Author: Panithan Boonkert

Research Committee:

• Edward Colgate, Mechanical Engineering
• Michael Peshkin, Mechanical Engineering
• David Zaretsky, Electrical and Computer Engineering

Abstract:
An electro-adhesive clutch–based variable damper for wearable tremor suppression is investigated using high-frequency switching control. The study combines experimental evaluation of two clutch variants with analytical modeling of switching behavior, demonstrating that duty cycle enables tunable damping while identifying engagement speed and force capacity as key design considerations.

A Switched Electro-Adhesive Clutch as a Variable Damper: Proof of Concept and Initial Characterization for Wearable Tremor Suppression

Author: Xinyu Cao

Research Committee:

• Qi Zhu, Electrical and Computer Engineering
• David Zaretsky, Electrical and Computer Engineering

Abstract:
A constrained reinforcement learning framework for safe robotic manipulation is developed using runtime Linear Temporal Logic monitoring within a constrained Markov decision process framework. The approach evaluates PPO-Lagrangian, FOCOPS, and CPO algorithms in the OmniGibson/Isaac-Sim physics simulator, achieving 73% success under deployment conditions and identifying strategies for stable safety-aware policy learning.

Constrained Reinforcement Learning for Safe Contact-rich Manipulation under Temporal-logic Specifications

Author: Qixin Deng

Research Committee:

• Thrasyvoulos Pappas, Electrical and Computer Engineering
• Bryan Pardo, Electrical and Computer Engineering

Abstract:
A study of joint language-audio embedding models is conducted to evaluate their ability to represent perceptual timbre semantics. The approach compares audio-text similarity predictions with human timbre ratings and descriptor-based audio effect manipulations to assess alignment between learned embeddings and auditory perception. The results show that current models capture timbre characteristics only partially, motivating the development of more perceptually grounded audio-language representations.

Evaluating The Encoding of Perceptual Timbre Semantics of Joint Language-Audio Embedding Models

Author: Daniel Donnelly

Research Committee:

• David Zaretsky, Electrical and Computer Engineering
• Selim Shahriar, Electrical and Computer Engineering

Abstract:
A quantum cascade laser simulator is developed through theoretical modeling, design, and implementation to enable analysis of QCL architectures. The simulator is validated by modeling an existing quantum cascade laser design and comparing simulation results with experimental measurements.

Quantum Cascade Laser Simulator

Author: Ruofan Huang

Research Committee:

• Hai Zhou, Electrical and Computer Engineering
• Jie Gu, Electrical and Computer Engineering

Abstract:
A graph neural network framework, PlacementGNN, is developed to predict per-cell wirelength gradients for VLSI global placement by approximating analytical gradients used in the OpenROAD RePlAce engine. The approach leverages a star-expanded hypergraph representation and scale-invariant features to support transfer across designs of varying sizes, with evaluation demonstrating its effectiveness for quality-preserving macro legalization in academic and commercial place-and-route flows.

GNN-Based Wirelength Gradient Prediction for VLSI Global Placement

Author: Tonghe Liu

Research Committee:

• Ying Wu, Electrical and Computer Engineering
• David Zaretsky, Electrical and Computer Engineering

Abstract:
A vision-based pipeline for automated MBSImP assessment is developed to enable objective analysis of videofluoroscopic swallowing study (VFSS) videos. The framework consists of anatomical landmark localization followed by component-level scoring based on detected landmarks, providing an automated approach for evaluating swallowing function.

Vision-Based Analytics of Dysphagia in Videofluoroscopic Swallowing Studies

Author: Xiaoyi Lu

Research Committee:

• R. James Cotton, Feinberg School of Medicine, Physical Medicine and Rehabilitation
• David Zaretsky, Electrical and Computer Engineering

Abstract:
A framework for image-based biomechanical human pose prediction is developed by bridging SMPL-X mesh annotations and a MuJoCo-based biomechanical skeleton through a shared virtual keypoint representation. The approach adapts a MultiHMR-style architecture with a DINOv3 backbone and uses differentiable forward kinematics to estimate biomechanical joint angles, body scale, root orientation, and depth without requiring direct biomechanical pose labels. Experiments on BEDLAM2 demonstrate improved feasibility for biomechanical pose estimation from monocular images.

