This paper presents a novel framework for constructing a diabetes-specific knowledge graph (KG) using a streamlined multi-agent system powered by Gemini-based Large Language Models (LLMs). Leveraging insights from the 2016 National Diabetes Survey (NNDS) conducted by the National Diabetes Education Program (NDEP), the framework extracts critical variables related to diagnosis, risk perception, medical advice, and self-management practices across diverse U.S. populations. By processing data from the NNDS’s extensive 94-question survey, the methodology performs adaptive ontology mapping using APIs for six major medical standards (e.g., SNOMED CT, ICD-11), ensuring semantic interoperability. Relationships between variables are identified and structured using RDF, RDFS, and OWL standards. The integration of LLMs with ontology tools like Protégé enhances automation and scalability. Results demonstrate the framework’s effectiveness in generating contextually rich and clinically relevant knowledge graphs, providing a robust foundation for advancing healthcare informatics and personalized diabetes management.

Training a neural network model from scratch is a computationally intensive operation. To alleviate this issue, researchers often employ "transfer learning" that transfers knowledge from a source data distribution to a target data distribution, instead of training the whole model from scratch. Typically, the last few layers of a pretrained convolutional neural network (CNN) are chosen for many transfer learning tasks where the outputs of those selected layers are combined to construct a feature space based on which a task-specific classification n etwork i s t rained o r fi ne-tuned. Th is arbitrary way of selecting layers, however, often fails to achieve the desired accuracy for the target task. What we need is an intelligent way of selecting layers from a pretrained model for a given task so that the additional overhead of successive training remains low. To this end, we propose a novel method, called TransJury, to find t he m ost s ignificant la yers from a pretrained mo del for transfer learning along with preserving the knowledge for the source domain. Through extensive experimentation on several target domain datasets, we show the supremacy of our approach in terms of lower training overhead and improved accuracy. By deploying MobileNet-v2, a lightweight CNN model pretrained on the ImageNet dataset on an edge device, we also discuss the future direction of this research.

There are increasingly leveraging public data with cities increasingly interested in driving both responsiveness to citizen demands and cost savings through data analytics. As public managers seek to augment existing data sources, such as 311 complaints, with existing secondary data, such as US Census products, severe challenges exist. This paper considers the problems inherent in data being collected at divergent geographic levels over different time horizons. An inductive analytical methodology is developed to create units of analysis that are both useful and analytically appropriate for public managers and policy leaders in urban areas. A big data analytics framework for public data, called BDAP, was presented predictive analytics for community need considering data the spatial and temporal location while addressing the data issues such as missing values, privacy-preserving, and predictive modeling. The findings illustrate the power of inductive data curation and privacy-preserving leading to benefits to the big data community. An application for the Open Data Platform was developed using KCMO’s 311 data, crime data and census data.

Many people around the world suffer from neurological impairments and face many difficulties in performing their routine daily tasks. Caregiver services can become exhausted, expensive, and depressing for the patient over time. In the last few decades, numerous research is conducted on developing BCIs that can help a neurological impaired individual to communicate and functionally rehabilitate. Advances in the field of the Internet of Things (IoT) has opened the door for new opportunities in smart assistive technologies, like smart home, elderly care, or robotic applications. In this paper, we present an IoT-based non-invasive brain-computer interface (BCI) system. This system uses electroencephalography (EEG) signals to help severely paralyzed and locked-in patients regain the ability to do a simple daily task, such as locking/unlocking the door, turning on/off the lights or TV, and communicate with others, caregivers or healthcare providers, by sending text notifications. For our BCI, we have also built a deep learning classifier to understand the patient intent and trigger the corresponding workflow on connected IoT devices. This model has achieved a global accuracy of 72.82% for 103 subjects and 93.33% subject specific transfer learning for 10 subjects on the EEG Motor Movement/Imagery (MI) dataset [1] for MI tasks for 2 classes. Our BCI is also integrable with EEG headsets (like Emotiv) and IoT devices like Google Home and Alexa.

