The widespread adoption of artificial intelligence (AI) and Intelligent Transportation Systems (ITS) technologies has led to the increasing application of AI-based ITS controllers, with the Traffic Signal Controller (TSC) being a prominent example. Reinforcement learning (RL) models have shown promising results for adaptively adjusting traffic light schedules in urban environments through RL-based TSCs (RL-TSCs). The real-world deployment of RL-TSCs involves three key aspects: performance, security, and data privacy. In terms of performance, RL-TSC models need to be designed with consideration for various metrics, such as fair traffic scheduling and air quality impact. To address this, our approach takes into account a multi-objective constrained learning formulation to optimize performance. However, the use of RL-TSCs for automation, by leveraging external inputs, introduces security concerns that require active research to mitigate. We address these security challenges by introducing an innovative defense mechanism. Additionally, the training of RL-TSCs relies on real-world mobility datasets, necessitating the protection of data privacy at different levels of granularity. To minimize the constraints associated with limited real data availability or privacy concerns, we introduce two distinct directions: synthetic trajectory data generation using recent generative AI methods, and location privacy models for raw mobility datasets based on differential privacy, which safeguard individual trajectories and aggregated mobility datasets.This research provides a valuable tool for evaluating the practical deployment of RL-TSCs, particularly in real-world settings where the last mile of implementation and security is paramount. By addressing the key challenges of performance, security, and data privacy, this work aims to facilitate the successful real-world deployment of AI-powered ITS controllers.

Initial fault detection and diagnostics are essential elements to improve the efficiency, safety, and stability of vehicle operation. Diagnostics can make direct and indirect financial impacts on service and support entities in place for the vehicle. Remote diagnostics can reduce vehicle downtime in service centers and increase customer satisfaction, primarily when conducting over-the-air updates or telephone lines. In order to troubleshoot a vehicle, specific tools can be used to look up failure codes stored in vehicle controllers or manually gather failure symptoms through customer service hotlines and remote service technicians. The overall gathering of data, deciphering, and execution of any repairs still consumes precious time and may suffer from potential human errors. Recently, numerous studies have investigated data-driven approaches to improve vehicle diagnostics using available vehicle data. This study investigates a machine learning pipeline to improve automated vehicle diagnostics and prognostics. Using Natural Language Processing (NLP), we demonstrate a comprehensive model to extract the customer and agent interactions from repair-service call transcriptions. This dissertation applies Machine Learning (ML) algorithms to identify accurate failure reports and claims. Also, it classifies the service requests to the proper service department and utilizes the historical service information along with current customer claims to identify possible failed vehicle parts. First, NLP techniques are used to automate the task of crucial information extraction from free-text failure reports (generated within customers' calls to the service department). We have introduced an NLP taxonomy in the automotive domain since known NLP techniques had a weak performance on such texts. We have shown that domain-based NLP processing and feature extraction can help to extract meaningful information from the reports. Deep learning algorithms are employed to validate service requests and filter vague or misleading claims. Various classification algorithms are implemented to classify service requests so that valid service requests can be directed to the relevant service department. We proposed to employ Bidirectional Long Short-term Memory (BiLSTM), along with Convolution Neural Network (CNN) model, which shows more than 18% performance improvement in validating service requests compared to average technicians' capabilities. Furthermore, using domain-based NLP techniques at preprocessing and feature extraction stages along with CNN-BiLSTM-based request validation enhanced the performance of the Gradient Tree Boosting (GTB) service classification model. The performance parameter of the Receiver Operating Characteristic Area Under the Curve (ROC-AUC) reached 0.82. Next, we performed automated failure classification on extracted data to route the claims to the proper service departments. By introducing optimized feature extraction and classification methods, requests can be forwarded to the correct departments with 80% accuracy. This method exceeds the 60% baseline accuracy for an average customer service technician. NLP analysis can also generate technical information from the text report for vehicle and component prognostics that have not been previously studied. Finally, we proposed a novel network structure that employs a multi-variant high-dimensional Markov chain to predict the possible failed component of the next service interval to enhance the CNN-LSTM model performance. The Markov model takes advantage of historical records to identify the most efficient CNN kernels in the network structure. The proposed model significantly improved data classification efficiency in correlated historical records such as vehicle service reports. Compared to conventional CNN-LSTM models, the introduced model demonstrated significant performance enhancement of 8% accuracy, 9% sensitivity, 11% specificity, 10% precision, and 12% f-score by reducing the false positive cases in customer claim classification.

