Computer science and systems (Tacoma)

Permanent URI for this collectionhttps://digital.lib.washington.edu/handle/1773/38512

Browse

Now showing 1 - 20 of 68

Distributed FaaSRunner: Enabling Reproducible Multi-node, Multi-threaded Function-as-a-Service Endpoint Testing
(2025-10-02) Kondo, Tomoki; Lloyd, Wes J
Today's cloud native applications are often built using a service-oriented architecture supported by many microservices hosted using serverless Function-as-a-Service (FaaS) platforms. The vast majority of serverless function benchmarks and tests are performed using a single-client machine to generate workloads against highly scalable serverless backends. However, single-client test engines can lack sufficient computational resources and network bandwidth to adequately stress serverless backends. Testing tools such as Apache JMeter support orchestrating tests using multiple client nodes, but lack the ability to orchestrate sophisticated tests with distributed workload patterns common with real-world serverless workloads. In this thesis, we introduce Distributed FaaSRunner, a distributed test tool which supports the ability to reproduce multi-node serverless workloads using traces derived by ingesting serverless function log files and randomly generated workload traces. We test Distributed FaaSRunner's ability to precisely reproduce serverless function request dispatch and arrival time using various test cluster configurations. Using globally distributed clients, we predict latency to adjust workload trace event dispatch times to reproduce original request arrival latency. We demonstrate that Distributed FaaSRunner can reproduce both temporal and spatial characteristics of serverless workloads, enabling new capabilities to assess performance of FaaS platforms beyond traditional load testing.
Enhancing Privacy in AI: Differential Privacy in Multiparty Computation
(2025-08-01) Pentyala, Sikha; De Cock, Martine
Artificial Intelligence (AI) based applications provide a lot of convenience but often rely on sensitive data of a personal nature to work well. AI pipelines raise input privacy concerns when AI models need to be trained across combined data from multiple data holders who may not be willing or even legally allowed to disclose their data to each other. Similarly, output privacy concerns arise when trained AI models are deployed in production and inadvertently leak private information about the individuals in the training data. A popular approach to address input privacy is Federated Learning (FL), a paradigm in which models are trained in a distributed manner so that raw personal data never leaves the source. State-of-the-art techniques to mitigate output privacy risks use Differential Privacy (DP), which obfuscates the presence of individual records in the training data by adding noise. Existing solutions combining traditional FL and DP to provide input and output privacy at the same time typically cater to specific data partitions (horizontal or vertical) and sacrifice a lot of accuracy to achieve privacy. In this dissertation, we focus on providing both input and output privacy guarantees when data is distributed across multiple data holders irrespective of the data partitioning. We provide solutions to preserve privacy when (a) training discriminative machine learning models for prediction; (b) mitigating biases in model predictions; (c) training AI models for synthetic data generation. All our solutions are grounded in novel Secure Multi-Party Computation (MPC) protocolsto provide input privacy for any data partition -- offering a single solution for horizontal, vertical, or mixed partitions. To simultaneously provide output privacy while maintaining high utility, we leverage the idea of a ''DP-in-MPC'' paradigm through the development of MPC protocols that emulate centralized DP even when the data resides with multiple data holders in a distributed manner. Our research is characterized by the quest for such solutions that (1) enable training of high utility AI models, (2) in a manner sufficiently efficient for use in practice, (3) without compromising individual privacy.
GraphQL vs. REST: Performance and Scalability Analysis for Serverless Applications
(2025-08-01) Jin, Runjie; Lloyd, Wes J
This thesis presents a comprehensive performance, scalability, and cost comparison of GraphQL and Representational State Transfer (REST) APIs within the context of serverless computing. While REST is the conventional choice for API implementation, its architectural style which is designed for network-based applications, specifically client-server applications, can lead to inefficiencies, such as over-fetching and under-fetching, leading to potential performance and price penalties in pay-per-use serverless environments. This work investigates GraphQL as a flexible and efficient interface alternative for two distinct and representative serverless application use cases: a CPU-bound image processing pipeline and a data-intensive relational database application.For the CPU-bound pipeline, experimental results demonstrate that GraphQL reduces client-perceived Round Trip Time (RTT) by eliminating network latency associated with multiple client-to-server round trips required to orchestrate the workflow with REST. For the data-intensive workload, GraphQL implementations show content-dependent performance compared to REST, with Apollo demonstrating 25-67\% performance improvements over REST on most operations, but worse scalability than REST under very high workloads. Collectively, these findings illustrate that GraphQL provides advantages for serverless applications. The nature of these advantages is context-dependent, from orchestrating tasks in multi-step CPU-bound workflows to data-fetching from a relational database, establishing GraphQL as a compelling architectural alternative for modern cloud-native applications.