Learning Biomechanical 3D Pose Representations from Surface Mesh Annotations

Author: Liming Luo

Research Committee:

• Jiancheng Ye, Cornell University (Original Northwestern Faculty)
• David Zaretsky, Electrical and Computer Engineering

Abstract:
A text-guided medical image segmentation framework, DescriptorMedSAM, is developed to enable abdominal CT organ segmentation using radiology language descriptions instead of manual prompts. The approach extends MedSAM with a CLIP text encoder and cross-attention module, demonstrating that richer anatomical descriptions improve generalization to unseen organs and provide a reproducible platform for evaluating SAM-based segmentation methods.

DescriptorMedSAM: Language–Image Fusion for Click-Free Medical Image Segmentation and a Reproducible Engineering Platform for Its Deployment

Author: Bhavana Panchumarthi

Research Committee:

• Prem Kumar, Electrical and Computer Engineering
• Gregory Scott Kanter, Electrical and Computer Engineering
• Hooman Mohseni, Electrical and Computer Engineering

Abstract:
An on-chip asymmetric Mach-Zehnder interferometer approach for extracting optical waveguide parameters is investigated to improve photonic chip characterization. Monte Carlo simulations are used to quantify error thresholds from deviations in single-mode coupling, providing insights for more accurate measurement of propagation loss and coupling ratios in diagnostic photonic devices.

Toward On-Chip Squeezing: Design and Characterization of TFLN Photonic Devices

Author: Aleksa Pjanic

Research Committee:

• Selim Shahriar, Electrical and Computer Engineering
• Seng-Tiong Ho, Electrical and Computer Engineering
• Mahdi Hosseini, Electrical and Computer Engineering
• Timothy Light Kovachy, Physics

Abstract:
A low-cost, high-resolution spectroscopy platform spanning ultraviolet to mid-infrared wavelengths is developed through innovative spectrometer architectures and precision mechanical designs. The work introduces a highly stable echelle spectrometer using low thermal expansion polymers and carbon fiber structures, along with a novel FTIR system featuring a diamagnetically levitated delay line to overcome mid-infrared dispersion limitations. These approaches enable improved stability, resolution, and accessibility for advanced spectroscopy applications.

Low CTE UV-VIS-IR Spectrometers 0.2–25μm: Design, Fabrication, and Frictionless Delay Lines

Author: Bowen Qi

Research Committee:

• Selim Shahriar, Electrical and Computer Engineering
• Hooman Mohseni, Electrical and Computer Engineering

Abstract:
A hybrid photonic and atomtronic computing approach for accelerating convolutional neural networks is investigated through physical correlator architectures. The study maps trained CNN convolutional kernels onto two-dimensional photonic and three-dimensional opto-atomic spatio-temporal holographic correlators for image and video classification tasks. The results demonstrate the potential of optical and atomic systems to efficiently perform convolution operations while integrating electronic processing for nonlinear activation and final classification.

Photonic and Atomtronic Correlators for Hardware Acceleration of Two- and Three-Dimensional Convolutional Neural Networks

Author: Yang Qian

Research Committee:

• Stephen Xia, Electrical and Computer Engineering
• David Zaretsky, Electrical and Computer Engineering

Abstract:
A stochastic modeling framework for ultimate frisbee strategy analysis is developed to examine how team strength and risk-taking influence optimal gameplay decisions. The model demonstrates that weaker teams benefit from aggressive, high-variance strategies, while stronger teams achieve better outcomes through conservative possession-based play. The work also introduces a computer vision pipeline using YOLO and Swin Transformer models for automated match analysis and scalable ultimate frisbee analytics.

DISC POSSESSION, FIELD POSITION, AND STRATEGY: A SIMULATION AND VIDEO ANALYSIS FRAMEWORK IN ULTIMATE FRISBEE

Author: Brendan Tan-Fahed

Research Committee:

• Daniel Kim, Feinberg School of Medicine
• Aggelos Katsaggelos, Electrical and Computer Engineering

Abstract:
A deep learning-based respiratory motion correction method for myocardial blood flow quantification in cardiac magnetic resonance imaging is developed to enable faster and accurate pixel-wise analysis. The approach addresses limitations of traditional optical flow–based correction methods by reducing processing time while maintaining comparable accuracy. The proposed solution supports more efficient inline implementation of quantitative CMR for coronary artery disease assessment.

