Search

Skin cancer has a significant impact on the lives of many individuals annually and is recognized as the most prevalent type of cancer. In the United States, an estimated annual incidence of approximately 3.5 million people receiving a diagnosis of skin cancer underscores its widespread prevalence. Furthermore, the prognosis for individuals afflicted with advancing stages of skin cancer experiences a substantial decline in survival rates. This paper is dedicated to aiding healthcare experts in distinguishing between benign and malignant skin cancer cases by employing a range of machine learning and deep learning techniques and different feature extractors and feature selectors to enhance the evaluation metrics. In this paper, different transfer learning models are employed as feature extractors, and to enhance the evaluation metrics, a feature selection layer is designed, which includes diverse techniques such as Univariate, Mutual Information, ANOVA, PCA, XGB, Lasso, Random Forest, and Variance. Among transfer models, DenseNet-201 was selected as the primary feature extractor to identify features from data. Subsequently, the Lasso method was applied for feature selection, utilizing diverse machine learning approaches such as MLP, XGB, RF, and NB. To optimize accuracy and precision, ensemble methods were employed to identify and enhance the best-performing models. The study provides accuracy and sensitivity rates of 87.72% and 92.15%, respectively. [ABSTRACT FROM AUTHOR]

The COVID-19 pandemic prompted governments worldwide to implement a range of containment measures, including mass gathering restrictions, social distancing, and school closures. Despite these efforts, vaccines continue to be the safest and most effective means of combating such viruses. Yet, vaccine hesitancy persists, posing a significant public health concern, particularly with the emergence of new COVID-19 variants. To effectively address this issue, timely data is crucial for understanding the various factors contributing to vaccine hesitancy. While previous research has largely relied on traditional surveys for this information, recent sources of data, such as social media, have gained attention. However, the potential of social media data as a reliable proxy for information on population hesitancy, especially when compared with survey data, remains underexplored. This paper aims to bridge this gap. Our approach uses social, demographic, and economic data to predict vaccine hesitancy levels in the ten most populous US metropolitan areas. We employ machine learning algorithms to compare a set of baseline models that contain only these variables with models that incorporate survey data and social media data separately. Our results show that XGBoost algorithm consistently outperforms Random Forest and Linear Regression, with marginal differences between Random Forest and XGBoost. This was especially the case with models that incorporate survey or social media data, thus highlighting the promise of the latter data as a complementary information source. Results also reveal variations in influential variables across the five hesitancy classes, such as age, ethnicity, occupation, and political inclination. Further, the application of models to different MSAs yields mixed results, emphasizing the uniqueness of communities and the need for complementary data approaches. In summary, this study underscores social media data's potential for understanding vaccine hesitancy, emphasizes the importance of tailoring interventions to specific communities, and suggests the value of combining different data sources. [ABSTRACT FROM AUTHOR]

This research utilizes the Long-Term Pavement Performance database, focusing on devel-oping a predictive model for flexible pavement performance in the Southern United States. Analyzing 367 pavement sections, this study investigates crucial factors influencing asphaltic concrete (AC) pavement deterioration, such as structural and material components, air voids, compaction density, temperature at laydown, traffic load, precipitation, and freeze–thaw cycles. The objective of this study is to develop a predictive machine learning model for AC pavement wheel path cracking (WpCrAr) and the age at which cracking initiates (WpCrAr) as performance indicators. This study thoroughly investigated three ensemble machine learning models, including random forest, extremely randomized trees (ETR), and extreme gradient boosting (XGBoost). It was observed that XGBoost, optimized using Bayesian methods, emerged as the most effective among the evaluated models, demonstrating good predictive accuracy, with an R2 of 0.79 for WpCrAr and 0.92 for AgeCrack and mean absolute errors of 1.07 and 0.74, respectively. The most important features influencing crack initiation and progression were identified, including equivalent single axle load (ESAL), pavement age, number of layers, precipitation, and freeze–thaw cycles. This paper also showed the impact of pavement material combinations for base and subgrade layers on the delay of crack initiation. [ABSTRACT FROM AUTHOR]

