Background and objectives: The COVID-19 pandemic raised awareness of the complexities of the patient, the disease, and the practice of medicine. The impact of these reaches beyond healthcare (e.g., supply chains, politics, socioeconomic factors) to include nations, individuals, and molecules. In personalized medicine, "accurate diagnosis" is critical as it affects patient management, clinical trial recruitment, regulatory approval, and reimbursement policies for payers. Conventional statistics evaluate hypothesis-driven reductionist practices in medicine, e.g., the use of "scores" combining individual measurements, and are often limited by the data:variables ratio. True personalization (N of 1) is not practical but better stratification of diseases and patients can improve diagnoses. This work describes our approach and tests its ability to identify patient complexity and clinical markers in the trial of a candidate HFpEF drug better than prior methods. Methods: This study evaluated discovery or data-driven approaches, by applying community detection (CD), forgoing statistical significance to identify unknown relationships. We reanalyzed data from the I-PRESERVE study of heart failure with preserved-ejection fraction, where subgroup analysis was unsuccessful. We initially performed unipartite CD analysis and evolved to address the complexity in real-world data using a bipartite model. The mathematically grounded modularity metric enabled greater confidence in community assignments. Results: This reanalysis with CD revealed novel patient subgroups with stronger supporting rationale for group assignments, pointing to further refined and targeted studies. Conclusions: We believe that generalization of the CD approach to higher-dimensional data can lead to a "next generation of phenotyping" that encompasses the temporal progression of the patient

Nowadays, a lot of data is in the form of Knowledge Graphs aiming at representing information as a set of nodes and relationships between them. This paper proposes an efficient framework to create informative embeddings for node classification on large knowledge graphs. Such embeddings capture how a particular node of the graph interacts with his neighborhood and indicate if it is either isolated or part of a bigger clique. Since a homogeneous graph is necessary to perform this kind of analysis, the framework exploits the metapath approach to split the heterogeneous graph into multiple homogeneous graphs. The proposed pipeline includes an unsupervised attentive neural network to merge different metapaths and produce node embeddings suitable for classification. Preliminary experiments on the IMDb dataset demonstrate the validity of the proposed approach, which can defeat current state-of-the-art unsupervised methods.

Modern mobile devices are able to provide context-aware and personalized services to the users, by leveraging on their sensing capabilities to infer the activity and situation in which a person is currently involved. Current solutions for context-recognition rely on annotated data and experts' knowledge to predict the user context. In addition, their prediction ability is strongly limited to the set of situations considered during the model training or definition. However, in a mobile environment, the user context continuously evolves, and it cannot be merely restricted to a set of predefined classes. To overcome these limitations, we propose COMPASS, a novel unsupervised and online clustering algorithm aimed at identifying the user context in mobile environments based on the stream of high-dimensional data generated by smartphone sensors. COMPASS can distinguish an arbitrary number of user's contexts from the sensors' data, without defining a priori the collection of expected situations. This key feature makes it a general-purpose solution to provide context-aware features to mobile devices, supporting a broad set of applications. Experimental results on 18 synthetic and 2 real-world datasets show that COMPASS correctly identifies the user context from the sensors' data stream, and outperforms the state-of-the-art solutions in terms of both clusters configuration and purity. Eventually, we evaluate its performances in terms of execution time and the results show that COMPASS can process 1000 high-dimensional samples in less than 20 s, while the reference solutions require about 60 min to evaluate the entire dataset.

This paper presents a complete framework aimed to nondestructive inspection of composite materials. Starting from the acquisition, performed with lock-in thermography, the method flows through a set of consecutive blocks of data processing: input enhancement, feature extraction, classification and defect detection. Experimental results prove the capability of the presented methodology to detect the presence of defects underneath the surface of a calibrated specimen made of Glass Fiber Reinforced Polymer (GFRP). Results are also compared with those obtained by other techniques, based on different features and unsupervised learning methods. The comparison further proves that the proposed methodology is able to reduce the number of false positives, while ensuring the exact detection of subsurface defects.

Currently available technological solutions do not allow to reliably detect falls in the elderly, due to still-open issues on both sensing and processing sides. In fact, the most promising sensing approaches raise concerns for privacy issues (e.g., vision-based approaches) or low acceptability rate (e.g., wearable-based approaches); whereas on the processing side, commonly used methodologies are based on supervised techniques trained with both positive (falls) and negative (non-fall) samples, both simulated by healthy young subjects. As a result of such a training protocol, fall detectors inevitably exhibit lower performance when used in real-life conditions, in which monitored subjects are older adults. In order to address the problem of fall detection under real-life conditions, this study investigates privacy-preserving unobtrusive sensing technologies together with an unsupervised methodology trained exclusively on daily activity (non-fall) data from the monitored elderly subject. Preliminary results are very encouraging, showing the effectiveness to achieve a good detection performance and, most importantly, which is more reproducible in real-life scenarios.