In metropolitan cities, outdoor air pollution and ambient outdoor noise transmission are significant environmental hazards, degrading indoor environmental quality. Natural ventilation is often a viable option for diluting indoor air pollutants, but transportation noise transmission is a crucial ecological conflict. Therefore, personal control over the concentration of outdoor/indoor air pollutants and noise levels is a significant threshold for natural ventilation availability. This study proposes real-time data detection and notification solutions using sensors and artificial intelligence (AI) to improve indoor air quality and outdoor air quality and outdoor noise transmission. An intelligent real-time detection and notification system was implemented in a distributed computing framework with cloud and edge computing. The objective of this study is placed on three facets: (1) development of sensors and NVIDIA Jetson Nano prototype for air quality and noise level detection, (2) application of machine learning for air quality and noise level prediction and classification, and (3) web interface for real-time monitoring and prediction for air quality and noise level detection. The results showed that the proposed user interface provides building occupants with real-time data of outdoor/indoor Air Quality Index (AQI) and noise levels for the optimized occupant control over Indoor Air Quality (IAQ). The personal control over indoor environmental quality (IEQ) enables occupants to promote natural ventilation behaviors and integrate with the existing building system on optimized IEQ by interacting with AI-based real-time data.

Heart diseases are considered the leading cause of death globally. Early diagnosis of disease may help give appropriate prescribing medicines, which may help control and reduce conditions. The current clinical diagnosis methods such as Electrocardiograms, computed tomography, echocardiogram, Magnetic Resonance Imaging, etc. provide valuable information for diagnosis and treatment. However, these techniques are time-intensive, operator-dependent, and expensive. In this paper, we propose a low-cost solution real-time solution to diagnose heart diseases. We suggest an integrated automatic multimodal heart disease classification (AMHDC) system using the Phonocardiogram (PCG) signal. For this purpose, first, we have developed an advanced fusion method using pre-processing techniques such as Data Normalization and Data Augmentation. Secondly, we have extracted the spectrograms from heart sound and used them as features and images for signal and image processing. Finally, we created a real-time integrated Convolutional Neural Network (CNN) model for high-performance heart disease classification. The results show our model outperformed the state-of-art research, whose accuracy is 89.7%., while our model reported accuracy of 93% for audio and 96% for image-based approach.

Human-in-the-Loop has been receiving special attention from the data science and machine learning community. It is essential to realize the advantages of human feedback and the pressing need for manual annotation to improve machine learning performance. Recent advancements in natural language processing (NLP) and machine learning have created unique challenges and opportunities for digital humanities research. In particular, there are ample opportunities for NLP and machine learning researchers to analyze data from literary texts and use these complex source texts to broaden our understanding of human sentiment using the human-in-the-loop approach. This paper presents our understanding of how human annotators differ from machine annotators in sentiment analysis tasks and how these differences can contribute to designing systems for the "human in the loop" sentiment analysis in complex, unstructured texts. We further explore the challenges and benefits of the human-machine collaboration for sentiment analysis using a case study in Greek tragedy and address some open questions about collaborative annotation for sentiments in literary texts. We focus primarily on (i) an analysis of the challenges in sentiment analysis tasks for humans and machines, and (ii) whether consistent annotation results are generated from multiple human annotators and multiple machine annotators. For human annotators, we have used a survey-based approach with about 60 college students. We have selected six popular sentiment analysis tools for machine annotators, including VADER, CoreNLP's sentiment annotator, TextBlob, LIME, Glove+LSTM, and RoBERTa. We have conducted a qualitative and quantitative evaluation with the human-in-the-loop approach and confirmed our observations on sentiment tasks using the Greek tragedy case study.

