the case of professional athletes, who usually have a higher pain tolerance as regular individuals. Hence, several approaches have been proposed during the past decades for the implementation of an autonomous and effective pain assessment system. These approaches range from more conventional supervised and semi-supervised learning techniques applied on a set of carefully hand-designed feature representations, to deep neural networks applied on preprocessed signals. Some of the most prominent advantages of deep neural networks are the ability to automatically learn relevant features, as well as the inherent adaptability of trained deep neural networks to related inference tasks. Yet, some significant drawbacks such as requiring large amounts of data to train deep models and over-fitting remain. Both of these problems are especially relevant in pain intensity assessment, where labeled data is scarce and generalization is of utmost importance. In the following work we address these shortcomings by introducing several novel multi-modal deep learning approaches (characterized by specific supervised, as well as self-supervised learning techniques) for the assessment of pain intensity based on measurable bio-physiological data. While the proposed supervised deep learning approach is able to attain state-of-the-art inference performances, our self-supervised approach is able to significantly improve the data efficiency of the proposed architecture by automatically generating physiological data and simultaneously performing a fine-tuning of the architecture, which has been previously trained on a significantly smaller amount of data.

. In this study, we evaluate a person independent pain intensity recognition task, based on the BioVid Heat Pain Database. Previous works show that for such classification tasks, the overall performance can be increased by reducing the training data, based on certain criteria, such as different distance measures. This results in considering only a certain amount of participants from the training set, whose data distributions are defined to be the most similar to the data distribution of the participant from the test set. Counterintuitively, we propose to remove participants, which are identified as central points, from the training set, completely independent from the test set. Our evaluations show that this approach can lead to significant improvement of classification accuracy.

Several approaches have been proposed for the analysis of pain-related facial expressions. These approaches range from common classification architectures based on a set of carefully designed handcrafted features, to deep neural networks characterised by an autonomous extraction of relevant facial descriptors and simultaneous optimisation of a classification architecture. In the current work, an end-to-end approach based on attention networks for the analysis and recognition of pain-related facial expressions is proposed. The method combines both spatial and temporal aspects of facial expressions through a weighted aggregation of attention-based neural networks’ outputs, based on sequences of Motion History Images (MHIs) and Optical Flow Images (OFIs). Each input stream is fed into a specific attention network consisting of a Convolutional Neural Network (CNN) coupled to a Bidirectional Long Short-Term Memory (BiLSTM) Recurrent Neural Network (RNN). An attention mechanism generates a single weighted representation of each input stream (MHI sequence and OFI sequence), which is subsequently used to perform specific classification tasks. Simultaneously, a weighted aggregation of the classification scores specific to each input stream is performed to generate a final classification output. The assessment conducted on both the BioVid Heat Pain Database (Part A) and SenseEmotion Database points at the relevance of the proposed approach, as its classification performance is on par with state-of-the-art classification approaches proposed in the literature.

Pain is the result of a complex interaction among the various parts of the human nervous system. It plays an important role in the diagnosis and treatment of patients. The standard method for pain recognition is self-report; however, not all patients can communicate pain effectively. In this work, the task of automated pain recognition is addressed using para-linguistic and physiological data. Hand-crafted and automatically generated features are extracted and evaluated independently. Several state-of-the-art machine learning algorithms are applied to perform subject-independent binary classification. The SenseEmotion dataset is used for evaluation and comparison. Random forests trained on hand-crafted features from the physiological modalities achieved an accuracy of 82.61%, while support vector machines trained on hand-crafted features from the para-linguistic data achieved an accuracy of 63.86%. Hand-crafted features outperformed automatically generated features.

Pain intensity recognition still constitutes a challenging classification task. In this work, we focus on the physiological signals of the publicly available BioVid Heat Pain Database, which was collected at Ulm University. The BioVid Heat Pain Database consists of different recordings of healthy test subjects that were exposed to various short-time heat stimuli. The results reported in the literature, which are based on those short-time sequences do not justify the implementation of automated pain detection systems, due to unsatisfactory accuracy rates. In the current study, we show that the outcomes, which are stated in the literature, most likely represent lower bound estimations. For this purpose, we transfer the classification task, which is provided by the BioVid Heat Pain Database, to a real-world scenario. More precise, according to an expected hospital setting, we analyse the automated pain intensity recognition approach in combination with different sets of short-time sequences. Our outcomes indicate, that in real-world applications, where the detection of pain intensity is based on more than one single short-time sequence, the accuracy values can be significantly improved. In the current study, the classification performance of bagged decision tree ensembles is evaluated, based on a person-independent scenario.