Rapid Supervised Multi-Stage Deep Learning Framework for Respiratory Motion Correction in First-Pass Perfusion Cardiovascular Magnetic Resonance Imaging

Author: Chi Wan

Research Committee:

• Manling Li, Computer Science
• David Zaretsky, Electrical and Computer Engineering

Abstract:
A unified state-action prediction model and test-time world modeling approach for Vision Language Action models are developed to improve rapid adaptation in new environments. The framework enhances model flexibility by enabling efficient learning and adjustment during deployment, supporting more robust performance across diverse operating conditions.

WorldAgen: Unified State-Action Prediction with Test-Time World Model Training

Author: Zinan Wang

Research Committee:

• Qi Zhu, Electrical and Computer Engineering
• David Zaretsky, Electrical and Computer Engineering

Abstract:
A multi-level formal safety evaluation framework, SENTINEL, is developed for large language model–based embodied agents. The framework introduces modular evaluation infrastructure with funnel-style error filtering, plan-level Linear Temporal Logic safety verification, and scalable batch processing to assess agent behavior across multiple models and safety configurations.

Safety Evaluation System for LLM-based Embodied Agents

Author: Hongjun Xu

Research Committee:

• Stephen Xia, Electrical and Computer Engineering
• David Zaretsky, Electrical and Computer Engineering

Abstract:
A runtime-free deployment pipeline for executing Mamba neural network models on resource-constrained microcontrollers is developed to enable efficient embedded inference. The approach converts trained PyTorch models into optimized standalone C implementations using operator fusion and memory optimization, reducing memory requirements while maintaining numerical fidelity and classification accuracy on ESP32S3 and STM32H7 platforms.

A Runtime-Free Deployment Pipeline for Mamba on Microcontrollers

Author: Jincheng Yu

Research Committee:

• Stephen Xia, Electrical and Computer Engineering
• David Zaretsky, Electrical and Computer Engineering

Abstract:
A wearable magnetic haptic platform for augmented interaction is developed, integrating actuator design, wireless power operation, host-side visual perception, and experimental evaluation of user perception and task performance. The work also investigates an ear-level active noise cancellation prototype and analyzes architecture-level latency characteristics. The results advance wearable systems for immersive interaction and low-latency human-computer interfaces.

WIRELESS FINGERTIP MAGNETIC HAPTICS FOR AUGMENTED TABLETOP INTERACTION

Author: Xiaohan Yu

Research Committee:

• Manijeh Razeghi, Electrical and Computer Engineering
• David Zaretsky, Electrical and Computer Engineering

Abstract:
A study of structural parameters in InGaAs/InAlAs/InP ridge waveguide quantum cascade lasers is conducted to understand their impact on device performance. The research investigates the effects of cavity length and mirror loss through combined experimental and theoretical analysis, providing insights into the relationship between QCL design and performance characteristics.

Analysis and Optimization of InGaAs/InAlAs/InP Quantum Cascade Lasers

Author: Sinong Zhan

Research Committee:

• Qi Zhu, Electrical and Computer Engineering
• Stephen Xia, Electrical and Computer Engineering

Abstract:
A delay-robust offline reinforcement learning framework, DT-CORL, is developed to learn effective policies from static, delay-free datasets under delayed observation conditions. The approach combines a transformer-based belief model to infer latent states from delayed observations with a constrained policy optimization objective for robust decision-making. The framework enables improved policy performance in environments affected by observation delays.

Delayed Policy Optimization for Offline RL Setting

Author: Yulong Dai

Research Committee:

• Aggelos Katsaggelos, Electrical and Computer Engineering

Abstract:
A deep learning framework for classifying seven crystal systems from scanning transmission electron microscopy images is developed to address challenges from variable material orientations. Using a diverse dataset of 6,670 simulated HAADF-STEM images, the approach introduces Fractal Average Pooling to capture structural complexity through fractal-based feature representations. The method achieves high classification accuracy across multiple datasets, demonstrating the value of fractal features for reliable materials image analysis.

STEM Image Classification Using Fractal-Based Neural Network

Author: Hao Dang

Research Committee:

• Lee Cooper, Electrical and Computer Engineering
• David Zaretsky, Electrical and Computer Engineering

Abstract:
A deep learning framework for predicting prostate cancer subtypes (PAM50 and PSC) directly from H&E-stained whole-slide images is developed using a patch-wise attention-based multiple instance learning model with a MobileNetV2 backbone and a three-branch attention mechanism. Trained on 641 whole-slide images from 323 patients, the approach demonstrates promising performance, particularly for luminal subtype identification, and visualization analyses reveal subtype-specific tissue patterns. The predicted subtypes further correlate with treatment response, underscoring the potential of this noninvasive and cost-effective method to complement genomic testing in personalized prostate cancer care.