Acute respiratory distress syndrome (ARDS) is a severe form of acute hypoxemic respiratory failure affecting 10% of patients admitted to the intensive care unit (ICU) globally. In-hospital mortality of 29.7%-42.9% has been reported across the spectrum of mild-severe ARDS, and one third of patients with initially mild ARDS will progress to moderate or severe ARDS. Over the last 20 years, multiple studies have reported improved outcomes for ARDS patients using ARDS-targeted therapies. However, ARDS remains persistently under-recognized and challenging to diagnose. Only one third of ICU providers correctly identify ARDS on the first day when diagnostic criteria are met, and less than two thirds ever recognize the diagnosis in the ICU. This under recognition of ARDS may prevent some patients from receiving lifesaving therapies necessary for treating the disease. Attempts to automate ARDS diagnosis using rule-based algorithms have seen limited success, and require subjective analysis of infrequently sampled patient data, like chest radiographs, which limit diagnosis automation, timeliness, and study reproducibility. To improve the current state of the art of ARDS detection technology, we intend to utilize objective and readily available ventilator waveform data (VWD) to improve the recognition of ARDS and develop next generation clinical decision support systems (CDSS) that can function as a pervasive monitoring system for mechanical ventilation management. For this task, we make use of a novel dataset consisting of only VWD from patients receiving mechanical ventilation. We analyze this data using both classical and deep machine learning models and show results that suggest that ARDS can be detected in absence of a chest scan or medical history. This finding highlights that VWD shows promise for use as a future digital biomarker of ARDS pathophysiology. This dissertation is broadly aimed at showing the research challenges we solved and software that we built to create our ML-based ARDS detection models. Our work encompasses multiple sub-studies including: 1) designing software for annotating large amounts of VWD for our ML datasets, 2) developing and training machine learning models to extract important metadata (such as ventilation mode) from VWD, 3) performing a comprehensive clinical validation for existing, but clinically unvalidated respiratory compliance estimation algorithms, 4) training an ML model for early recognition of ARDS using only VWD, 5) improving our initial ML model for ARDS recognition by applying deep learning, and 6) discussing how to integrate all our work into a comprehensive monitoring system for patient ventilation. While our work is specifically focused on ARDS detection, this dissertation broadly highlights the utility of applying modern machine learning tools to underutilized high-frequency physiologic waveform data as a surrogate for manually charted EHR data types in disease detection and early health warning. These warnings and alerts can then be continuously monitored and returned to clinicians at the point of care in the form of actionable advice and disease prognostication. Such improvements will enable pervasive clinical decision support systems that promise to reduce the cognitive burden on care providers, improve quality of care, reduce patient suffering, and reduce overall hospital mortality from deadly conditions such as ARDS.