In this paper, we propose a new framework, DeepLite, for real-time deep learning on the edge. In DeepLite, a network of multiple deep learning models is designed to conduct the context-aware inferencing using real-time deep learning technologies. Comprehensively, DeepLite has several innovative concepts as follows: 1) an inference network of deep learning models/containers, 2) invocation of a new model based on the output of the previous model's inferencing, 3) intelligent containers for models, 4) plug-and-play model/container, and 5) microservice architecture on the edge. DeepLite has been evaluated via a case study, NeighborNets which was based on an inference network of several deep learning models deployed to edge devices for computer vision, e.g., object detection algorithms to compute a diverse range of aspects, such as house types, level of greenery, house age, traffic conditions, and the types of recreational facilities.

In this paper, we propose a new Internet of things (IoT) solution, called the Urban Noise Monitoring (UNM) system, which can classify real-time environmental audio sound using an embedded system such as Raspberry pi 4 and log the data in the Google Cloud. The reported events will be available for future usage, i.e., selection of the safe area for living. The real-time audio classification has been a big challenge for deep learning in environmental sounds due to the high noisy nature of sound. We have implemented a real-time IoT system for urban sound classification and monitored the historical reports generated. We have developed an advanced fusion method using normalization techniques such as peak, RMS, and EBU and an efficient data augmentation method using various factors, including time stretch, pitch shifting, and dynamic range compression. Further, we have integrated the normalization and the augmentation methods into 2D Convolutional Neural Network (CNN) with the TensorFlow framework on Raspberry pi 4 for urban sound classification. Our classification model outperformed the state of the art performance: 95% accuracy with the Urban sound dataset. The outstanding performance confirmed the effectiveness of the proposed method on the IoT system for urban noise monitoring.

Human emotions are essential to recognize the behavior and state of mind of a person. Emotion detection through speech signals has started to receive more attention lately. This paper proposes the method for detecting human emotions using speech signals and its implementation in real-time using the Internet of Things (IoT) based deep learning for the care of older adults in nursing homes. The research has two main contributions. First, we have implemented a real-time system based on audio IoT, where we have recorded human voice and predicted emotions via deep learning. Secondly, for advance classification, we have designed a model using data normalization and data augmentation techniques. Finally, we have created an integrated deep learning model, called Speech Emotion Detection (SED), using a 2D convolutional neural networks (CNN). The best accuracy that was reported by our method was approximately 95%, which outperformed all state-of-the-art approaches. We have further extended to apply the SED model to a live audio sentiment analysis system with IoT technologies for the care of older adults in nursing homes.

Understanding software evolution is an imperative prerequisite for software related activities such as testing, debugging, and maintenance. As a software system evolves, it increases in size and complexity, introducing new challenges of understating the inner system interactions and subsequently hinders the overall system comprehension. While tools that construct and visualize call graphs have been used to facilitate software comprehension, they are still limited to capturing the functionality of a single software system at a time. However, understanding the similarities and differences across multiple releases becomes an imperative and challenging task during software evolution. To this end, we present a tool, named GraphEvo, that focuses on automating the process of quantifying and visualizing the changes across multiple releases of a software system based on an information-theoretic approach to compare the call graphs. Specifically, GraphEvo can automatically (1) construct and visualize the call graph for one or more software releases, (2) calculate and display a set of graph-based metrics, and (3) construct color-coded call graphs to visualize system evolution. The main goal of GraphEvo is to assist software developers and testers in exploring and tracking software changes over time. We demonstrate the functionality of GraphEvo by analyzing and studying five real software systems throughout their entire lifespan. The tool, evaluation results, and a video demo are available at https://goo.gl/8edZ64.

A parallel and distributed machine learning framework are in need to deal with a large amount of data. We have seen unsatisfactory classification performance especially with increasing the number of classes. In this paper, we propose a distributed deep learning framework, called Multi-Class Discriminative Distribution (MCDD) that aims to distribute classes while improving the accuracy performance of the deep learning models with large scale datasets. The MCDD framework works on an evidence-based learning model for the optimal distribution of classes by computing a misclassification cost (i.e., confusion factor). These observations about learning attempts have been used to extend a classifier into a classification model hierarchy by learning an optimal distribution of classes. As a result, a distributed deep neural network model with multi-class classifiers (MCDD) was built to optimize the accuracy and performance of the learning process. The MCDD model runs on parallel environments, such as Apache Spark and Tensor Flow using large real-world datasets (Caltech-101, CIFAR-100, ImageNet-1K) showing that MCDD can build a class distribution model with higher accuracy compared to existing models.