Deep Learning-Based Framework for Predicting Prostate Cancer Molecular Subtypes

Author: Guoting Ding

Research Committee:

• Randy Freeman, Electrical and Computer Engineering
• David Zaretsky, Electrical and Computer Engineering

Abstract:
A diffusion policy framework for two-dog shepherding is developed by extending a single-dog model to coordinate multi-agent herd guidance. Experimental evaluation investigates the effects of flock distributions and herd sizes on performance, demonstrating their influence on shepherding success rates and task completion times.

Diffusion Policy for Shepherding: Extending from One To Two Shepherds

Author: Yuxin Du

Research Committee:

• Nabil Alshurafa, Feinberg School of Medicine, Preventive Medicine

Abstract:
A multimodal wearable sensing system for privacy-conscious detection of eating behavior is developed to support dietary monitoring and health interventions. The approach employs an upward-facing thermal infrared camera to identify hand-to-mouth gestures and activates a forward-facing RGB camera only when needed, combining compact hardware with efficient on-device machine learning. The system achieves 98% gesture recognition accuracy while maintaining energy efficiency and scalability.

A Wearable Platform for Privacy-Aware Eating Detection via Multimodal Sensing

Author: Weihe Gao

Research Committee:

• Qi Zhu, Electrical and Computer Engineering

Abstract:
A deep reinforcement learning approach for multi-agent autonomous exploration of Martian caves is developed and compared with the state-of-the-art UDZRSR algorithmic strategy. Employing custom reward functions and phased exploration policies, the method demonstrates superior performance in larger cave environments through greater adaptability, efficient energy utilization, and dynamic role allocation. The findings indicate that integrating algorithmic precision with machine learning flexibility provides a particularly effective framework for planetary and analogous terrestrial exploration tasks.

Deep Reinforcement Learning Versus Algorithmic Approaches in Multi-Agent Mars Cave Exploration

Author: Robert Hayek

Research Committee:

• Igor Kadota, Electrical and Computer Engineering

Abstract:
Federated learning over a 5G-NR Standalone testbed is investigated using resource-constrained IoT devices and a central server based on software-defined radio and O-RAN technologies. Implemented with the Flower framework and a custom instrumentation tool, the study compares performance across 5G, Wi-Fi, and Ethernet networks, demonstrating that 5G uplink latency substantially slows convergence—by factors of 33.3 relative to Ethernet and 17.8 relative to Wi-Fi—and exacerbates the straggler effect during training.

Federated Learning over 5G, WiFi, and Ethernet: Measurements and Evaluation

Author: Mengnan He

Research Committee:

• Seda Ogrenci, Electrical and Computer Engineering

Abstract:
A hardware-efficient method for deploying quantized neural networks on resource-constrained FPGAs is developed by replacing sigmoid activations with piecewise-linear approximations on the Zynq XC7Z020 platform. The approach reduces memory requirements and enables shared computational resources between dense and activation layers, achieving timing closure while maintaining competitive accuracy despite scheduling and timing tradeoffs. The resulting design is well suited for real-time edge applications.

A Linearization-Based Approach to Cross-Layer Resource Sharing in Neural Network FPGA Deployment

Author: Hanbang Hua

Research Committee:

• John Rogers, Materials Science & Engineering

Abstract:
A compact wireless near-infrared spectroscopy system for continuous tissue oxygen saturation monitoring in pediatric patients is developed to overcome limitations of traditional bulky, wired devices. The platform integrates low-power components, dual-wavelength LED illumination, and optimized hardware to achieve accurate, long-term monitoring with extended battery life. Bench-top and clinical evaluations demonstrate its reliability and potential for noninvasive pediatric monitoring in hospital and home environments.

A compact, wireless Low-power functional Near Infrared Spectroscopy Device for continuous StO2 monitoring in Pediatric Application

Author: Pengxiang Huang

Research Committee:

• David Zaretsky, Electrical and Computer Engineering

Abstract:
APILOT, a tool for mitigating the use of outdated and vulnerable APIs in large language model–generated code, is developed by maintaining a real-time repository of deprecated APIs and steering code generation away from them. Evaluation across multiple language models demonstrates an average 89.42% reduction in outdated API usage with minimal impact on overall performance.