Supporting bioinformatics analysis using a hybrid cloud and HPC architecture
(2025-05-12) McKeever, Patrick; Yeung, Ka Yee
The exponential growth of next-generation sequencing data requires novel strategies for storage, transfer, and processing of said data. We present a scheduler a based on the Temporal.io workflow framework which enables two key optimizations of bioinformatics workflows. Firstly, we enable users to transparently map workflow steps to diverse execution environments, including high-performance computing (HPC) resources managed by the SLURM resource manager. When tested on a Bulk RNA sequencing workflow, this feature allows a 26% reduction in credit consumption on the NSF Bridges 2 supercomputer by performing adapter trimming locally and all other steps on the supercomputer. Secondly, we enable asynchronous execution of workflows, a feature which guarantees that workflows will achieve reasonable resource utilization even when the scheduler cannot make use of a system's full RAM and CPU resources. When benchmarked on the same Bulk RNA sequencing workflow, this optimization facilitates a reduction in workflow makespan of between 13% and 23%, depending on the exact workflow configuration. Taken together, these features will enable reductions in the cost and time requirements of bioinformatics pipelines for researchers.
From Therapy to Treatment: Transforming Healthcare Support with Large Language Models
(2025-05-12) Filienko, Daniil; De Cock, Martine
The rapid advancement of Large Language Models (LLMs) has opened new avenues for AI-assisted healthcare, particularly in chronic disease management. This study explores the application of in-context learning methods to enhance LLMs' ability to deliver Problem Solving Therapy (PST) and support tuberculosis (TB) treatment adherence. We investigate how LLMs can improve the quality and empathy of AI-driven therapy sessions. Additionally, we propose the integration of LLMs into digital adherence technologies to facilitate interactive patient-provider communication during TB treatment. We leverage prompt engineering, Retrieval Augmented Generation (RAG), and multi-agent systems. Our evaluation across both projects employs both automatic metrics and expert human assessment to analyze the effectiveness of these AI-driven interventions. Findings indicate that while LLMs provide a promising tool for enabling better ongoing care for people with chronic disease across different fields, challenges remain in maintaining privacy, safety, and ethical considerations. This research contributes to the growing field of AI-enhanced healthcare, highlighting the potential and limitations of LLMs in bridging mental health and infectious disease treatment gaps.
A Support System for Diacritic-aware Classical Arabic Language Processing: Integration of Speech, Text, and Vision Modalities
(2025-01-23) Aly, Adel AbdelSabour Ahmad; Ali, Mohamed
Artificial intelligence (AI) impresses us daily, outperforming humans in complex games and tasks. Yet, AI and Large Language Models (LLMs) stumble to grasp a language that is thousands of years old. It's Arabic, where subtle diacritical marks can completely alter a word's meaning. Top language models like ChatGPT and Google's Gemini face challenges with Arabic's unique features, potentially leading to critical misunderstandings. The main obstacle is adapting successful NLP systems from other languages to Arabic without understanding its distinct nuances. This dissertation presents a support system—a Multimodal Integration System—for diacritic-aware Classical Arabic language processing. The system integrates speech, text, and vision modalities to address the unique challenges of Arabic's rich linguistic features, such as diacritics and linguistic styles. Arabic has multiple correct linguistic styles, each preserving the same text but with different diacritics. These variations reflect regional dialects, adding meaning, and altering grammar and rhetoric. The Holy Quran, with its 20 linguistic styles based on a single core text, serves as our ideal dataset. We developed innovative databases and models that push the boundaries of Arabic language processing. Our scalable databases store texts in 7 different Arabic linguistic styles. QR-Vision excels at recognizing precise diacritics in images. QRDiaRec adds diacritics to un-diacritical text in various styles. QRSR and DASAM specialize in speech processing and alignment for Arabic diacritical text and voice. SemSim stands out as a dual-space similarity explorer, analyzing numeric and semantic data. Our methodology involves advanced techniques in data modeling, data quality validation, deep learning, computer vision, signal processing, and interactive data visualization.Our findings demonstrate improvements in addressing key challenges in Arabic Natural Language Processing (NLP). For text processing, we created an Automated Diacritization Deep Learning model. The model supports multiple Arabic diacritical styles, a unique feature in