Longitudinal analysis of high-dimensional biomedical data offers a transformative window into the transition from health to metabolic disease. Despite advancements in predictive modeling, the integration of temporal proteomic shifts with clinical phenotypes for type 2 diabetes mellitus (T2DM) risk assessment remains underdeveloped, particularly within underrepresented Middle Eastern populations. We developed an interpretable, deep-learning pipeline utilizing tabular network (TabNet), an attention-based architecture designed for tabular data, to analyze a longitudinal cohort from the Qatar biobank (QBB). The framework integrates multimodal inputs, including proteomic signatures, dual-energy X-ray absorptiometry (DXA) body composition metrics, and electronic health records (EHR). Model transparency was ensured through built-in feature selection and SHAP (SHapley Additive exPlanations) values. The performance of the TabNet model was rigorously benchmarked against gradient-boosting machines, support vector machines, and random forest. Our findings demonstrate that the TabNet significantly outperforms conventional machine learning benchmarks in capturing the non-linear trajectories of diabetes progression. The model identified a distinct cluster of proteomic markers and visceral adiposity metrics that serve as early indicators of T2DM. Age-and-gender-stratified analyses revealed significant demographic divergence in risk profiles; specifically, the predictive weight of inflammatory markers and visceral fat mass exhibited profound variance across age cohorts, suggesting the need for life-stage-specific screening protocols. This study establishes an explainable artificial intelligence (XAI)-driven framework that bridges the gap between molecular discovery and clinical risk stratification. By uncovering the demographic-specific drivers of T2DM, our results provide a foundation for personalized preventative medicine and highlight the efficacy of attention-based deep learning in managing complex, longitudinal biobank data.

This paper introduces a novel federated learning framework termed LoRa-FL designed for training low-rank one-shot image detection models deployed on edge devices. By incorporating low-rank adaptation techniques into one-shot detection architectures, our method significantly reduces both computational and communication overhead while maintaining scalable accuracy. The proposed framework leverages federated learning to collaboratively train lightweight image recognition models, enabling rapid adaptation and efficient deployment across heterogeneous, resource-constrained devices. Experimental evaluations on the MNIST and CIFAR10 benchmark datasets, both in an independent-and-identically-distributed (IID) and non-IID setting, demonstrate that our approach achieves competitive detection performance while significantly reducing communication bandwidth and compute complexity. This makes it a promising solution for adaptively reducing the communication and compute power overheads, while not sacrificing model accuracy.

Automated diabetic retinopathy (DR) grading is crucial for disease monitoring and personalized treatment, challenged by high intra-class variation and data imbalance. This paper presents a novel approach to enhancing DR grading detection by integrating linear-Radon sinogram-based images with original retinal images to provide a multimodal network using different convolutional neural network (CNN) architectures. Using the Kaggle Aptos dataset, we evaluated the performance of this multimodal integration. Our findings reveal a significant improvement in multi-class classification performance compared to unimodal retina-only images, underscoring our method's ability to detect subtle patterns among different DR grades. This study underscores the potential of sinogram-based images as a valuable modality for DR grading and paves the way for future research to validate this approach across diverse datasets and explore the application of curve-based sinograms,

Many applications seek to make sense of high volume streaming data from social media by identifying spatiotemporal patterns. Events, representing topics that emerge and decay over time, are detected by monitoring for changes in the language being used, but typical approaches do not consider the localisation of events in cities and countries, and within hours, days, and weeks. This work develops and evaluates a new approach to event localisation and ranking that can be applied to Twitter data streams. The proposed approach models the use of language in tweets per city per hour to produce a model that can be used to detect the magnitude of unexpected changes in the use of the language. The approach uses a spatiotemporal lattice structure and a method for traversing between hours, days, and weeks, as well as cities, regions, and countries to identify anomalies in the language used across millions of tweets. The output is a ranked list of events comprising a list of tweets posted within a location and period of time, and characterized by language features of interest. The approach was implemented and tested by comparing events detected across five example domains (suicide, shooting, elections, sports, and sentiment) using 11.7 million tweets from users located in 100 cities and posted within the 203-day study period. Experiments demonstrate that the approach can detect events across a range of application domains.

In a smart grid distribution management system, operation, planning, forecasting and decision making relies on demand-side management functions, which require real-time smart grid data. This data has significant dollar value because it is extremely useful for efficient control and intelligent prediction of the energy consumption, and expert management of residential and commercial load. However, the huge amount of (smart grid) data generated at a very high velocity poses a number of challenges. Utility companies have a huge demand for efficient summarization techniques to mine interesting patterns and extracting useful and actionable intelligence. Research from various domains has shown that data summarization can significantly improve the scalability and efficiency of various data analytic tasks (e.g., transactional database mining, data streams mining, network monitoring). This paper proposes a summarization approach (i.e., a set of algorithms, data structures, and query mechanisms) that enables the utility company to accurately infer various energy consumption patterns in real-time by automatic monitoring of smart grid data using significantly less computational resources. The proposed summarization approach is suitable for processing spatiotemporal streams, and it can also provide answers in real-time to various smart grid applications (e.g., demand-side management, direct load control, smart pricing and Volt-VAr control). Both theoretical bound and experimental evaluation are presented in this paper, which shows that the memory required for the proposed data structure grows linearly for the first 52 weeks; but interestingly, after the first year, the memory growth is negligible. The experimental results show that the proposed approach can process around 4 million smart meter readings every second or 120 million readings every minute. The proposed approach outperforms widely commercially used Database Management Systems (DBMSs) in terms of update and query costs: it is about 200 times faster than DBMSs in terms of update time, and about 340 times faster than DBMSs in terms of query time.

