Due to rapid growth in computing power, the collection of high dimensional and complex datasets is increasingly feasible. To reap their full benefit, novel analytic strategies may be required. Application of such methods remains limited in certain epidemiologic research areas. The overarching aim of this dissertation was to apply novel analytic strategies with close ties to causal inference and statistical learning theory to gain new insights into well-studied areas of perinatal epidemiology. In Study 1, we explored whether the association between short interpregnancy intervals (i.e., the end of one pregnancy to the start of the next) and increased risk of preterm birth may be due to residual confounding in three populations (n=693 American Indian and n=728 white women from the Northern Plains, U.S., and n=783 mixed ancestry women from the Western Cape, South Africa). Using data from the prospective Safe Passage cohort (2007-2015), we applied propensity score methods to control for a variety of sociodemographic and reproductive factors. A third-to-half of women with <6 months intervals had propensity scores that largely did not overlap with those of women with 18-23 months intervals. Since the propensity score models included factors related to both interpregnancy interval and preterm birth, these findings suggest the possibility of strong confounding in all three populations. The pooled associational estimate with preterm birth was attenuated in the propensity score trimmed and weighted data (risk ratio 1.4, 95% CI 0.75-2.6) compared with the crude results (risk ratio 1.7, 95% CI 1.1-2.7). However, the sample size and precision were reduced after propensity score trimming, and several covariates remained imbalanced. The data demonstrated the complexity of the processes leading to interpregnancy interval length. These issues may have been difficult to identify without comprehensive confounder data and with other methods, such as traditional regression adjustment. In Study 2, we examined the relative importance of timing (first trimester versus second/third trimesters) and degree of gestational weight gain in relation to infant size at birth (small-and-large-for-gestational age) among women with obesity using data from a medical records-based case-cohort study (Pittsburgh, PA, 1998-2010). We operationalized serial antenatal weight measurements as above, below, or within the current recommended ranges for U.S. pregnancies, i.e., 0.2-2.0 kg total gain in the first trimester and 0.17-0.27 kg per week in the second and third trimesters (based on group based trajectory modeling). Data were analyzed by obesity class (n=1290 in the class I subcohort, n=1247 class II, n=1198 class III). Our findings supported the current clinical guidelines, except for women with class III obesity. Among women with class III obesity, lower than recommended gain in the second and third trimesters was associated with decreased risk of having a large-for-gestational age infant (adjusted risk ratio 0.76, 95% CI 0.51-1.1), while not increasing small-for-gestational age (SGA) risk (adjusted risk ratio 1.0, 95% CI 0.63-1.7). Our results were in agreement with findings from several other studies of women with obesity using other methodologies to operationalize gestational weight gain. In Study 3, we used hierarchical clustering to explore latent groups of placental pathology features. We also investigated whether the placental clusters, in addition to birthweight percentiles, were beneficial to explain the variability of select adverse pregnancy outcomes. Data were from the Safe Passage Study (same as Study 1, n=2005). We identified one cluster with low prevalence of abnormalities (60.9%) and three clusters that mapped well to the expert consensus-based Amsterdam criteria: severe maternal vascular malperfusion (5.8%), fetal vascular malperfusion (11.1%), and inflammation (22.1%). The clusters were weakly-to-moderately associated with certain antenatal risk factors, pregnancy complications, and neonatal outcomes. Birthweight percentiles plus the placental clusters was better able to explain the variance of select adverse outcomes, compared with using small-for-gestational age only. This study serves as proof-of-concept that machine learning methods, and placental data, may aid in the identification and etiologic study of certain adverse pregnancy outcomes. In sum, all three studies support that the application of novel analytic methods to high-dimensional datasets may expand our understanding of certain causal questions, even ones that have been broached before, although, as seen in Study 2, such research may not always yield novel insights.