the field. Our best-performing model achieved a 94.2% accuracy rate in adding correct diacritics to Arabic text. In image processing, we built a specialized Optical Character Recognition (OCR) model for diacritic-aware Arabic text. Our OCR model reached an accuracy rate of 91.67% showing an improvement over the existing models. For speech processing, we developed two key systems. The first system, operating at the sentence level, combines our novel FuzTPI algorithm with machine learning models. This hybrid approach achieved up to 96% accuracy in audio segmentation and text-audio classification. Our second system focuses on word-level segmentation and alignment for Arabic diacritic-based speech. It achieved RÂ² values of 0.959 for word start times and 0.957 for end times. These results show an improvement over existing Arabic speech recognition technologies. The dissertation is structured around these three modalities, with each section detailing the challenges, methodologies, and results achieved in processing Arabic with diacritics. They have far-reaching implications for applications such as machine translation, information retrieval, speech recognition, natural language understanding, educational technology, and the preservation of linguistic heritage. By addressing the unique challenges of Arabic diacritics across multiple modalities, this research paves the way for more nuanced and culturally sensitive AI applications in Arabic-speaking contexts.
Privacy Preserving Machine Learning for Next Day Pain Prediction
(2024-04-26) Engavle, Ashutosh Vilas; De Cock, Martine
The availability of healthcare data is critically limited due to stringent privacy regulations, ethical considerations, and the intrinsic sensitivity of medical information. This scarcity hampers research and development in medical science, ultimately affecting the advancement of healthcare services and patient outcomes. Synthetic data emerges as a potent solution to this challenge, offering a pathway to bolster data accessibility while safeguarding patient privacy. This thesis explores the multifaceted issue of next day pain prediction with machine learning models and the limited availability of patient data to train such models, and delves into the potential of synthetic data to bridge this gap. By generating realistic, non-personal data that mimics the statistical properties of real healthcare datasets, synthetic data provides a viable alternative for research and analysis, circumventing privacy concerns. We use methodologies for synthetic data generation with and without privacy, and evaluate their effectiveness and utility for next day pain prediction in patients with Juvenile Idiopathic Arthritis and lupus. We compare the utility of synthetic data with that of real data by training models on both kinds of data and evaluating the trained models on real data. The findings indicate that machine learning models are able to do next day pain prediction. We also see that marginal based synthetic data generation methods can create synthetic data with good utility with substantial privacy guarantee for this task.
TETRA: Time- and Energy-Aware TOPSIS-based Resource Allocation
(2024-04-26) Paruchuri, Sri Vibhu; Al-Masri, Eyhab
With the exponential growth of IoT devices, there has been an increasing demand for distributed computing paradigms such as edge computing and fog computing to address the limitations of cloud computing. Resource scheduling is a critical aspect across the different layers, as it ensures that the available resources are efficiently utilized and allocated to different tasks. Most of the existing resource scheduling algorithms for fog computing environments focus primarily on performance metrics such as makespan, resource utilization, and cost separately. However, there is a need for dynamic multi-objective optimization techniques that can be energy-aware while not compromising on makespan. In this thesis, we introduce a novel resource scheduling algorithm for fog computing environments that optimizes time and energy consumption, which ensures higher performance and lower data center costs. The algorithm considers all the available Virtual Machines (VM) in the fog computing environment. Then, it uses the Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS), which is a multi-criteria decision analysis (MCDA) method, to identify the optimal resources. Our algorithm considers multiple computational parameters such as Million Instructions Per Second (MIPS), the number of processing cores, and thermal design power (TDP) to rank available resources. We conducted a series of experiments, and our algorithm achieves a multi-objective optimization for scheduling IoT tasks on higher-ranked resources resulting in a 7%, 19% and 25% optimization rates in makespan over Best-Fit, Greedy and First-Fit algorithms respectively. In addition, the optimizations in energy consumption over the Best-Fit, Greedy and First-Fit algorithms from our experiments were 1%, 41% and 27%, respectively.