APILOT: Securing LLM-Generated Code by Avoiding Outdated API Usage

Author: Lijia Jiang

Research Committee:

• Stephen Xia, Electrical and Computer Engineering

Abstract:
A Kolmogorov-Arnold Network–based approach for speech bandwidth extension is investigated to reconstruct high-frequency components in narrowband speech and improve audio quality. The proposed architecture integrates KANs into a Transformer-augmented U-Net framework, combining global context modeling with learnable activation functions for spectral reconstruction. While demonstrating potential for capturing complex speech patterns, the approach highlights computational and training challenges that require further optimization.

Analysis for application of KAN for Bandwidth Extension

Author: Qinze Jiang

Research Committee:

• Peter Dinda, Computer Science

Abstract:
A custom RISC-V Rocket coprocessor, VIR-V, is developed to accelerate VCODE, an intermediate vector dataflow representation for expressing nested data-parallel operations. The architecture efficiently maps vector operations, including arithmetic, scans, reductions, and permutations, to hardware for parallel execution. Evaluated using FireSim, VIR-V achieves a 1.89× speedup over baseline CPU performance, demonstrating effective integration within the RISC-V ecosystem.

VIR-V: A RISC-V RoCC Accelerator for VCODE Computing

Author: Yaxing Jing

Research Committee:

• Jeffery Goldstein, Feinberg School of Medicine, Pathology

Abstract:
A modular content-based image retrieval framework for large-scale pathology image curation is developed to enable efficient and scalable dataset refinement. The approach combines triplet-loss-based feature extraction, weighted similarity fusion, and a closed-loop retrieval process to improve identification of relevant pathology samples across diverse datasets. Applied to a funisitis dataset, the system demonstrates that iterative retrieval with minimal manual input can enrich rare and clinically relevant samples for pathology research and discovery.

Closed-Loop CBIR for Scalable Pathology Image Curation: A Case Study on Umbilical Cord Funisitis

Author: Gen Li

Research Committee:

• Lirong Yan, Feinberg School of Medicine, Radiology

Abstract:
The relationship between periventricular and deep white matter hyperintensities and cognitive function in older adults is investigated using automated lesion segmentation and cerebrovascular measurements. Deep white matter hyperintensity burden is found to be more strongly associated with cognitive decline than periventricular burden, while both subtypes exhibit moderate negative correlations with regional cerebral blood flow, particularly in the frontal and insular regions. These findings indicate that increased lesion burden reflects broader cerebrovascular compromise and support the value of subregional white matter hyperintensity analysis for improving diagnostic precision in cognitive impairment.

Deep and Periventricular White Matter Hyperintensity Segmentation and Their Relationship with Cognitive Decline and Vessel Hemodynamics

Author: Haijie Li

Research Committee:

• Qi Zhu, Electrical and Computer Engineering

Abstract:
An adaptive communication strategy for multi-agent patrol systems under partial observability is evaluated using MAPPO with Graph Neural Networks. The study compares Full Communication, Bernoulli, and Smart Communication approaches, demonstrating that Smart Communication achieves the lowest average idleness with reduced communication overhead. The results highlight the importance of intelligent, adaptive communication mechanisms for efficient multi-agent coordination in real-world environments.

Selective Communication Strategies for Multi-Agent Patrolling Systems

Author: Qitong Li

Research Committee:

• Nivedita Arora, Electrical and Computer Engineering

Abstract:
A low-cost bioelectronic sensing system, PhytoBits, is developed to enable accessible monitoring and education of CAM photosynthesis. Using implanted electrodes and microcontrollers, the system captures bioelectrical signals associated with CAM acid accumulation cycles and integrates them into hands-on educational modules. The platform makes complex plant physiological processes observable and supports improved understanding of plant adaptation to water stress.

Phytobits: A Bioelectronic Sensor and Microcontroller-Based Curriculum for Cam Photosynthesis Education

Author: Xiangyu Li

Research Committee:

• David Zaretsky, Electrical and Computer Engineering

Abstract:
A multimodal framework for autonomous vehicle behavior classification is developed by integrating numerical features with large language model–based semantic reasoning. The LLM-MLFFN framework combines statistical and behavioral data with LLM-generated descriptions through a dual-channel attention network to improve prediction accuracy. Evaluated on the Waymo dataset, the approach achieves 94% accuracy, demonstrating the effectiveness of combining numerical and semantic information for robust autonomous vehicle behavior analysis.