The 3D body scan technology has recently innovated the way of measuring human bodies and generated a large volume of body measurements. However, one inherent issue that plagues the use of the resultant database is the missing data usually caused by using automatic data extractions from the 3D body scans. Tedious extra efforts have to be made to manually fill the missing data for various applications. To tackle this problem, this paper proposes a machine learning (ML)-based approach for 3D body measurement estimation while considering the measurement (feature) importance. The proposed approach selects the most critical features to reduce the algorithm input and to improve the ML method performance. In addition, a Bayesian search is further used in fine-tuning the hyperparameters to minimize the mean square error. Two distinct ML methods, i.e., Random Forest and XGBoost, are used and tested on a real-world dataset that contains 3D body scans of 212 participants in the Kansas-Missouri area of the United States. The results show the effectiveness of the proposed methods with roughly 3% of Mean Absolute Percentage Errors in estimating the missing data. The two ML methods with the proposed hybrid feature selection and the Baysian search are comprehensively compared. The comparative results suggest that the Random Forest method performs better than the XGBoost counterpart in filling missing 3D body measurements. [ABSTRACT FROM AUTHOR]

Annual turnover of home care workers represents a huge loss of revenue and is a key source of inefficiency in the home health care industry. In this article, we propose a data-driven approach to monitor employee churn and to capture the evolution of employee intent to leave. Unlike most papers in the literature, we use machine learning techniques to analyze over 2 million visits in the US, Canada, and Australia between 2016 and 2019. Results show that the gap between the number of hours worked and in the contract is the most important factor to predict employee intent to leave, which means an employee should be given as many hours as requested in the contract to improve retention. Secondary results show that having diverse shift lengths and continuity in services and patients seem to be associated with less turnover. [ABSTRACT FROM AUTHOR]

Mobile device location data (MDLD) contains abundant travel behavior information to support travel demand analysis. Compared to traditional travel surveys, MDLD has larger spatiotemporal coverage of the population and its mobility. However, ground truth information such as trip origins and destinations, travel modes, and trip purposes are not included by default. Such important attributes must be imputed to maximize the usefulness of the data. This paper targets at studying the capability of MDLD on estimating travel mode share at aggregated levels. A data-driven framework is proposed to extract travel behavior information from MDLD. The proposed framework first identifies trip ends with a modified Spatiotemporal Density-based Spatial Clustering of Applications with Noise algorithm. Then three types of features are extracted for each trip to impute travel modes using machine learning models. A labeled MDLD dataset with ground truth information is used to train the proposed models, resulting in a 95% recall rate in identifying trip ends and over 93% tenfold cross-validation accuracy in imputing the five travel modes (drive, rail, bus, bike and walk) with a random forest (RF) classifier. The proposed framework is then applied to two large-scale MDLD datasets, covering the Baltimore-Washington metropolitan area and the United States, respectively. The estimated trip distance, trip time, trip rate distribution, and travel mode share are compared against travel surveys at different geographies. The results suggest that the proposed framework can be readily applied in different states and metropolitan regions with low cost in order to study multimodal travel demand, understand mobility trends, and support decision making. [ABSTRACT FROM AUTHOR]

Two-vehicle crashes resulting from distracted driving led to a higher number of fatalities and serious injuries over time. This study utilized machine learning and econometric models to investigate two-vehicle-involved distracted driving crashes from the Crash Report Sampling System within the United States. XGBoost and Random Forest were utilized to identify the top variables based on SHAP value, although mixed logit with unobserved heterogeneity was used to model injury severity. The model results indicate that there is a complex interaction of driver characteristics, such as demographics (male drivers), driver actions (careless driving, driving more than the speed limit of more than 15 mph, hitting a stopped vehicle), a driver without violation history, turning violation, drinking, roadway characteristics (non-interstate highways, undivided and divided roadways with positive barrier, curved roadways, dry surface), environmental conditions (rainy weather), vehicle attributes (motorcycle, displacement volume up to 2500 cc, newer vehicle within five years of crash-involvement), temporal characteristics (4–6 PM, July–September, and year 2017). These findings underscore the importance of driving behavior and roadway design. As such, prioritizing efforts to address distracted driving behavior through driver training and law enforcement, as well as considering its implications for roadway design and maintenance, becomes crucial. [ABSTRACT FROM AUTHOR]