Program comprehension is an imperative and indispensable prerequisite for several software tasks, including testing, maintenance, and evolution. In practice, understanding the software system requires investigating the high-level system functionality and mapping it to its low-level implementation, i. e. source code. The implementation of a software system can be captured using a call graph. A call graph represents the system's functions and their interactions at a single level of granularity. While call graphs can facilitate understanding the inner system functionality, developers are still required to manually map the high-level system functionality to its call graph. This manual mapping process is expensive, time-consuming and creates a cognitive gap between the system's highly-level functionality and its implementation. In this paper, we present an innovative approach that can automatically (1) construct and visualize the static call graph for a system written in Python, (2) cluster the execution paths of the call graph into hierarchal abstractions, and (3) label the clusters according to their major functional behaviors. The goal is to bridge the cognitive gap between the high-level system functionality and its call graph, which can further facilitate system comprehension. To validate our approach, we conducted four case studies including code2graph, Detectron, Flask, and Keras. The results demonstrated that our approach is feasible to construct call graphs and hierarchically cluster them into abstraction levels with proper labels.

The aim of the paper is to perform audio analytics based on the audio sensor data that is continuously monitoring the home environment automatically through an audio Internet of Things (IoT) system. Domestic violence is one of the major problems in many cities nowadays. We have proposed a home automation system where IoT sensors records the audio in home environment continuously and the audio is sent to machine learning server where the audio is split into small clips and classified into different categories. The need of an automatic detection system is urgent for enforcing home safety and safe neighborhood. If IoT system detects any suspicious sound, it generates an emergency notification to nearest emergency services for possible action to be taken. The classification of audio such as gunshots, explosion, glass breaking, screaming and siren is based on shallow learning (Support vector machine, Decision tree, Random forest and Naïve Bayes) and deep learning (Convolutional neural network and Long short-term memory). Our experiments validated that Convolutional Neural Network shows the best performance (89% accuracy) compared to other machine learning algorithms.

An integral step in the discovery of new treatments for medical conditions is the matching of potential subjects with appropriate clinical trials. Eligibility criteria for clinical trials are typically specified as inclusion and exclusion criteria for each study in freetext form. While this is sufficient for a human to guide a recruitment interview, it cannot be reliably and computationally construed to identify potential subjects. Standardization of the representation of eligibility criteria can enhance the efficiency and accuracy of this process. This article presents a semantic framework that facilitates intelligent matchmaking by identifying a minimal set of eligibility criteria with maximal coverage of clinical trials. In contrast to existing top-down manual standardization efforts, a bottom-up data driven approach is presented to find a canonical nonredundant representation of an arbitrary collection of clinical trial criteria. The methodology has been validated with a corpus of 709 clinical trials related to Generalized Anxiety Disorder containing 2,760 inclusion and 4,871 exclusion eligibility criteria. This corpus is well represented by a relatively small number of 126 inclusion clusters and 175 exclusion clusters, each of which corresponds to a semantically distinct criterion. Internal and external validation measures provide an objective evaluation of the method. An eligibility criteria ontology has been constructed based on the clustering. The resulting model has been incorporated into the development of the MindTrial clinical trial recruiting system. The prototype for clinical trial recruitment illustrates the effectiveness of the methodology in characterizing clinical trials and subjects and accurate matching between them.