Fluids in computer generated imagery can add an impressive amount of realism to a scene, but are particularly time-consuming to simulate. In an attempt to run fluid simulations in real-time, recent efforts have attempted to simulate fluids by using machine learning techniques to approximate the movement of fluids. We explore utilizing machine learning to simulate fluids while also integrating the Fluid-Implicit-Particle (FLIP) simulation method into machine learning fluid simulation approaches.

In this thesis, a methodology and framework were created to detect the seawalls accurately and efficiently in low coastal areas and was evaluated in the study area of Hallandale Beach City, Broward County, Florida. Aerial images collected from the Florida Department of Transportation (FDOT) were processed using eCognition Developer software for Multi-Resolution Segmentation and Classification of objects. Two classification approaches, pixel-based image analysis, and the object-based image analysis (OBIA) method were applied for image classification. However, Pixel based classification was discarded for having less accuracy in output. Three techniques within object-based classification-machine learning technique, knowledge-based technique and machine learning followed by knowledge-based technique were used to compare the most efficient method of classification. While performing the machine learning technique, three algorithms: Random Forest, support vector machine and decision tree were applied to test the best algorithm. Of all the approaches used, the combination of machine learning and a knowledge-based method was able to map the sea wall effectively. / Includes bibliography. / Thesis (M.S.)--Florida Atlantic University, 2021. / FAU Electronic Theses and Dissertations Collection

Many recent successes of machine learning went hand in hand with advances in optimization. The exchange of ideas between these fields has worked both ways, with ' learning building on standard optimization procedures such as gradient descent, as well as with new directions in the optimization theory stemming from machine learning applications. In this thesis, we discuss new developments in optimization inspired by the needs and practice of machine learning, federated learning, and data science. In particular, we consider seven key challenges of mathematical optimization that are relevant to modern machine learning applications, and develop a solution to each. Our first contribution is the resolution of a key open problem in Federated Learning: we establish the first theoretical guarantees for the famous Local SGD algorithm in the crucially important heterogeneous data regime. As the second challenge, we close the gap between the upper and lower bounds for the theory of two incremental algorithms known as Random Reshuffling (RR) and Shuffle-Once that are widely used in practice, and in fact set as the default data selection strategies for SGD in modern machine learning software. Our third contribution can be seen as a combination of our new theory for proximal RR and Local SGD yielding a new algorithm, which we call FedRR. Unlike Local SGD, FedRR is the first local first-order method that can provably beat gradient descent in communication complexity in the heterogeneous data regime. The fourth challenge is related to the class of adaptive methods. In particular, we present the first parameter-free stepsize rule for gradient descent that provably works for any locally smooth convex objective. The fifth challenge we resolve in the affirmative is the development of an algorithm for distributed optimization with quantized updates that preserves global linear convergence of gradient descent. Finally, in our sixth and seventh challenges, we develop new VR mechanisms applicable to the non-smooth setting based on proximal operators and matrix splitting. In all cases, our theory is simpler, tighter and uses fewer assumptions than the prior literature. We accompany each chapter with numerical experiments to show the tightness of the proposed theoretical results.

No description available.

In the last decade, deep learning models have been successfully applied to a variety of applications and solved many tasks. The ultimate goal of this study is to produce deep learning models to improve the skills of forecasting ocean dynamic events in general and those of the Loop Current (LC) system in particular. A specific forecast target is to predict the geographic location of the (LC) extension and duration, LC eddy shedding events for a long lead time with high accuracy. Also, this study aims to improve the predictability of velocity fields (or more precisely, velocity volumes) of subsurface currents. In this dissertation, several deep learning based prediction models have been proposed. The core of these models is the Long-Short Term Memory (LSTM) network. This type of recurrent neural network is trained with Sea Surface Height (SSH) and LC velocity datasets. The hyperparameters of these models are tuned according to each model's characteristics and data complexity. Prior to training, SSH and velocity data are decomposed into their temporal and spatial counterparts.A model uses the Robust Principle Component Analysis is first proposed, which produces a six-week lead time in forecasting SSH evolution. Next, the Wavelet+EOF+LSTM (WELL) model is proposed to improve the forecasting capability of a prediction model. This model is tested on the prediction of two LC eddies, namely eddy Cameron and Darwin. It is shown that the WELL model can predict the separation of both eddies 10 and 14 weeks ahead respectively, which is two more weeks than the DAC model. Furthermore, the WELL model overcomes the problem due to the partitioning step involved in the DAC model. For subsurface currents forecasting, a layer partitioning method is proposed to predict the subsurface field of the LC system. A weighted average fusion is used to improve the consistency of the predicted layers of the 3D subsurface velocity field. The main challenge of forecasting of the LC and its eddies is the small number of events that have occurred over time, which is only once or twice a year, which makes the training task difficult. Forecasting the velocity of subsurface currents is equally challenging because of the limited insitu measurements. / Includes bibliography. / Dissertation (PhD)--Florida Atlantic University, 2021. / FAU Electronic Theses and Dissertations Collection