It has become crucial to have an early prediction model that provides accurate assurance for users about the financial situation of consumers. Recent studies have focused on predicting corporate bankruptcies and credit defaults, not personal bankruptcies. Due to this situation, the present study fills the literature gap by comparing different machine learning algorithms to predict personal bankruptcy. The main objective of the study is to examine the usefulness of machine learning models such as SVM, random forest, AdaBoost, XGBoost, LightGBM, and CatBoost in forecasting personal bankruptcy. The study relies on two samples of households (learning and testing) from the Survey of Consumer Finances, which was conducted in the United States. Among the models estimated, LightGBM, CatBoost, and XGBoost showed the highest effectiveness. The most important variables used in the models are income, refusal to grant credit, delays in the repayment of liabilities, the revolving debt ratio, and the housing debt ratio. [ABSTRACT FROM AUTHOR]

Regression algorithms are regularly used for improving the accuracy of satellite precipitation products. In this context, satellite precipitation and topography data are the predictor variables, and gauged-measured precipitation data are the dependent variables. Alongside this, it is increasingly recognised in many fields that combinations of algorithms through ensemble learning can lead to substantial predictive performance improvements. Still, a sufficient number of ensemble learners for improving the accuracy of satellite precipitation products and their large-scale comparison are currently missing from the literature. In this study, we work towards filling in this specific gap by proposing 11 new ensemble learners in the field and by extensively comparing them. We apply the ensemble learners to monthly data from the PERSIANN (Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks) and IMERG (Integrated Multi-satellitE Retrievals for GPM) gridded datasets that span over a 15-year period and over the entire contiguous United States (CONUS). We also use gauge-measured precipitation data from the Global Historical Climatology Network monthly database, version 2 (GHCNm). The ensemble learners combine the predictions of six machine learning regression algorithms (base learners), namely the multivariate adaptive regression splines (MARS), multivariate adaptive polynomial splines (poly-MARS), random forests (RF), gradient boosting machines (GBM), extreme gradient boosting (XGBoost) and Bayesian regularized neural networks (BRNN), and each of them is based on a different combiner. The combiners include the equal-weight combiner, the median combiner, two best learners and seven variants of a sophisticated stacking method. The latter stacks a regression algorithm on top of the base learners to combine their independent predictions. Its seven variants are defined by seven different regression algorithms, specifically the linear regression (LR) algorithm and the six algorithms also used as base learners. The results suggest that sophisticated stacking performs significantly better than the base learners, especially when applied using the LR algorithm. It also beats the simpler combination methods. [ABSTRACT FROM AUTHOR]

Background: The accurate encoding of operation notes is essential for activity‐based funding and workforce planning. The aim of this project was to evaluate the procedural coding accuracy of vitrectomy and to develop machine learning, natural language processing (NLP) models that may assist with this task. Methods: This retrospective cohort study involved vitrectomy operation notes between a 21‐month period at the Royal Adelaide Hospital. Coding of procedures were based on the Medicare Benefits Schedule (MBS)—the Australian equivalent to the Current Procedural Terminology (CPT®) codes used in the United States. Manual encoding was conducted for all procedures and reviewed by two vitreoretinal consultants. XGBoost, random forest and logistic regression models were developed for classification experiments. A cost‐based analysis was subsequently conducted. Results: There were a total of 1724 procedures with individual codes performed within 617 vitrectomy operation notes totalling $1 528 086.60 after manual review. A total of 1147 (66.5%) codes were missed in the original coding that amounted to $736 539.20 (48.2%). Our XGBoost model had the highest classification accuracy (94.6%) in the multi‐label classification for the five most common procedures. The XGBoost model was the most successful model in identifying operation notes with two or more missing codes with an AUC of 0.87 (95% CI 0.80–0.92). Conclusions: Machine learning has been successful in the classification of vitrectomy operation note encoding. We recommend a combined human and machine learning approach to clinical coding as automation may facilitate more accurate reimbursement and enable surgeons to prioritise higher quality clinical care. [ABSTRACT FROM AUTHOR]

Human injury in a vehicle crash is a critical subject of analysis. The injury severity of a person in crashes helps the transportation agency to determine crash conditions. This will in turn helps road safety manager or engineer to implement the counter measures and to improve and enhance the level of safety at the roadside. In this research study, the injury severity of a person in crash analysis across angles is analyzed using various machine learning models. For the injury severity prediction 2018 Fatality Analysis Reporting System (FARS) NHTSA (National Highway Traffic Safety Administration) dataset of United States is used. The person injury severity is predicted with the help of machine learning models like Multinomial logistic regression, Naive Bayes Classifier, Random Forest, Extra Trees, XGB Classifier and optimized XGBoost model. It is observed that optimized XGBoost method performs better than other machine learning models in terms of performance metrics like Accuracy, Error rate, Cohen-kappa-score, Loss and Misclassified samples. [ABSTRACT FROM AUTHOR]

