Data fragmentation across IoT healthcare devices makes identifying and tracing causal health events challenging. In this work, we proposeCasFly, a decentralized, on-the-fly causal chain tracing protocol for healthcare event sequences. We design a specialized data structure,Temporal Probabilistic Health Graph (TPHG), to dynamically trace event dependencies across heterogeneous healthcare devices. When a device detects an anomalous parameter at the edge, it queries its local TPHG to identify the most probable temporal precursors. We propose incremental chain formation at each of the devices usingLaVE Algorithm, a Lag-aware Viterbi Expansion Algorithm. The algorithm transmits data only when new causal paths are identified, thus reducing the latency of the dependency path construction.CasFlyoperates device-to-device by expanding the chain until no further expansion is possible. Finally, the chain is returned to the initiating device for visualization and analysis. We observe that on emulated IoT hardware, TPHGs load in < 0.1 s and full chains develop in ≤5 s using ∼170MB RAM. TPHGs reduce storage by ∼97.6% versus Conditional Probability Tables. As compared to PCMCI+, LaVE improves macro AUPRC by ≈18% and reduces log-loss by ≈8%, with similar AUROC. These results indicate that on-the-fly chain construction supports real-time monitoring and clinician interpretation of probable cross-device event sequences.

This work proposes a marketplace model, EMsaaS , for sharing microservices on the IoT network on edge. Containerized microservices are well-suited for real-time IoT applications on edge because of their lightweight nature, independent lifecycle, and scalability. However, deploying microservices on demand is often failure-prone because of edge bandwidth limitations and resource constraints. Additionally, centralized image registries hosting containerized application images are often far away from the edge site and incur latency for the download of the images. They are also prone to a single point of failure. This work proposes a decentralized mechanism for sharing microservices in the edge network using a peer-to-peer approach. We introduce a component Edge Microservice Tracker and deploy it on the participating edge nodes. This component communicates with the other peer nodes and discovers microservices in the network. When requested, it fetches the requested microservice from the host node and deploys the same on the requesting edge node. We implement the proposed system in an IoT network of edge devices and evaluate the model using several parameters such as the deployment time, CPU and memory consumption, internet upload/download rate, and metadata synchronization time for sharing microservices on the edge. We observe that the system performs better than the cloud or centralized registry based approach with respect to the microservice deployment time with almost a 30% reduction in the deployment time even without any bandwidth limit for internet download from the edge node. It further surpasses the cloud registry-based microservice deployment procedure, with edge nodes having bandwidth limits. Moreover, EMsaaS proposes the concept of microservice sharing in an edge network, which enhances collaboration between edge service providers and fosters faster innovations.

In this work, we propose a smartphone-based system for detecting diabetic foot ulcers. Existing solutions mostly rely on clinical tests or costly imaging systems such as MRI or X-rays. These face challenges in resource-constrained environments and also do not provide immediate results. To address these limitations, we propose IgniSole, an edge-based offline solution for the detection of diabetic foot ulcers. We first train a fine-tuned MobileNetV2 Convolutional Neural Network (CNN) on the plantar thermogram dataset using transfer learning. We also develop an Android application to allow users to test their diabetic foot images using a smartphone. We deploy the trained model on the Android application to reduce dependency on the internet and to secure the user’s data. We evaluate the model’s performance in two phases: the training phase and the inference phase on the smartphone. We observe that our model achieves an accuracy of 96.24%, specificity of 100%, and sensitivity of 92.85% in the training phase. In contrast, the model’s accuracy, specificity, and sensitivity are 96.16%, 98.88%, and 97.96%, respectively, in the inference phase on the smartphone. Moreover, the model predicts the results within a span of 1-2 seconds. Therefore, our work provides a smartphone-based platform to users for the detection of diabetic foot ulcers with affordable, scalable, and accurate diagnostic tools. The source codes are available at https://github.com/anshita510/IgniSole.

Heterogeneous healthcare IoT edge devices operate in isolation. This limits their ability to make reliable collaborative decisions in real-time clinical settings. Existing systems lack lightweight inference-fusion mechanisms for decentralized coordination across sensors. In this work, we propose ZhiSync, a plugand-play, metadata-driven collaborative inference framework for heterogeneous healthcare IoT systems. In this system, we allow the devices to deploy their lightweight neural models for modality-specific prediction. We develop a framework that allows devices to exchange confidence, urgency, and timestamp metadata, known as ZhiTag, over asynchronous User Datagram Protocol (UDP). We deploy an inference algorithm, ZhiAware, which integrates peer ZhiTag with local input to adaptively refine predictions in real-time. We test our system on several parameters using a simulation framework with real datasets that includes three devices, namely, an Electrocardiogram (ECG) monitor, a motion sensor, and a breath analyzer, using embedded neural networks. ZhiSync increases local prediction confidence in 99.6% of decisions, with an average relative gain of 20.6% and stronger effects under high-urgency events and reduces false positives. CPU RAM and runtime is less than 2% as compared to baseline. Communication remains sub-kilobyte per second per node. Overall, lightweight metadata exchange shows that ZhiSync provides a reliable and scalable collaboration across decentralized healthcare IoT.