Privacy Vulnerabilities in Marginals-based Synthetic Data
(2024-04-26) Golob, Steven; De Cock, Martine
Synthetic data generation (SDG) lauds the benefit of augmenting, enhancing, and safeguarding real data, which in many applications is scarce. When acting as a privacy-enhancing technology, SDG aims to exclude any personally identifiable information from the underlying real data, all while maintaining important statistical properties that keep it useful to data consumers. Many SDG algorithms provide robust differential privacy guarantees. However, we show that those that preserve marginal probability statistics of the underlying data leak more information about individuals than has been previously understood. We demonstrate this by conducting a novel membership inference attack, MAMA-MIA, on three state-of-the-art differentially private SDG algorithms: MST, PrivBayes, and RAP. We present the heuristic for our attack on marginals-based SDG algorithms here. It assumes knowledge of auxiliary "population" data, and also assumes knowledge of which SDG algorithm was used. We use this information to adapt the recent DOMIAS attack to MST, PrivBayes, and RAP. Our approach went on to win the international SNAKE challenge in November 2023.
Selective Metric Differential Privacy for Language Models
(2024-02-12) Maratkhan, Anuar; De Cock, Martine
Recent advancements in pre-trained language models (LMs) have led to many breakthroughs in Natural Language Processing (NLP). When applied for downstream tasks, such as text classifiers or chatbots, LMs can leak information about the large text corpora they were trained on. In privacy-preserving machine learning, it is common to apply Differential Privacy (DP) mechanisms that mitigate such leakage. The traditional notion of DP, where each record in the data is treated as sensitive, does not translate well to NLP tasks since some token sequences - such as addresses and social security numbers - may be sensitive while others are not. We introduce the new notion of Selective Metric Differential Privacy (SMDP) and a concrete mechanism to realize SMDP. To this end, we draw upon the recently proposed notions of Selective DP, in which records are treated as sensitive or not, and Metric DP, in which the notion of adjacent inputs is relaxed through the use of a metric. Our experiments show that GPT models trained on data privatized with our SMDP approach have higher utility than with Metric DP while preserving the same level of privacy protection.
Text-Supervised Local Feature Mixup Towards Long-Tailed Image Categorization
(2023-09-27) Franklin, Richard Samuel; Hu, Juhua
In many real-world applications, the frequency distribution of class labels for training deep visual models can exhibit a long-tailed distribution that challenges traditional approaches of training deep neural networks, which require heavy amounts of balanced data. Gathering and labeling data to balance out the class label distribution can be both costly and time-consuming. Many existing solutions that enable ensemble learning, re-balancing strategies, and fine-tuning applied to deep neural networks are limited by the inert problem of few class samples across a subset of classes. Recently, vision-language models like CLIP have been observed as effective solutions to zero-shot or few-shot learning by grasping a similarity between vision and language features for image and text pairs. Considering that large pre-trained vision-language models may contain valuable side textual information for minor classes, in this work, we propose to leverage text supervision to tackle the challenge of long-tailed learning for visual recognition. Furthermore, we propose a novel local feature mixup technique that takes advantage of the semantic relations between classes recognized by the pre-trained text encoder to further help alleviate the long-tailed problem. Our empirical study on several benchmark long-tailed tasks demonstrates the effectiveness of our proposal with a theoretical guarantee.