The World Wide Web (WWW) has become a major platform for hosting, discovering, and composing web services. Web service clustering is a technique for efficiently facilitating web service discovery. Most web service clustering approaches are based on suitable semantic similarity distance measure and a threshold. Threshold selection is essentially difficult and often leads to unsatisfactory accuracy. In this paper we propose a taxonomic clustering algorithm for grouping functionally similar web services. We have tested the algorithm on both simulation based randomly generated test data and the standard OWL-S TC test data set. We have observed promising results both in terms of accuracy and performance.

Semantic web search involves retrieval of user-specific web artifacts by utilizing their semantic descriptions. A very specific web artifact that has evolved in recent times is web services. Web services can be semantically described in languages like OWL-S. However, such languages are limited with respect to their expressivity of context. They also lack a formal ontological framework where efficient web service retrieval can be conducted. In this paper we model a service request scenario within the web as an event-driven system. Services as well as user requests are modeled as events. We propose an ontological framework called Context-Aware Ontology Framework for Events and Services (CAOFES) where such event-driven web service retrieval can be efficiently executed by a novel reasoning technique.

Today's Web becomes a platform for services to be dynamically interconnected to produce a desired outcome. It is important to formalize the semantics of the contextual elements of web services. In this paper, we propose a novel technique called Semantic Genome Propagation Scheme (SGPS) for measuring similarity between semantic concepts. We show how SGPS is used to compute a multi-dimensional similarity between two services. We evaluate the SGPS similarity measurement in terms of the similarity performance and scalability.

Peer-to-peer (P2P) systems and Semantic Web are two novel technologies that face a lot of shortcomings if considered as isolated paradigms. We present an approach that utilizes ontologies to set up a peer profile containing all the data, necessary for peer-to-peer interoperability. Using this profile can help eliminate some major issues persistent in current P2P networks, such as security, resource aggregation, group management. We also consider applications of peer profiling for Semantic Web built on P2P networks, such as an improved semantic search for resources, not explicitly published on the Web, but available in a P2P system. We develop the ontology-based peer profile in RDF format and demonstrate its manifold benefits for peer communication and knowledge discovery in both P2P networks and Semantic Web.

Emerging peer-to-peer computing provides new possibilities but also challenges for distributed applications. Despite their significant potential, current peer-to-peer networks lack efficient knowledge discovery and management. This paper addresses this deficiency and proposes the Intelligent File Sharing framework, which provides an effective and flexible query for P2P file sharing. The IFS is based on powerful schema and flexible inference, as well as efficiently integrated and extensible retrieval algorithms. Experimental results have provided evidence of the high performance and scalability of the Intelligent File Sharing (IFS) system in peer-to-peer environments.

Dynamic and complex computation in the area of Geographic Information System (GIS) or Mobile Computing System involves huge amount of spatial objects such as points, boxes, polygons, etc and requires a scalable data structure and an efficient management tool for this information. In this paper, for a dynamic management of spatial objects, we construct a hierarchical dynamic data structure, called an IST/OPG hierarchy, which may overcome some limitations of existing Spatial Access Methods (SAMs). The hierarchy is constructed by combining three primary components: (1) Minimum Boundary Rectangle (MBR), which is the most widely used method among SAMs; (2) the population-based domain slicing, which is modified from the Grid File [14]; (3) extended optimal Integer Searching algorithm [4]. For dynamic management of spatial objects in the IST/OPG hierarchy, a number of primary and supplementary operations are introduced. This paper includes a comparative analysis of our approach with previous SAMs, such as R-Tree, R+-Tree and R*-Tree and QSF-Tree. The results of analysis show that our approach is better than other SAMs in construction and query time and space requirements. Specifically, for a given search domain with n objects, our query operations yield $O($ \scriptsize $\sqrt {\frac {\log n} {\log\log n}}$\normalsize $)$ compared to $O(\log n)$ of the fast SAM and an IST/OPG hierarchy containing $n$ objects can be constructed in $O(n$ \scriptsize $\sqrt {\frac {\log n}{\log\log n}}$\normalsize $)$ time and O(n) space.