Large Language Model-Enhanced Multi-Level Feature Fusion Network for Autonomous Driving Behavior Classification

Author: Yuxuan Li

Research Committee:

• Stephen Xia, Electrical and Computer Engineering

Abstract:
A Time-Weighted Contrastive Reward Learning (TW-CRL) framework for improving episodic reinforcement learning is developed to address sparse rewards and hidden failure states. The approach leverages successful and failed demonstrations with temporal context to learn dense reward functions that emphasize critical success and failure states. Experiments on navigation and robotic manipulation tasks demonstrate improved learning efficiency and robustness compared with existing inverse reinforcement learning methods.

Time-Weighted Contrastive Reward Learning for Efficient Inverse Reinforcement Learning

Author: Zhiyao Li

Research Committee:

• David Zaretsky, Electrical and Computer Engineering

Abstract:
A Multiscale Dilated Fusion Attention module integrated into a UNet framework is developed to improve cell segmentation under challenging conditions with complex morphologies, overlapping boundaries, and limited labeled data. The approach leverages dilated convolutions and multiscale feature fusion to capture diverse spatial information and enhance segmentation performance. Results demonstrate improved accuracy, particularly in low-data and limited-training scenarios, compared with baseline methods.

Cell Image Segmentation System Based on Improved UNet

Author: Lambert Lin

Research Committee:

• Jens Koch

Abstract:
A reduced-basis Eigenvector Continuation approach for accelerating superconducting qubit energy spectrum simulations is developed to address the computational cost of repeated Hamiltonian diagonalization. The method introduces monolevel and multilevel schemes with PCA-based stabilization for fluxonium and 0-π qubit systems, achieving exponential error convergence and up to 30× speedup for sparse 0-π qubit Hamiltonians while maintaining high accuracy.

Eigenvector Continuation for Superconducting Qubits

Author: Wei Liu

Research Committee:

• Aggelos Katsaggelos, Electrical and Computer Engineering

Abstract:
A variational deep learning framework for robust face recognition under atmospheric turbulence is developed by integrating a Variational Autoencoder with an ArcFace-based recognition backbone. The proposed VAE-FR and JDVAE-FR models improve feature extraction by modeling image degradation from blur and noise, while maintaining efficient inference through decoder removal after training. Evaluated on the IJB-C benchmark, JDVAE-FR achieves AUC values of 99.6932% and 99.7317% under challenging blur-and-noise and noise-only conditions, respectively, demonstrating enhanced recognition robustness.

Variational Deep Atmospheric Turbulence Restoration and Correction for Face Recognition with InsightFace

Author: Zheng Liu

Research Committee:

• Russ Joseph, Electrical and Computer Engineering

Abstract:
A configurable RISC-V vector architecture simulator based on gem5 is used to evaluate the impact of scalar core design on system performance. Experiments across scalar and vectorized benchmarks demonstrate that scalar core configurations significantly influence execution time and instruction throughput, even in vector-intensive workloads. The results highlight the importance of scalar processing elements and architectural flexibility in optimizing vector system performance.

Vector Processor Performance Model

Author: Suet Ching Luk

Research Committee:

• Stephen Xia, Electrical and Computer Engineering

Abstract:
A fine-tuned AV-HuBERT framework for robust audio-visual speech recognition in drone noise environments is developed by replacing the original Transformer encoder with a Conformer and incorporating Compression and Recovery modules. The enhanced architecture improves dependency modeling and noise resilience, achieving better performance than the original AV-HuBERT model on the LRS2 dataset and custom drone noise recordings. The results demonstrate the effectiveness of architectural modifications for challenging acoustic conditions.

Finetuning Audio-visual model in noisy environments

Author: Yunyi She

Research Committee:

• Zachary Wood-Doughty, Computer Science

Abstract:
A comparative analysis of structured and unstructured data approaches for predicting ICU length of stay within the first 24 hours is conducted using the MIMIC-IV dataset. Structured clinical data combined with classical machine learning models, including XGBoost, achieves the highest performance with an AUROC of 0.805 and R² of 0.352, while transformer-based models using clinical notes provide comparable accuracy with greater computational requirements. The findings highlight the tradeoffs between predictive performance, resource efficiency, and clinical integration when selecting data modalities for early ICU outcome prediction.