Efficient Transfer Learning using Pre-trained Models on CT/MRI
(2023-09-27) Guobadia, Nicole; Hu, Juhua
The medical imaging field has unique obstacles to face when performing computer vision classification tasks. The retrieval of the data, be it CT scans or MRI, is not only expensive but also limited due to the lack of publicly available labeled data. In spite of this, clinicians often need this medical imaging data to perform diagnosis and recommendations for treatment. This motivates the use of efficient transfer learning techniques to not only condense the complexity of the data as it is often volumetric, but also to achieve better results faster through the use of established machine learning techniques like transfer learning, fine-tuning, and shallow deep learning. In this paper, we introduce a three-step process to perform classification using CT scans and MRI data. The process makes use of fine-tuning to align the pretrained model with the target class, feature extraction to preserve learned information for downstream classification tasks, and shallow deep learning to perform subsequent training. Experiments are done to compare the performance of the proposed methodology as well as the time cost trade offs for using our technique compared to other baseline methods. Through these experiments we find that our proposed method outperforms all other baselines while achieving a substantial speed up in overall training time.
Neural Network Guided Variability Detection in Geospatial Data
(2023-09-27) Salama, Abdulrahman M; Ali, Mohamed
Geospatial data refers to data associated with a specific location on the earth's surface. It plays an important role in a wide range of applications, including environmental monitoring, agriculture planning, mapping, and routing. With the increasing availability of geospatial data from various sources, there is a growing need for methods to validate and verify the accuracy and consistency of this data. Variabilities in such data can have significant impacts on the reliability of the derived information and decision-making processes. Thus, detecting these variabilities is of extreme importance for ensuring the quality of geospatial data. This PhD dissertation focuses on the development of deep neural network methods for detecting variabilities in geospatial data. Variabilities refer to differences between datasets that are otherwise expected to be consistent. Variabilities in geospatial data can occur due to various reasons such as measurement errors, misalignments between datasets, different algorithms used in processing metadata, and changes in real-world phenomena over time. The main objective of this dissertation is to present methods for evaluating the accuracy and consistency of geospatial data, detecting and reporting variabilities found in such data, and providing insights into how data is evolving over time. The effectiveness of the proposed methods will be evaluated using real-world datasets in various applications. This dissertation contributes to advancing the field of geospatial data management by providing new and innovative methods for detecting and reporting variabilities in geospatial data empowering decision-making and future planning.
Dictionary-Guided Text Recognition for Smart Street Parking
(2023-08-14) Zhong, Deyang; Hu, Juhua
Smart detection and recognition of the driving environment are critical tasks in the automobile industry, while understanding road signs is a complicated task. When the traffic is heavy or the parking sign is unclear, drivers cannot finish street curbside parking efficiently, which blocks traffic and makes it worse. Numerous object detection and recognition techniques have been employed to address this issue, but the study for automatic street parking sign understanding, particularly street parking text recognition, is relatively limited. This work bridges the gap between scene text recognition and a smart street curbside parking system. Concretely, we propose a smart street parking sign text recognition method that utilizes a large synthetic data and one real parking sign text data. We focus on providing a multi-candidates technique built upon one general text recognition method and including specific parking sign text words in the candidates' dictionary. The former collects more text information and reduces potential errors, while the latter increases specificity and performance for the parking sign text recognition task. We compare the performance of leading text recognition engines with our proposed method in a real parking sign text data set. We show significant improvements, demonstrating the feasibility and superiority of our new proposal.
Distributed Task Scheduling on Cloud Infrastructure for Bioinformatics Workflows and Analysis
(2023-04-17) Morrow, Rick; Hung, Ling-Hong
The datasets analyzed in bioinformatics are large and numerous, requiring complex analysis. The size and complexity of bioinformatics data often makes it impractical for researchers to run analytical workflows on their personal laptops or PCs. Bioinformatics jobs, however, can benefit from greatly reduced compute times when parallelized across multiple CPUs or cores. Historically, researchers would run analytical workflows on local or static HPC clusters using batch schedulers like Slurm. Running jobs on HPC clusters can be complicated, as jobs must be created and using scheduler specific scripts. HPC clusters are also typically a shared resource with fixed scalability and the main purpose of schedulers was to queue requests and provide resources when available . Cloud computing presents a cost-effective, scalable, reproducible and on-demand resource for researchers to extend the computing resources available to them without the overhead of acquiring and maintaining on-site infrastructure. Although the cloud provides elastic and scalable resources, there remains the challenge of effectively utilizing cloud computing resources through efficient task scheduling. Specifically, there needs to be some method of queueing and scheduling tasks, and assigning those tasks to available workers. In this project we build a task scheduler that handles bioinformatics workflows, which are split into atomic tasks that can be run in parallel, distributed across an arbitrary mix of computing resources from local machines to cloud resources of arbitrary size. The tasks themselves will be processed by containerized workers that implement a standardized and reproducible execution environment. We assess the performance of our scheduler using a real-world bioinformatics task which aligns a set of short sequences (reads) to a human reference sequence using the Burrows-Wheeler Aligner (BWA). We benchmark and compare our scheduling methodology against sequential processing of bioinformatics data using a Dockerized Burrows-Wheeler Aligner (BWA) and against a script that processes this data in parallel without containerization.