Deep Learning is a branch of machine learning with a layered structure where each layer gets its input from the previous layer to analyze the data and has shown promising results to aid healthcare providers in various applications. This thesis presents the development and evaluation of various Deep Learning approaches to predict if a hemodynamically unstable patient would be responsive to infusion of intravenous fluids. This thesis also explores various meticulously designed experiments to thoroughly verify the predictive model’s generalization across various carefully curated datasets to represent shocks resulting from different physiologies in pigs. Treatment of patients suffering from shock often involves the infusion of intravenous fluids, also known as fluid bolus therapy (FBT). An increase in cardiac output (CO) of 15% or more after a supply of 500 ml of the fluid bolus indicates fluid responsiveness, and the ground truth labels were designed based on this rule. In addition, the arterial blood pressure (ABP) and central venous pressure (CVP) waveforms were recorded before and after the infusion of each bolus. The period before, during, and after the administration of fluid boluses is known as pre-macrobolus (Premac), macrobolus, post-macrobolus (Postmac), respectively. The deep learning model takes sequences of specific lengths obtained from the ABP and CVP during the premac period and the ground truth labels to classify each bolus to be fluid responsive or not. The models were tuned and trained using nested cross-validation accompanied by grid search algorithms. The results from our experiments suggest that deep learning can offer a satisfactory framework to classify boluses as fluid responsive or fluid non-responsive. In addition, this thesis presents a comprehensive guide to experimentation on various aspects that could potentially affect the performance of deep learning models while classifying one-dimensional data, including input sequence length, model’s architecture, sample weighting in loss functions, normalization, resampling the data, and various methods to sample the data to acquire meaningful inferences from the results. The experiments showcase the restricting nature of small-scale datasets on the deep learning model’s performance. The deep learning model fails to generalize when the training and the test sets contain different physiologies but generalizes better when both the training and the test sets contain a comparable mixture of physiologies.

Neurodegenerative diseases are defined by a progressive loss of selectively vulnerable neuronal populations. Hence, augmenting the ability of a pathologist/expert for the detection/localization of specific pathologies with data-driven deep learning (DL) approaches could have a transformative impact. DL models trained on high-quality annotated datasets can provide a cost-effective and reliable means for deeper phenotyping of neurodegenerative and neurologic diseases. Leveraging digital slide scanners such as the Leica Aperio AT2 and Zeiss Axio Z1, physical tissue slides can be digitized into gigapixel digital whole slide images (WSIs) containing a rich set of information, including diverse morphological patterns, cellular features, and tissue architecture. However, there are severe challenges to fully realize the digitalization and automation of detecting these pathologies and neurodegenerative features. First, the tissue sections can be scanned at ×20 or ×40 objective magnification, resulting in ultra-high resolution WSIs that require a large amount of memory space for storage and processing. To alleviate this issue, we design a general patch-based approach and use plaque quantification in grey and white matter as one example. In this study case, our automated framework can classify, localize, quantify, and visualize the distribution of each type of plaque in grey and white matter, separately. We have tested it across different scanners, brain areas, and storage formats to benchmark its generalizability.

Second, although the proposed supervised learning pipelines can achieve promising results, their performance heavily relies on the quality and quantity of the data and annotations. However, manual annotation to build comprehensive and well-annotated digital pathology datasets can be time-consuming and labor-intensive, which is not scalable for diverse pathological tasks. Therefore, to relieve the heavy reliance on the labeled data, we investigate semi-supervised learning (SSL) and propose four SSL frameworks to enhance the applicability and deployment of SSL algorithms. For example, we propose an imbalanced SSL but make no assumption on the distribution of the unlabeled data; we improve pseudo-labeling and design a deep fusion of SSL and active learning to further reduce the labeling efforts.

Third, besides the annotation cost, data collection can also be expensive and challenging due to the scarcity of pathology data and the heterogeneity of pathology datasets, which can vary significantly regarding image resolution, staining techniques, and scanning equipment. Therefore, to reduce the data collection effort, we study unsupervised domain adaptation (UDA) and efficient adaptation of large-scale pre-trained vision-language models (VLMs). UDA proposes to reduce data annotation costs by leveraging a labeled source domain to transfer the knowledge into an unlabeled target domain. We propose a UDA pipeline to bridge the domain gap and make domain adaptation data and label-efficient. On the other hand, VLMs have shown promising results in their transferability across different downstream tasks. However, traditional fine-tuning may involve billions of parameters to be tuned, which is computation-expensive and has overfitting issues when the downstream task with limited data availability. We propose a parameter-efficient pipeline to adapt VLMs to multiple downstream tasks in a scalable way in terms of processing time and computational cost. In conclusion, our proposed pipelines can potentially achieve data and label-efficient learning for diverse pathological imaging tasks.