Early ICU Length-of-Stay Prediction on MIMIC-IV: A Dual Approach with Clinical Features and Textual Notes

Author: Aoran Sun

Research Committee:

• Lee Cooper, Electrical and Computer Engineering
• Jeffery Goldstein, Feinberg School of Medicine, Pathology

Abstract:
A deep learning approach for predicting prostate cancer molecular subtypes from H&E-stained whole-slide images is developed to provide an alternative to molecular testing. Using a patch-wise attention-based multiple instance learning framework with a CNN backbone, the method achieves AUCs of 0.774 for PAM50 and 0.724 for PSC on 641 slides from 323 patients. Visualization analyses indicate clearer discrimination of basal than luminal subtypes, highlighting the promise of AI-based image analysis as a scalable and cost-effective tool for molecular subtyping.

A Deep learning approach for Prostate Molecular Subtype Prediction

Author: Jiachen Sun

Research Committee:

• Seng-Tiong Ho, Electrical and Computer Engineering

Abstract:
A hybrid plasmonic waveguide design for compact and efficient optical integrated circuits is developed to achieve strong light confinement while reducing intrinsic ohmic losses. The approach evolves from a basic Ag–InP structure to designs incorporating an InGaAs gain layer and etched InP ridge geometry to improve mode confinement and gain overlap. The optimized Ag–InGaAs–InP ridge waveguide demonstrates low loss, high confinement, and positive net gain, supporting its potential for advanced plasmonic photonic devices.

Optimization of Plasmonic Waveguide Structure with Optical Gain for Realization of Lossless Plasmonic Integrated Circuits

Author: Jack Thoene

Research Committee:

• Nivedita Arora, Electrical and Computer Engineering

Abstract:
A low-cost wireless sensor system for real-time monitoring of CAM photosynthesis is developed to overcome limitations of traditional bulky and expensive measurement tools. The platform integrates biocompatible electrodes, low-power electronics, LoRa communication, anomaly detection, and tinyML techniques to enable scalable, long-term data collection in field environments. The system provides researchers with real-time insights and improved data management capabilities for plant physiology studies.

A Novel Electrochemical IoT Sensor for Cheap, Scalable, and Real-Time Plant Metabolism Monitoring of CAM Plants

Author: Jiayu Wang

Research Committee:

• Mohammed Elbaz, Feinberg School of Medicine, Radiology

Abstract:
A cascade enhancement framework for cine cardiac MRI analysis is developed to improve automated biventricular segmentation by progressively refining image quality across contrast, sharpness, and inhomogeneity domains. The approach enhances images using high-quality references from within the dataset before applying a 2D U-Net segmentation model, achieving improved image quality, segmentation accuracy, and functional measurement consistency on the ACDC and M&M datasets. The framework demonstrates potential for more reliable automated cardiac MRI analysis.

Cascade Disentangled Quality Enhancement for Improved Biventricular Segmentation in Cine CMR

Author: Zihan Wang

Research Committee:

• Lirong Yan, Feinberg School of Medicine, Radiology

Abstract:
A multi-scale cerebrovascular image analysis framework combining MRI and deep learning is developed for improved vascular segmentation and physiological assessment. The 4DST U-Net spatiotemporal convolutional model enables accurate and generalizable cerebral vessel segmentation in 4D MRA, while analysis of arterial damping capacity and perivascular space enlargement reveals associations with vascular changes and cognitive function. The findings provide tools for advancing early diagnosis and personalized treatment of neurovascular and cognitive disorders.

Advanced Neurovascular Image Processing and Analysis Using Deep Learning: Non-Contrast Enhanced 4DMRA Segmentation and Quantitative Assessment of Perivascular Spaces

Author: Ruiqin Wu

Research Committee:

• Mark Hersam, Materials Science & Engineering

Abstract:
A finite element simulation framework for neuromorphic devices is developed to analyze memtransistors and organic electrochemical transistors (OECTs) using COMSOL Multiphysics. The models capture key electrostatic and transport mechanisms, demonstrating improved gate control in back-gated memtransistors and accurate reproduction of OECT experimental behavior, including high on/off ratios and p/n-type characteristics. The validated framework supports optimization of low-power, scalable neuromorphic devices for bioelectronics and brain-inspired computing applications.