Scalable and cloud-enabled analysis of long read sequencing data
(2023-01-21) Reddy, Shishir; Yeung, Ka Yee
Long-read sequencing has great promise in enabling portable, rapid molecular-assisted diagnoses. Applications of long-read sequencing include improved prognosis of critically ill patients through variant detection along with rapid genetic diagnoses. A key challenge in democratizing long-read sequencing technology in the biomedical and clinical community is the lack of graphical bioinformatics software tools which can efficiently process the raw data, support graphical output and interactive visualizations for interpretations of results. Another obstacle is that high performance software tools for long-read sequencing data analyses often leverage graphics processing units (GPU), which is challenging and time-consuming to configure, especially on the cloud. Many solutions can be explored in long-read sequencing including the addition of graphical bioinformatics software tools, hardware acceleration such as Graphics Processing Units (GPUs), or optimization with Tensor Processing Units (TPUs). Long-read sequencing workflows for diagnosis involve several steps that can be hardware-accelerated and optimized using various processing methods. Optimizing long-read sequencing workflows through hardware-acceleration can reduce turnaround times of diagnoses from days to hours. Our goal is to create and optimize long-read sequencing workflows to build rapid, cost-effective solutions for cancer detection and diagnosis on the cloud. This thesis introduces two containerized, hardware-accelerated long-read sequencing analysis workflows for fusion analysis and variant-calling. The fusion analysis workflow introduces a fusion finding tool -- the Biodepot Fusion Finder (BFF) -- capable of rapidly detecting fusions and calculating sample enrichment. This fusion workflow is benchmarked for accuracy and compared to the fusion finding software LongGF on cell-line and patient samples of nanopore data. The variant-calling workflow uses PEPPER-Margin-Deepvariant to call structural variants in a cloud-based GPU-enabled environment. This workflow is benchmarked for accuracy between GPU and CPU versions of the variant-calling software for better visibility in which specific stages of the pipeline benefit from hardware acceleration.
Public Cloud Virtual Machine Co-residency: Prediction and Implications
(2022-09-23) Sharma, Madhuri Suresh; Lloyd, Wes J
Infrastructure-as-a-Service (IaaS) cloud platforms provide virtual machines (VMs) on demand to users hosted on the public cloud using shared or private physical servers. Reserving VMs on private dedicated hosts is expensive compared to renting the same on public shared servers, thus public shared servers are a widely preferred option for application deployment by cloud consumers. Despite considerable efforts to improve the performance of VMs on the public cloud, significant performance variation is still possible when a large number VMs share a single physical server. The information about VM co-residency is abstracted by the cloud provider and remains unknown to the cloud consumer. Due to this abstraction cloud consumers tend to make less informed decisions and choices when creating VMs on public clouds. This research evaluates performance degradation due to resource contention and utilizes memory benchmarks as performance metrics to predict VM co-residency evaluated across different Amazon Elastic Compute Cloud (EC2) VM placement groups. We conducted performance experiments leveraging memory benchmarks, to study performance implications for memory-intensive workloads executing on co-resident VMs. The benchmarking results obtained were used to train machine learning models to predict VM co-residency. The VM co-residency predictions are evaluated by launching VMs using EC2 placement groups, a feature that allows users to influence the physical placement of VMs on EC2.
