In this paper, we present an ESD characterization and sensitivity analysis methodology using a Technology Computer-Aided Design (TCAD)-based gate-grounded NMOS (GGNMOS) digital twin. First, the virtual GGNMOS device is created by simulating the fabrication process of a $0.35 \mu \mathrm{~m}$ CMOS technology. The digital twin is then calibrated to a known device by matching the transmission line pulse (TLP) IV curves. Second, a one-at-a-time (OAT) sensitivity analysis is performed for various device/process parameters. In particular, the impact of device/process parameters on the trigger point, breakdown point, and on-resistance was quantified. Statistical box plots are used to interpret the results of the sensitivity analysis. Finally, the OAT analysis is leveraged to generate a targeted labeled dataset that includes all the failure modes with respect to the ESD design window. This proposed platform opens the door to applying machine learning (ML) to the design and refinement of the ESD protection circuit/window while accounting for process variation.

We propose a video-based monitoring platform that combines computer vision and AI techniques to provide a comprehensive patient behavioral analysis for clinical monitoring and risk assessment. To achieve this, a dual-camera platform is implemented to identify, track, and analyze interactions between patients and clinically relevant objects. Our methodology includes a CNN-based person/object detection & tracking, posture classification, ambulation assessment, depth estimation, and interaction analysis. A customized clinically relevant image dataset is constructed for training and validation. Our platform has achieved promising performance, i.e., an average F1 score of 98% for posture classification, and an average accuracy of 93% for patient-object interaction. The proposed platform is scalable across the number of cameras, patients, and objects of interest. It produces excellent results in spite of patient occlusion and non-ideal lighting conditions and generates informative summary reports for clinical decision-making.

Images captured in low-light environments suffer from reduced visibility, low contrast, and noise, which degrade visual quality and hinder reliability in automated object detection. This paper presents a deep learning illuminationaware object detection framework that integrates a lightweight encoder-decoder enhancer with an object detection architecture to improve performance under poor lighting conditions. A threestep training strategy further improves the model's robustness and transferability. The proposed framework provides a scalable, real-time, and resource-efficient approach to illuminationaware object detection in low-light environments. The proposed architecture achieves 69.1% mAP on ExDark dataset with approximately 50% fewer parameters, demonstrating that an integrated architecture with joint training leads to an efficient solution for low-light visual monitoring.

Epilepsy patients with drug-resistant seizures emanating from two or more distinct regions of left and right hemispheres are the primary candidates for neurostimulation treatment. Stereo-electroencephalography (SEEG) is a minimally invasive technique to monitor and evaluate brain activities during seizures before stimulator implantation. This work proposes a seizure network modeling method using SEEG to analyze the functional connectivity of epileptogenic zone during bilateral seizures. Network nodes are selected subset of SEEG contact points, and network edges are directed signal correlations calculated from directed transfer function. Based on signal directionality, four connectivity values are extracted to measure the intra- and inter-activities that are within or between the left and right hemispheres, respectively. Statistical difference between connectivity values is used to quantify the seizure impact of each hemisphere. A subset of network nodes is selected from impactful side as stimulation target candidates. Experimental results are validated on ten patients having different seizure types with bilateral onset. Each seizure type has specific connectivity patterns that show different importance from each brain side. Selection of neurostimulation targets from primary side are consistent with clinicians’ decision. Relationships are found among connectivity differences, seizure types and stimulation outcomes. Using SEEG signals, we can capture specific connectivity differences associated with bilateral seizure networks. Such differences are related with corresponding neurostimulation targets and stimulating outcomes. The proposed work elucidates the difference of network connectivity for bilateral patients, and assists clinicians to choose the stimulation targets and to predict the potential outcomes.

We propose a performance analysis and ranking methodology for track and field athletes, based on an AI model trained using historical data collected from actual races over several years. Track coaches will then have a data-driven tool to (i) quantify their athletes’ performances and progress during the training period and (ii) predict their level of consistency and competitiveness for upcoming competitions. To achieve this, we have utilized the public domain data of races recorded in TFRRS database, curated data, extracted key statistical features, and trained a neural network clustering (ART2) technique. The records of an athlete (e.g. during the training phase) can then be mapped to a few clusters to rank and explain their overall performance.

We propose a characterization method for evaluating the ESD behavior of multi-finger Grounded Gate NMOS or ggNMOS transistors using TLP stress data (V,I,I-leak). By employing a hierarchical clustering with Mahalanobis distance, the approach addresses challenges in analyzing data with noise and variability using real-world techniques, such as conventional Transmission Line Pulsing (TLP) testing platform. The proposed method enhances data integrity by identifying and excluding the unexpected behavior of samples caused by measurement or setup issues while providing statistical analyses the characterization of the devices that demonstrate normal behavior.