There are many application domains, where hierarchical data is inherent, but surprisingly, there are few techniques for mining patterns from such important data. Hierarchical Heavy Hitters (HHH) and multilevel and Cross-Level Association Rules (CLAR) mining are well-known hierarchical pattern mining techniques. The problem in these techniques; however, is that they focus on capturing only global patterns from data but cannot identify local contextual patterns. Another problem in these techniques is that they treat all data items in the transaction equally and do not consider the sequential nature of the relationship among items within a transaction; hence, they cannot capture the correlation semantic within the transactions of the data items. There are many applications such as clickstream mining, healthcare data mining, network monitoring, and recommender systems, which require to identify local contextual patterns and correlation semantics. In this work, we introduce a new concept, which can capture the sequential nature of the relationship between pairs of hierarchical items at multiple concept levels and can capture local contextual patterns within the context of the global patterns. We call this notion Hierarchically Correlated Heavy Hitters (HCHH). Specifically, the proposed approach finds the correlation between items corresponding to hierarchically discounted frequency counts. We have provided formal definitions of the proposed concept and developed algorithmic approaches for computing HCHH in data streams efficiently. The proposed HCHH algorithm have deterministic error guarantees, and space bounds. It requires O(η/ϵpϵs) memory, where h is a small constant, and ϵp∈ [0,1], ϵs∈ [0,1] are user defined parameters on upper bounds of estimation error. We have compared the proposed HCHH concept with existing hierarchical pattern mining approaches both theoretically as well as experimentally.

Distributed data stream mining in a sliding window has emerged recently, due to its applications in many domains including large Telecoms and Internet Service Providers, financial tickers, ATM and credit card operations in banks and transactions in retail chains. Many of these large-scale applications prohibit monitoring data centrally at a single location due to their massive volume of the data; therefore, data acquisition, processing, and mining tasks are often distributed to a number of processing nodes, which monitor their local streams and exchange only the summary of data either periodically or on demand. While this offer many advantages, distributed stream applications possess significant challenges including problems related to an online analysis of the recent data, communication efficiency and various estimation of various complex queries. There are few existing techniques which solve problems related to distributed sliding window data stream; however, those techniques are focused on solving only simple problems and require high space, query, and communication cost, which can be a bottleneck for many of these large scale applications. In this paper, we propose an efficient query estimation technique by constructing a small sketch of the data stream. The constructed sketch uses a deterministic sliding window model and can estimate various complex queries, for both centralized and distributed applications; including point queries (i.e., range queries and heavy hitter queries), quantiles, inner product, and self-join size queries, with deterministic guarantees on the precision. The proposed approach improves upon recent existing work for these problems, in terms of the memory and query cost in a centralized setting and in terms of communication cost and merge complexity in a distributed setting. It requires O(1/ε21 log (εN)) memory (where 0 <; ε <; 1 is a user defined parameter), can provide estimates in O(1) time, and processes each incoming record in O(1) amortized time. Detailed experimental analysis, both in centralized and distributed settings demonstrates that in practice the proposed approach uses about six times less memory, and has about eight times less query time when compared to ECM sketches. In a distributed application, the proposed technique also significantly improves (around seven times) on the communication cost between distributed sites.

Recently, there has been significant increase in interest on Smart Grid security. Researchers have proposed various techniques to detect cyber-attacks using sensor data. However, there has been little work to distinguish a cyber-attack from a power system physical fault. A serious operational failure in physical power grid may occur from the mitigation strategies if fault is wrongly classified as a cyber-attack or vice-versa. In this paper, we utilize a data-driven approach to accurately differentiate the physical faults from cyber-attacks. First, we create a realistic dataset by generating different types of faults and cyber-attacks on the IEEE 30 bus benchmark test system. With extensive experiments, we observe that most of the established supervised methods perform poorly for the classification of faults and cyber-attacks specially for the practical datasets. Hence, we provide a data-driven approach where labelled data are projected in a new low-dimensional subspace using Principal Component Analysis (PCA). Next, Sequential Minimal Optimization (SMO) based Support Vectors are trained using the new projection of the original dataset. With both simulated and practical datasets, we have observed that the proposed classification method outperforms other existing popular supervised classification approaches considering the cyber-attack and fault datasets.

We have developed an efficient summarization paradigm for data drawn from hierarchical domain to construct a succinct view of important large-valued regions (“heavy hitters”). It requires one pass over the data with moderate number of updates per element of the data and requires lesser amount of memory space as compared to existing approaches for approximating hierarchically discounted frequency counts of heavy hitters with provable guarantees. The proposed technique is generic that can make use of existing state-of-the-art sketch-based or count-based frequency estimation approaches. Any algorithm from both of these families can be coupled as a subroutine in the proposed framework without any substantial modifications. Experimental as well as theoretical justifications have been provided for its significance.