Enhanced Early Sepsis Onset Prediction: A Multi-Layer Approach
(2022-07-14) Ewig, Kevin; Hu, Juhua
Sepsis is a life-threatening organ malfunction caused by the host's inability to fight infection. Without proper and immediate treatment, sepsis can lead to death. Early diagnosis and medical treatment of sepsis in critically ill populations at high risk for sepsis and sepsis-associated mortality are vital to providing the patient with rapid therapy. The mortality rate increases with each hour that antibiotic treatment is delayed. Studies show that advancing sepsis detection by 6 hours leads to earlier administration of antibiotics, which is associated with improved mortality. However, clinical scores like SequentialOrgan Failure Assessment (SOFA) are not applicable for early sepsis onset prediction, while machine learning algorithms may be able to capture the progressing pattern for early prediction. Therefore, this thesis aims to develop a machine learning model that predicts sepsis onset 6 hours before it is suspected clinically. Although some machine learning algorithms have been applied to sepsis prediction, many of them did not consider the fact that six hours is not a small gap. To overcome this big gap challenge for early sepsis detection, this thesis explores a multi-layer approach in which the likelihood of sepsis occurring earlier than 6 hours is output from the 1st layer and fed to the 2nd layer as features to help predictions for the 6-hour horizon. Moreover, we use the hourly sampled data like vital signs in an observation window to derive a temporal change trend to further assist in sepsis prediction, which however is often ignored by previous traditional machine learning algorithms. Our empirical study shows that both the multi-layer approach to alleviating the 6-hour gap and the added features to capture the temporal trend can help improve the performance of early sepsis prediction.
Sets of Sub-Sequences based Sepsis Prediction for ICU Trauma Patients
(2022-07-14) Huang, Sijin; Teredesai, Ankur
Sepsis is an extreme inflammatory response of the body to an infection. It is one of the leading causes of death in ICUs worldwide, resulting in approximately 25% mortality in critically ill populations. Early identification and intervention are crucial to reducing sepsis-associated mortality and improving patient prognosis because severe sepsis cases can lead to organ failure and other life-threatening complications. Diagnosis of Sepsis is challenging in terms of diagnostic accuracy and timeliness due to ambiguous symptoms and individual differences. In recent years, numerous efforts have been made using machine learning methods for sepsis prediction. However, there are still very limited successful implementations due to limitations in consistency of data input, which cannot fit the characteristics of uncertain time intervals and large number of missing values in real-world scenarios. In this study, we propose an innovative approach to predict sepsis occurrence in real-time, using a flexible graph structure to model patient health records, predicting future sepsis risk at any time after the first 48 hours using observations data from the past 12 hours. To the best of our knowledge, our proposed approach is the first ever implementation that uses a graph representation to overcome the problem of irregular input features and continuous risk prediction thereby improving the compatibility of the model in clinical settings. Experiments on multi-year longitudinal data from a large level-1 trauma center demonstrate the effectiveness of our approach.
Real-Time Parking Sign Detection for Smart Street Parking
(2022-04-19) Jin, Yin; Hu, Juhua
Correctly interpreting the complex parking sign and finding a suitable parking spot is always a challenging task to do in a short time. An automatic parking sign understanding assistance can improve the efficiency of a driver’s daily life. Besides, as Tesla becomes popular, automatic driving is also becoming a prevalent task. However, the current studies on self-driving steering command mainly focus on the driving part. The smart parking task is still an open task. In this thesis, we aim to handle a subtest of the automatic parking sign understanding problem, that is, real-time parking sign detection, to provide real-time parking sign localization for further sign interpretation. Specifically, with a real-time video streaming of the street view as input, we aim to detect parking signs on the street that will inform the parking rules over the current road. In this thesis, we achieved two main goals: 1) generating a diverse parking sign dataset with a size over 4,000 that covers complex street view; 2) training a well-performance real-time parking sign detection model with Yolov5. Our parking sign detection model achieves a mean average precision at Intersection over Union threshold 50% (mAP@.5) of 0.968 and an inference speed of 6.1ms per image that meets the real-time detection requirement. Moreover, the model size is only 14.4 MB which is small enough to fit in small and less powerful devices like mobile phones or autonomous cars.

Browse

Recent Submissions