In this paper, we propose a machine learning-powered approach to explore analysis of electrostatic discharge (ESD) protection device measurements taken with high-current pulse method. Specifically, the data analyzed was collected under a very fast transmission line pulse (VF - TLP) on a grounded gate NMOS (ggNMOS) protection device. The analysis is done by applying K-means clustering in two complementary steps. First, we use the raw data for a holistic pattern-based clustering to filter out the possible measurement errors. In the second step, the subset that seems error-free is further analyzed through a feature-based clustering that provides insight into parametric variations and stability of the protection design. We also employ visualization techniques, i.e. t-SNE dimensionality reduction and statistical boxplots, to make interpretation of results easier for ESD test and quality engineers. We have validated the proof of concept using measurement data obtained in Kelvin setup for 18 devices.

We propose a novel approach of AI-driven analysis of data from both TLP and VF-TLP of identical ESD devices. Hierarchical clustering with Mahalanobis distance is shown to deal with challenges such as noise and variability. The proposed approach detects anomalies, outliers, measurement artifacts and performance trends while identifying inconsistent samples.

Video-based patient behavior analysis offers deep insight into patient activities, adherence, risk assessment, and clinical decision-making. By integrating computer vision and machine learning, healthcare providers can monitor and analyze patient behaviors, improving quality of care, recovery progress, and treatment. In this paper, we propose a method to systemati- cally monitor patient behavior by analyzing interactions between a patient (in various postures) and clinically significant objects. Our methodology includes a CNN-based person/object detection & tracking, posture classification, ambulation assessment, depth estimation, and interaction analysis. A customized clinically relevant image dataset is constructed for training and validation. Our platform has achieved promising performance, i.e. an average F1 score of 98% for posture classification and an average accuracy of 90% for interaction detection.

Stereo-electroencephalography (SEEG) is a technique to monitor and evaluate the spatial and temporal properties of ictal EEG changes in patients with epilepsy during pre-surgery evaluation. When patients have different types of seizures, the intra-seizure patterns across different ictal episode provide extra and meaningful information for personalized treatment planning. This work introduces a patient-specific framework to capture the intra-seizure patterns in a seizure-specific way. After defining a Pre-seizure Plus Seizure (PPS) window as period of interest, SHapley Additive exPlanations (SHAP) is applied to quantify the contributing score of each SEEG channel (i.e. spatial correlation) based on XGBoost classifier. These SHAP scores are segmented and compared via Soft-Dynamic Time Warping (Soft-DTW) to characterize their dissimilarities (i.e. temporal pattern). Then, k-medoids clustering is exploited to divide seizure episodes into groups (episodes) based on Soft-DTW variations, and three clinically meaningful stages, namely trigger, transient, and steady stages, are consistently identified. Validated on SEEG data from eight patients, our results demonstrate high classification performance, reliable epileptogenic zone localization, and robust intra-seizure stage segmentation which are consistent with clinicians annotations.

We propose a vision-language framework designed for clinical environments to monitor patient behavior and generate clinically informative reports using advanced computer vision and large language models (LLMs). Our framework analyzes camera images, identifies patient's postures, activities, and interactions with objects, and provides insights into a pretrained summarization LLM to produce context-aware reports. This is done by leveraging a custom-trained Convolutional Neural Network (CNN). This vision-language approach enables real-time or offline monitoring and report generation without the need for wearable devices, offering a non-intrusive, automatic, and scalable monitoring for healthcare settings.

Stereo-electroencephalography (SEEG) is an invasive technique to implant depth electrodes and collect data for pre-surgery evaluation. Visual inspection of signals recorded from hundreds of channels is time consuming and inefficient. We propose a machine learning approach to rank the impactful channels by incorporating clinician’s selection and computational finding. A classification model using XGBoost is trained to learn the discriminative features of each channel during ictal periods. Then, the SHapley Additive exPlanations (SHAP) scoring is utilized to rank SEEG channels based on their contribution to seizures. A channel extension strategy is also incorporated to expand the search space and identify suspicious epileptogenic zones beyond those selected by clinicians. For validation, SEEG data for five patients were analyzed showing promising in terms of accuracy, consistency and explainability.

We propose an ambulation assessment platform in healthcare that employs a deep neural network for body/head detection and further recognizes the position, postures, and motion of a person in a video stream. To achieve this, we: (i) find the head’s 3D coordinates, (ii) measure distance from the camera, (iii) track body movement, and (iv) detect postures. Based on testing using human volunteers, our method has achieved promising results with accuracy up to 90% for body and head detection, 92% for movement distance measurement, and 99% for posture classification.

This paper explores the critical task of analyzing running postures and recognizing their key roles in enhancing athletic performance and averting injuries. Our approach integrates AI technologies, with video pose estimation and kinematic analysis, to offer comprehensive and detailed feedback on a runner's posture and technique. Leveraging MediaPipe, an advanced deep learning framework, we conduct thorough kinematic analyses by tracking essential body landmarks and analyzing four fundamental postures: toe-off, maximal vertical projection, touch-down, and full-support. The classification of these postures is facilitated by a random forest tree classifier, which consistently demonstrates high accuracy, and is validated by experienced coaches. Our system goes beyond simple classification; it provides detailed feedback, highlights distinctions in form and technique, and assists runners and coaches in refilling performance. This methodology not only optimizes athletic prowess but also analyzes real-world track running data, as opposed to unnatural data collected by laboratory-based systems.