Introduction: The placenta is a complex, disk-shaped organ vital to a successful pregnancy and responsible for materno-fetal exchange of vital gases and biochemicals. Instances of compromised placental development or function – collectively termed placenta dysfunction - underlies the most common and devastating pregnancy complications observed in North America, including preeclampsia (PE) and fetal growth restriction (FGR). A comprehensive histopathology examination of the placenta following delivery can help clarify obstetrical disease etiology and progression and offers tremendous potential in the identification of patients at risk of recurrence in subsequent pregnancies, as well as patients at high risk of chronic diseases in later life. However, these types of examinations require a high degree of specialized training and are resource intensive, limiting their availability to tertiary care centers in large city centres. The development of machine learning algorithms tailored to placenta histopathology applications may allow for automation and/or standardization of this important clinical exam – expanding its appropriate usage and impact on the health of mothers and infants. The primary objective of the current project is to develop and pilot the use of machine learning models capable of placental disease classification using digital histopathology images of the placenta. Methods: 1) A systematic review was conducted to identify the current methods being applied to automate histopathology screening to inform experimental design for later components of the project. Of 230 peer-reviewed articles retrieved in the search, 18 articles met all inclusion criteria and were used to develop guidelines for best practices. 2) To facilitate machine learning model development on placenta histopathology samples, a villi segmentation algorithm was developed to aid with feature extraction by providing objective metrics to automatically quantify microscopic placenta images. The segmentation algorithm applied colour clustering and a tophat transform to delineate the boundaries between neighbouring villi. 3) As a proof-of-concept, 2 machine learning algorithms were tested to evaluated their ability to predict the clinical outcome of preeclampsia (PE) using placental histopathology specimens collected through the Research Centre for Women’s and Infant’s Health (RCWIH) BioBank. The sample set included digital images from 50 cases of early onset PE, 29 cases of late onset PE and 69 controls with matching gestational ages. All images were pre-processed using patch extraction, colour normalization, and image transformations. Features of interest were extracted using: a) villi segmentation algorithm; b) SIFT keypoint descriptors (textural features); c) integrated feature extraction (in the context of deep learning model development). Using the different methods of feature extraction, two different machine learning approaches were compared - Support Vector Machine (SVM) and Convolutional Neural Network (CNN, deep learning). To track model improvement during training, cross validation on 20% of the total dataset was used (deep learning algorithm only) and the trained algorithms were evaluated on a test dataset (20% of the original dataset previously unseen by the model). Results: From the systematic review, 5 key steps were found to be essential for machine learning model development on histopathology images (image acquisition and preparation, image preprocessing, feature extraction, pattern recognition and classification model training, and model testing) and recommendations were provided for the optimal methods for each of the 5 steps. The segmentation algorithm was able to correctly identify individual villi with an F1 score of 80.76% - a significantly better performance than recently published methods. A maximum accuracy of 73% for the machine learning experiments was obtained when using textural features (SIFT keypoint descriptors) in an SVM model, using onset of PE disease (early vs. late) as the output classification of interest. Conclusion: Three major outcomes came of this project: 1) the range of methods available to develop automated screening tools for histopathology images with machine learning were consolidated and a set of best practices were proposed to guide future projects, 2) a villi segmentation tool was developed that can automatically segment all individual villi from an image and extract biologically relevant features that can be used in machine learning model development, and 3) a prototype machine learning classification tool for placenta histopathology was developed that was able to achieve moderate classification accuracy when distinguishing cases of early onset PE and late onset PE cases from controls. The collective body of work has made significant contributions to the fields of placenta pathology and computer vision, laying the foundation for significant progress aimed at integrating machine learning tools into the clinical setting of perinatal pathology.