Context: Predictive modeling is an integral part of broad-scale conservation efforts, and machine-learning (ML) models are increasingly being used for this purpose. But like all other predictive methods, ML models are susceptible to the problem of extrapolation. Objectives: Our objectives were to promote the quantification of spatial extrapolation risk associated with ML models, and to encourage application of this information to help assess the reliability of ML model predictions for conservation decisions. Methods: Using mean abundance and presence/absence data for Brewer's Sparrow (Spizella breweri)—a declining sagebrush obligate in the western United States—along with spatially optimized measures of landscape conditions, we demonstrated how to quantify the probability of spatial extrapolation by random forest and boosted regression tree models. Results: For both model types and both bird response variables, models used to predict extrapolation risk were highly accurate, implying that the approach we demonstrated can provide sound results for assessing the spatial extrapolation risk for new observations that were not involved in model training. Conclusions: Spatial extrapolation risk should be quantified and accounted for in decisions about whether to use ML model predictions in conservation efforts because these actions will help to avoid the use of unreliable predictions. [ABSTRACT FROM AUTHOR]

Background: Predicting survival of recipients after liver transplantation is regarded as one of the most important challenges in contemporary medicine. Hence, improving on current prediction models is of great interest.Nowadays, there is a strong discussion in the medical field about machine learning (ML) and whether it has greater potential than traditional regression models when dealing with complex data. Criticism to ML is related to unsuitable performance measures and lack of interpretability which is important for clinicians.Methods: In this paper, ML techniques such as random forests and neural networks are applied to large data of 62294 patients from the United States with 97 predictors selected on clinical/statistical grounds, over more than 600, to predict survival from transplantation. Of particular interest is also the identification of potential risk factors. A comparison is performed between 3 different Cox models (with all variables, backward selection and LASSO) and 3 machine learning techniques: a random survival forest and 2 partial logistic artificial neural networks (PLANNs). For PLANNs, novel extensions to their original specification are tested. Emphasis is given on the advantages and pitfalls of each method and on the interpretability of the ML techniques.Results: Well-established predictive measures are employed from the survival field (C-index, Brier score and Integrated Brier Score) and the strongest prognostic factors are identified for each model. Clinical endpoint is overall graft-survival defined as the time between transplantation and the date of graft-failure or death. The random survival forest shows slightly better predictive performance than Cox models based on the C-index. Neural networks show better performance than both Cox models and random survival forest based on the Integrated Brier Score at 10 years.Conclusion: In this work, it is shown that machine learning techniques can be a useful tool for both prediction and interpretation in the survival context. From the ML techniques examined here, PLANN with 1 hidden layer predicts survival probabilities the most accurately, being as calibrated as the Cox model with all variables.Trial Registration: Retrospective data were provided by the Scientific Registry of Transplant Recipients under Data Use Agreement number 9477 for analysis of risk factors after liver transplantation. [ABSTRACT FROM AUTHOR]

Aim: To develop and validate models based on logistic regression and artificial intelligence for prognostic prediction of molar survival in periodontally affected patients. Materials and Methods: Clinical and radiographic data from four different centres across four continents (two in Europe, one in the United States, and one in China) including 515 patients and 3157 molars were collected and used to train and test different types of machine‐learning algorithms for their prognostic ability of molar loss over 10 years. The following models were trained: logistic regression, support vector machine, K‐nearest neighbours, decision tree, random forest, artificial neural network, gradient boosting, and naive Bayes. In addition, different models were aggregated by means of the ensembled stacking method. The primary outcome of the study was related to the prediction of overall molar loss (MLO) in patients after active periodontal treatment. Results: The general performance in the external validation settings (aggregating three cohorts) revealed that the ensembled model, which combined neural network and logistic regression, showed the best performance among the different models for the prediction of MLO with an area under the curve (AUC) = 0.726. The neural network model showed the best AUC of 0.724 for the prediction of periodontitis‐related molar loss. In addition, the ensembled model showed the best calibration performance. Conclusions: Through a multi‐centre collaboration, both prognostic models for the prediction of molar loss were developed and externally validated. The ensembled model showed the best performance in terms of both discrimination and validation, and it is made freely available to clinicians for widespread use in clinical practice. [ABSTRACT FROM AUTHOR]