Responsive neurostimulation (RNS) is an effective device for patients with multifocal seizures whose ictal foci are independent of each other across left and right hemispheres. Accurate RNS placement is crucial to enhance seizure suppression outcomes. Stereo-electroencephalography (SEEG) is employed before RNS placement to collect deep brain activities and determine stimulation targets. To deal with different seizure types and semiologies for multifocal patients, this paper presents a functional seizure network model using SEEG to identify potential RNS targets. The network nodes are a subset of SEEG contact points, and directional weighted network edges are SEEG correlations quantified by directed transfer function (DTF). The network nodes are then ranked by their strength values, and the top-4 nodes are selected as RNS targets with respect to different seizure types in each hemisphere. The proposed methodology is validated based on five multifocal patients. Consistent results between our computational findings and clinicians' decisions are observed with 95.7% overlapping ratio.

Interictal epileptiform discharges (IEDs) are electrophysiological events that intermittently occur in between seizures in Epilepsy patients. Automated detection of IEDs is crucial for assisting clinicians in epilepsy diagnosis as they can help identify the extent of cortical irritations and may indicate an upcoming seizure, thus helping in preventing seizure. It also minimizes visual inspection of very long EEG signals by physicians. This paper presents a transfer-learning-based approach for analyzing time-frequency representations of different types of IEDs from scalp EEG data using a fine-tuned deep residual network. The proposed method was evaluated using the publicly available Temple University Events EEG dataset. Experimental results show that our method demonstrates promising performance, by achieving an F1-score of 88.52% on this dataset for binary classification of IEDs.

Electromyogram (EMG) signal is considered as an easy-to-capture (i.e. skin-mounted) and promissing biometric for the control of prosthetic hands. Despite the plethora number of researches on EMG-based control of prosthetics, two major challenges have not sufficiently been addressed , i.e. realtime classification, and robustness against noise. Our contribution in this paper is two fold. First, we have proposed a deep neural network (DNN) model with customized architecture to learn features directly from raw EMG data. In this model, a self-improvement module enhances the accuracy and robustness against artifact. Second, we utilize extreme value machine (EVM) for classification of the learnt feature. EVM models the statistical distribution of the latent space and classifies finger movements based on the maximum cumulative distribution probability to the fitted model. Our experimental results, using public domain EMG data, are very promising. They indicate the average accuracy of 98.5% for 750 msec window for 10-class classification. This result is superior or competitive compared with other online classifications reported in the literature on the same dataset. Additionally, our results illustrate that self-improvement feature learner (SIFL) is much more resistant against the noise than the models that employ engineered features. For example, in presence of severe white noise, the accuracy of our methodology degrades 10−15% compared to 30−50% of others.

Stereo-electroencephalography is a minimally invasive technique for patients with refractory epilepsy pursuing surgery to reduce or control seizures. Electrodes are implanted based on pre-surgery evaluations and can collect deep brain activities for surgery decisions. This paper presents a methodology to analyze stereo-electroencephalography and assist clinicians by recommending the optimal surgical option and target areas for focal epilepsy patients. A seizure network (graph) model is proposed to characterize the spatial distribution and temporal changes of ictal events. The network nodes and edges correspond to specific epileptogenic regions and propagation/impact pathways (weighted by directed transfer function), respectively. We then employ a K-means clustering strategy to group nodes into a few clusters, from which the target surgical areas can be identified. Ten patients with different types of focal seizures were thoroughly analyzed. Promising consistency between results of our method's recommendations, clinical decisions and surgery outcomes were observed.

This paper presents a pioneering examination of using Machine Learning (ML) for applications to ESD data analysis. It involves demonstrating how post ESD stress IV curve data can lead to machine learning opportunities. The proposed technique can significantly reduce the iterations between product test and ESD design engineers and thus shortening time to market.

Patients with drug-resistant focal seizures are usually recommended with surgery for seizure freedom or reduction. The traditional open resection, which is to surgically remove epileptogenic zones, requires open lobectomy that are risky to cause post-surgery complications. Minimal invasive surgery (MIS) approaches, including laser interstitial thermal therapy (LITT) and responsive neurostimulation (RNS), are effective to significantly reduce (or achieve zero) seizures without risk of open resection. Stereo-electroencephalography (SEEG) is a promising technique for pre-surgery evaluation by implanting depth electrodes to collect brain activities. This paper presents a SEEG analyzing method to determine the LITT/RNS targets for patients with focal seizures. A seizure network (graph) model is proposed to characterize the spatial distribution and temporal changes of ictal events. Two network metrics, the outflux and influx values, are defined to characterize the nodal contributions to seizures. A clustering strategy is then proposed to group nodes (a subset of selected SEEG contact points) into several clusters based on outflux/influx values. Retrospective data from six patients with both unilateral and bilateral focal seizures were analyzed. Promising consistency were observed between results of our computational method and clinical decisions.