During this study a model for predicting the next week's alarm codes based on the past week's alarm codes was developed. The model used alarm data from the location and its surroundings. The model was tuned using hyper parameter optimization to improve performance, this resulted in a model performing better than previous models used on this data set. The performance when adding weather data was evaluated and it was shown that it improved the performance for some alarm codes and the performance for the majority of other alarm codes was not compromised resulting in an improvement in the overall performance. The weather data consisted of temperature, precipitation, cloud coverage, air pressure and wind direction and speed data. Two labeling methods were trialed for the weather data, the first one used the data of the closest weather station for each type of data. The second labeling method used data of the ten closest weather stations within 100 km. The final model using weather data labeled with method 2 had a precision micro average of 0.90, a recall micro average of 0.86, a precision macro average of 0.80 and a recall macro average of 0.77.

Proteins play critical roles in cellular processes of living organisms. It is therefore important to identify and characterize their key properties associated with their functions. Correlating protein’s structural, sequence and physicochemical properties of its amino acids (aa) with protein functions could identify some of the critical factors governing the specific functionality. We point out that not all functions of even well studied proteins are known. This, complemented by the huge increase in the number of newly discovered and predicted proteins, makes challenging the experimental characterization of the whole spectrum of possible protein functions for all proteins of interest. Consequently, the use of computational methods has become more attractive. Here we address two questions. The first one is how to use protein aa sequence and physicochemical properties to characterize a family of proteins. The second one focuses on how to use transcription factor (TF) protein’s domains to enhance accuracy of predicting TF DNA binding sites (TFBSs). To address the first question, we developed a novel method using computational representation of proteins based on characteristics of different protein regions (N-terminal, M-region and C-terminal) and combined these with the properties of protein aa sequences. We show that this description provides important biological insight about characterization of the protein functional groups. Using feature selection techniques, we identified key properties of proteins that allow for very accurate characterization of different protein families. We demonstrated efficiency of our method in application to a number of antimicrobial peptide families. To address the second question we developed another novel method that uses a combination of aa properties of DNA binding domains of TFs and their TFBS properties to develop machine learning models for predicting TFBSs. Feature selection is used to identify the most relevant characteristics of the aa for such modeling. In addition to reducing the number of required models to only 14 for several hundred TFs, the final prediction accuracy of our models appears dramatically better than with other methods. Overall, we show how to efficiently utilize properties of proteins in deriving more accurate solutions for two important problems of computational biology and bioinformatics.

The main aim of this dissertation is to address the security issues of the physical layer of Cyber Physical Systems. The network security is first assessed using a 5-level Network Security Evaluation Scheme (NSES). The network security is then enhanced using a novel Intrusion Detection System that is designed using Supervised Machine Learning. Defined as a complete architecture, this framework includes a complete packet analysis of radio traffic of Routing Protocol for Low-Power and Lossy Networks (RPL). A dataset of 300 different simulations of RPL network is defined for normal traffic, hello flood attack, DIS attack, increased version attack and decreased rank attack. The IDS is a multi-model detection model that provides an efficient detection against the known as well as new attacks. The model analysis is done with the cross-validation method as well as using the new data from a similar network. To detect the known attacks, the model performed at 99% accuracy rate and for the new attack, 85% accuracy is achieved. / Graduate