Background and Objective: Computer-based neurocognitive tests are widely used in sport-related concussion management, but the performance of these tests is not well understood in the participant population with attention-deficit/hyperactivity disorder (ADHD) and/or learning disorder (LD). This research estimates the sensitivity and specificity performance of the Immediate Post-Concussion Assessment and Cognitive Testing (ImPACT) computer-based neurocognitive test in identifying concussion in this population. Methods: Study participants consisted of collegiate university athletes and military service academy cadets from the National Collegiate Athletic Association-Department of Defense CARE Consortium who completed the ImPACT test between 2014 and 2021. Participants who self-identified as belonging to one of the subgroups of interest (ADHD with or without LD [ADHD:LD+/−], LD with or without ADHD [LD:ADHD+/−], ADHD and/or LD [ADHD a/o LD]) and completed a baseline (1874 ADHD:LD+/−, 779 LD:ADHD+/−, 2338 ADHD a/o LD) or 24–48 h post-concussion (175 ADHD:LD+/−, 77 LD:ADHD+/−, 216 ADHD a/o LD) ImPACT assessment were included. Sensitivity and specificity were calculated using a normative data method and three machine learning classification methods: logistic regression, classification and regression tree, and random forest. Results: Using the four methods, participants with ADHD:LD+/− had sensitivities that ranged from 0.474 to 0.697, and specificities that ranged from 0.538 to 0.686. Participants with LD:ADHD+/− had sensitivities that ranged from 0.455 to 0.688, and specificities that ranged from 0.456 to 0.588. For participants with ADHD a/o LD, sensitivities ranged from 0.542 to 0.755, and specificities ranged from 0.451 to 0.724. Conclusions: For all subgroups and analytical methods, the results illustrate sensitivity and specificity values below typically accepted levels indicative of clinical utility. These findings support that using ImPACT alone may be insufficient to inform concussion diagnoses and encourages the use of a multi-dimensional concussion assessment. [ABSTRACT FROM AUTHOR]

Gridded satellite precipitation datasets are useful in hydrological applications as they cover large regions with high density. However, they are not accurate in the sense that they do not agree with ground-based measurements. An established means for improving their accuracy is to correct them by adopting machine learning algorithms. This correction takes the form of a regression problem, in which the ground-based measurements have the role of the dependent variable and the satellite data are the predictor variables, together with topography factors (e.g., elevation). Most studies of this kind involve a limited number of machine learning algorithms and are conducted for a small region and for a limited time period. Thus, the results obtained through them are of local importance and do not provide more general guidance and best practices. To provide results that are generalizable and to contribute to the delivery of best practices, we here compare eight state-of-the-art machine learning algorithms in correcting satellite precipitation data for the entire contiguous United States and for a 15-year period. We use monthly data from the PERSIANN (Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks) gridded dataset, together with monthly earth-observed precipitation data from the Global Historical Climatology Network monthly database, version 2 (GHCNm). The results suggest that extreme gradient boosting (XGBoost) and random forests are the most accurate in terms of the squared error scoring function. The remaining algorithms can be ordered as follows, from the best to the worst: Bayesian regularized feed-forward neural networks, multivariate adaptive polynomial splines (poly-MARS), gradient boosting machines (gbm), multivariate adaptive regression splines (MARS), feed-forward neural networks and linear regression. [ABSTRACT FROM AUTHOR]

Merging satellite products and ground-based measurements is often required for obtaining precipitation datasets that simultaneously cover large regions with high density and are more accurate than pure satellite precipitation products. Machine and statistical learning regression algorithms are regularly utilized in this endeavor. At the same time, tree-based ensemble algorithms are adopted in various fields for solving regression problems with high accuracy and low computational costs. Still, information on which tree-based ensemble algorithm to select for correcting satellite precipitation products for the contiguous United States (US) at the daily time scale is missing from the literature. In this study, we worked towards filling this methodological gap by conducting an extensive comparison between three algorithms of the category of interest, specifically between random forests, gradient boosting machines (gbm) and extreme gradient boosting (XGBoost). We used daily data from the PERSIANN (Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks) and the IMERG (Integrated Multi-satellitE Retrievals for GPM) gridded datasets. We also used earth-observed precipitation data from the Global Historical Climatology Network daily (GHCNd) database. The experiments referred to the entire contiguous US and additionally included the application of the linear regression algorithm for benchmarking purposes. The results suggest that XGBoost is the best-performing tree-based ensemble algorithm among those compared. Indeed, the mean relative improvements that it provided with respect to linear regression (for the case that the latter algorithm was run with the same predictors as XGBoost) are equal to 52.66%, 56.26% and 64.55% (for three different predictor sets), while the respective values are 37.57%, 53.99% and 54.39% for random forests, and 34.72%, 47.99% and 62.61% for gbm. Lastly, the results suggest that IMERG is more useful than PERSIANN in the context investigated. [ABSTRACT FROM AUTHOR]