Finite Element Simulation of Neuromorphic Devices

Author: Tianyu Xu

Research Committee:

• Hao Zhang, Biomedical Engineering

Abstract:
An integrated visible-light OCT and fluorescence SLO imaging system for in vivo validation of retinal ganglion cell axon imaging is developed to overcome limitations of postmortem validation methods. Using transgenic Eno2-YFP mice, the system demonstrates strong agreement between vis-OCT fibergraphy and SLO measurements, achieving a Pearson correlation coefficient of 0.991 for axon bundle widths. The multimodal platform shows promise for noninvasive, longitudinal assessment of retinal ganglion cell damage in optic neuropathies.

Integrated visible-light optical coherence tomography and fluorescence scanning laser ophthalmoscopy

Author: Xiaoyuan Zhang

Research Committee:

• Qi Zhu, Electrical and Computer Engineering

Abstract:
A large language model–based framework for decoding unvoiced electromyography signals into text is developed to enable silent speech recognition without requiring paired audio data. The approach introduces an EMG adaptor that maps articulatory biosignals into the LLM input space, achieving a 0.49 word error rate and outperforming specialized models by nearly 20% with only six minutes of training data. The results demonstrate the potential of LLMs for interpreting silent speech through surface EMG signals.

Multi-modality and Large Lanuage Model

Author: Yiting Zhang

Research Committee:

• Stephen Xia, Electrical and Computer Engineering

Abstract:
A machine learning framework for human-centered sensing systems is developed to enhance intelligence, usability, and robustness in real-world applications. The work introduces DomAIn for natural language–driven smart home automation, DUal-NET for transformer-based speech enhancement with bone-conduction microphones, and SoleSense for wearable outdoor sound event detection using a compact transformer-convolution model. Together, these systems demonstrate the potential of domain-aware machine learning architectures to enable accessible and deployable contextual computing solutions.

Human-Centered Sensing Across Wearables, Audio, and Smart Environments: A Multi-Modal, Model-Driven Approach

Author: Yubo Zhang

Research Committee:

• Igor Kadota, Electrical and Computer Engineering

Abstract:
A reinforcement learning framework for decentralized spectrum sharing in wireless networks is developed to enable independent source-destination pairs to learn optimal transmission strategies without coordination or shared information. The Fair Share RL approach incorporates state augmentation, risk-aware modeling, and fairness-oriented rewards to improve throughput and equitable resource allocation. Simulation results demonstrate up to 89% greater fairness in challenging conditions and a 48.1% average improvement over standard baselines.

Fair Dynamic Spectrum Access via Fully Decentralized Multi-Agent Reinforcement Learnin

Author: Zhuoyi Zhao

Research Committee:

• Igor Kadota, Electrical and Computer Engineering

Abstract:
A framework for optimizing Age of Information in real-world wireless networks is developed to improve freshness of time-sensitive data for applications such as autonomous vehicles and IoT systems. The approach addresses imperfect AoI knowledge and variable update sizes through MMSE-based estimation, randomized and Max-Weight scheduling policies, and a Lyapunov-based Age-Debt strategy. Analytical and simulation results demonstrate effective AoI reduction under realistic communication constraints.

Optimizing Age-of-Information in Real-World Networks

Author: Thomas Zheng

Research Committee:

• Mark Hersam, Materials Science & Engineering

Abstract:
A neuromorphic computing approach based on multi-terminal memtransistors fabricated from two-dimensional materials such as MoS₂ is investigated to address limitations of conventional AI hardware. The study explores improvements in MoS₂ synthesis, device fabrication, and contact engineering to enable integrated memory and processing while addressing challenges including scalability, operating voltage, and performance optimization. The work advances the development of efficient brain-inspired computing systems by bridging materials science, computer engineering, and neuroscience.

Nanomaterial Synthesis and Memtransistors Fabrication for Next-generation Microelectronics

Author: Yibo Zhou

Research Committee:

• John Rogers, Materials Science & Engineering

Abstract:
A wireless and flexible bioelectronic platform for minimally invasive and non-contact healthcare applications is developed to address clinical monitoring and stimulation needs. The work introduces a battery-free TAVI pacemaker using metasurface-assisted wireless power transfer for deep-tissue cardiac stimulation and a non-invasive breast milk volume monitoring system using flexible antennas and wireless sensing algorithms. These systems demonstrate the potential of advanced bioelectronics for wearable, real-time, and patient-centered healthcare technologies.

Design and Evaluation of Wireless and Flexible Bioelectronic Systems: From Implantable Pacemakers to Non-Invasive Milk Volume Monitoring