Editing model parameters directly in Transformers makes updating open-source transformer-based models possible without re-training (Meng et al., 2023).However, these editing methods have only been evaluated on statements about encyclopedic knowledge with a single correct answer.Commonsense knowledge with multiple correct answers, e.g., an apple can be green or red but not transparent, has not been studied but is as essential for enhancing transformers' reliability and usefulness.In this paper, we investigate whether commonsense judgments are causally associated with localized, editable parameters in Transformers, and we provide an affirmative answer.We find that directly applying the MEMIT editing algorithm results in sub-par performance, and propose to improve it for the commonsense domain by varying edit tokens and improving the layer selection strategy, i.e., MEMIT CSK .GPT-2 Large and XL models edited using MEMIT CSK outperform best-fine-tuned baselines by 10.97% and 10.73% F1 scores on PEP3k and 20Q datasets.In addition, we propose a novel evaluation dataset, PROBE SET, that contains unaffected and affected neighborhoods, affected paraphrases, and affected reasoning challenges.MEMIT CSK performs well across the metrics while fine-tuning baselines show significant trade-offs between unaffected and affected metrics.These results suggest a compelling future direction for incorporating feedback about common sense into Transformers through direct model editing. 1 * Co-first and last authors.Lorraine's work done at AI2. 1 Code and datasets for all experiments are available at https://github.com/anshitag/memit_csk

Stress has become an increasing concern among college students. Passive sensing techniques allow the extraction of stress-related parameters from a student. These techniques use highly resource-intensive machine learning algorithms to predict the stress levels of a student from these parameters. However, the current techniques do not provide any social communication solution for students suffering from stress. In this work, we propose StressAlly, a stress companion recommender system. The system comprises two modules, the stress score predictor and the stress companion recommender. The stress score predictor incorporates edge computing and deploys lightweight in-app inferences. The stress score predictor calculates the stress level in the scale 0–4 in the smartphone using the Artificial Neural Network (ANN) Regressor and sends it to the server. The Stress companion recommender provides similar stress levels of students to each student using the User-User Collaborative Filtering technique. We achieved training MAE and loss of 0.7478 and 1.0298, respectively. We get a test Mean Absolute Error (MAE) of 0.737 on the unseen data. We evaluate the CPU, memory, time delay, and network performance of StressAlly on the server and the Android smartphone. StressAlly utilizes 188 MB memory and 25% CPU on the smartphone.

In social IoMT systems, resource-constrained devices face the challenges of limited computation, bandwidth, and privacy in the deployment of deep learning models. Federated learning (FL) is one of the solutions to user privacy and provides distributed training among several local devices. In addition, it reduces the computation and bandwidth of transferring videos to the central server in camera-based IoMT devices. In this work, we design an edge-based federated framework for such devices. In contrast to traditional methods that drop the resource-constrained stragglers in a federated round, our system provides a methodology to incorporate them. We propose a new phase in the FL algorithm, known as split learning. The stragglers train collaboratively with the nearest edge node using split learning. We test the implementation using heterogeneous computing devices that extract vital signs from videos. The results show a reduction of 3.6 h in the training time of videos using the split learning phase with respect to the traditional approach. We also evaluate the performance of the devices and system with key parameters, CPU utilization, memory consumption, and data rate. Furthermore, we achieve 87.29% and 60.26% test accuracy at the nonstragglers and stragglers, respectively, with a global accuracy of 90.32% at the server. Therefore, FedCare provides a straggler-resistant federated method for a heterogeneous system for social IoMT devices.

The social IoMT-based activity-monitoring system comprises several devices with different datasets. It faces challenges like a collection of a global activity dataset which comprises a myriad of activities. In this article, we propose a federated Siamese network-based data-independent group activity segregator—Skipper—which aims to identify anomalies in an activity-monitoring social IoMT system. The novelty of this work is that Skipper does not require any dataset before its deployment, which removes the need for any prior training of the model for activity monitoring. As a proof of concept, we select activities pertaining to school environments to identify low-performing students in a classroom, who would require teachers’ close attention to ensure balanced growth and proper health. Skipper monitors the students independently for their motion signatures through a wearable device that consists of an accelerometer. A federated Siamese network calculates indices that signify the degree of similarity among the students’ activities. Skipper identifies the students who do not perform the same activity. With real-world implementations, we observe that Skipper requires network rates of 10 Kb/s, making it suitable for low bandwidth networks while we achieve just 20% CPU and 10 MB memory utilization on constrained edge devices. Further, with an increasing number of students up to 100, the time delay for final results is limited to 80 s. Hence, Skipper is a fast, easy, and accurate solution for recognizing outliers in IoMT social systems.

In this work, we propose an environment-aware electrical safety assistant using smartphones in pervasive domains like industry, homes, and healthcare. Conventional methods involve using eye shields, gloves, finger guards, and safety toe shoes. We depend on IoT-based solutions and propose Magdroid, an autonomous and standalone smartphone application that detects any electrical anomaly around the user and alerts all the users in the network. Magdroid supports edge computing and provides in-app inferences without any dependency on remote servers. It extracts the in-built magnetometer readings to detect any electrical anomaly in the environment. Since the readings vary with the environments, Magdroid first senses and then uses a cascaded deep learning technique to predict the electrical anomaly around the smartphone. We use two Convolution Neural Network (CNN) architectures and cascade the inference of one with the input of another to generate efficient results for detecting electrical anomalies that are particular to that environment. The first model achieves a test accuracy of 98.97% for the prediction of the environment and the cascaded CNN achieves a test accuracy of 81.88% with 7.82% and 37.27% loss, respectively. Additionally, Magdroid is a low resource-consuming application that utilizes 8% CPU and 126.MB memory of the smartphone.