• More accurate and explanatory model is proposed for natural gas storage deliverability. • A hybrid Sparse Autoencoder and Random Forest model shows on average of 5.7% increase in prediction accuracy. • The first use of machine learning and explanatory methods (XAI) in deliverability prediction for natural gas storage. Numerous subsurface factors, including geology and fluid properties, can affect the connectivity of the storage spaces in depleted reservoirs; hence, fluid flow simulations become more complicated, and predicting their deliverability remains challenging. This paper applies Machine Learning (ML) techniques to predict the deliverability of underground natural gas storage (UNGS) in depleted reservoirs. First, three baseline models were developed based on Support Vector Regression (SVR), Artificial Neural Network (ANN), and Random Forest (RF) algorithms. To improve the accuracy of the RF model as the best-performing baseline model, a unified framework, referred to as SARF, was developed. SARF combines the capabilities of Sparse Autoencoder (SA) and that of Random Forest (RF). To achieve this, the internal representations of the SA, which constitute extracted features of the input variables, are used in RF to develop the proposed SARF framework. The predictive capabilities of the baseline models and the proposed SARF model were validated using 3744 real-world storage data samples of 52 active storage reservoirs in the United States. The experimental result of this study shows that the proposed SARF model achieved an average 5.7% increase in accuracy on four separate data partitions over the baseline RF model. Furthermore, a set of eXplainable Artificial Intelligence (XAI) methods were developed to provide an intuitive explanation of which factors influence the deliverability of reservoir storage. The visualizations developed using the XAI method provide an easy-to-understand interpretation of how the SARF model predicted the deliverability values for separate reservoirs. [ABSTRACT FROM AUTHOR]

Using a machine-learning technique known as random forests, we analyze the role of investor confidence in forecasting monthly aggregate realized stock-market volatility of the United States (US), over and above a wide-array of macroeconomic and financial variables. We estimate random forests on data for a period from 2001 to 2020, and study horizons up to one year by computing forecasts for recursive and a rolling estimation window. We find that investor confidence, and especially investor confidence uncertainty has out-of-sample predictive value for overall realized volatility, as well as its "good" and "bad" variants. Our results have important implications for investors and policymakers. [ABSTRACT FROM AUTHOR]

Background: The current epidemic of opioid overdoses in the United States necessitates a robust public health and clinical response. We described patterns of non-fatal opioid overdoses (NFOODs) in a small western region using data from the 9-1-1 Computer Assisted Dispatch (CAD) record and electronic Patient Clinical Records (ePCR) completed by EMS responders. We determined whether CAD and ePCR variables could identify NFOOD cases in 9-1-1 data for intervention and surveillance efforts. Methods: We conducted a retrospective analysis of 1 year of 9-1-1 emergency medical CAD and ePCR (including naloxone administration) data from the sole EMS provider in the response area. Cases were identified based on clinician review of the ePCR, and categorized as definitive NFOOD, probable NFOOD, or non-OOD. Sensitivity, specificity, positive and negative predictive values (PPV and NPV) of the most prevalent CAD and ePCR variables were calculated. We used a machine learning technique—Random-Forests (RF) modeling—to optimize our ability to accurately predict NFOOD cases within census blocks. Results: Of 37,960 9-1-1 calls, clinical review identified 158 NFOOD cases (0.4%), of which 123 (77.8%) were definitive and 35 (22.2%) were probable cases. Overall, 106 (67.1%) received naloxone from the EMS responder at the scene. As a predictor of NFOOD, naloxone administration by paramedics had 67.1% sensitivity, 99.6% specificity, 44% PPV, and 99.9% NPV. Using CAD variables alone achieved a sensitivity of 36.7% and specificity of 99.7%. Combining ePCR variables with CAD variables increased the diagnostic accuracy with the best RF model yielding 75.9% sensitivity, 99.9% specificity, 71.4% PPV, and 99.9% NPV. Conclusion: CAD problem type variables and naloxone administration, used alone or in combination, had sub-optimal predictive accuracy. However, a Random Forests modeling approach improved accuracy of identification, which could foster improved surveillance and intervention efforts. We identified the set of NFOODs that EMS encountered in a year and may be useful for future surveillance efforts. [ABSTRACT FROM AUTHOR]

The United States' blood supply chain is experiencing market decline due to recent innovations in surgical practice, transfusion management, and hospital policy. These innovations strain US blood centers, resulting in cuts to surge capacities, consolidation, and reduced funding for research and outreach programs. In this study, we use data from a regional blood center to explore the application of contemporary machine learning algorithms for modeling donor retention. Such predictive models of donor retention can be used to design more cost effective donor outreach programs. Using data from a large US blood center paired with random forest classifiers, we are able to build a model of donor retention with a Mathews correlation of coefficient of 0.851. [ABSTRACT FROM AUTHOR]

Early childhood is an important development period for establishing healthy physical activity (PA) habits. The objective of this study was to evaluate PA levels in a representative sample of U.S. preschool-aged children. The study sample included 301 participants (149 girls, 3–5 years of age) in the 2012 U.S. National Health and Examination Survey National Youth Fitness Survey. Participants were asked to wear an ActiGraph accelerometer on their wrist for 7 days. A machine learning random forest classification algorithm was applied to accelerometer data to estimate daily time spent in moderate- and vigorous-intensity PA (MVPA; the sum of minutes spent in running, walking, and other moderate- and vigorous-intensity PA) and total PA (the sum of MVPA and light-intensity PA). We estimated that U.S. preschool-aged children engaged in 28 min/day of MVPA and 361 min/day of total PA, on average. MVPA and total PA levels were not significantly different between males and females. This study revealed that U.S. preschool-aged children engage in lower levels of MVPA and higher levels of total PA than the minimum recommended by the World Health Organization. [ABSTRACT FROM AUTHOR]

Natural streams longitudinal dispersion coefficient (Kx) is an essential indicator for pollutants transport and its determination is very important. Kx is influenced by several parameters, including river hydraulic geometry, sediment properties, and other morphological characteristics, and thus its calculation is a highly complex engineering problem. In this research, three relatively explored machine learning (ML) models, including Random Forest (RF), Gradient Boosting Decision Tree (GTB), and XGboost-Grid, were proposed for the Kx determination. The modeling scheme on building the prediction matrix was adopted from the well-established literature. Several input combinations were tested for better predictability performance for the Kx. The modeling performance was tested based on the data division for the training and testing (70–30% and 80–20%). Based on the attained modeling results, XGboost-Grid reported the best prediction results over the training and testing phase compared to RF and GTB models. The development of the newly established machine learning model revealed an excellent computed-aided technology for the Kx simulation. [ABSTRACT FROM AUTHOR]

The United States NRCS has a soil database that has data collected from across the country over the last several decades. This also includes soil spectral scans. This data is available, but it may not be used to its full potential. For this study, pedon, horizon and spectral data was extracted from the database for samples collected from 2011 to 2015. Only sites that had been fully described and horizons that had been analyzed for the full suite of desired properties were used. This resulted in over 14,000 samples that were used for modeling and eight soil properties: soil organic carbon (SOC); total nitrogen (TN); total sulfur (TS); clay; sand; exchangeable calcium (Caex); cation exchange capacity (CEC); and pH. Four chemometric methods were employed for soil property prediction: partial least squares (PLSR); Random Forest (RF); Cubist; and multivariable adaptive regression splines (MARS). The dataset was sufficiently large that only random subsetting was used to create calibration (70%) and validation (30%) sets. SOC, TN, and TS had the strongest prediction results, with an R2 value of over 0.9. Caex, CEC and pH were predicted moderately well. Clay and sand models had slightly lower performance. Of the four methods, Cubist produced the strongest models, while PLSR produced the weakest. This may be due to the complex relationships between soil properties and spectra that PLSR could not capture. The only drawback of Cubist is the difficult interpretability of variable importance. Future research should include the use of environmental variables to improve prediction results. Future work may also avoid the use of PLSR when dealing with large datasets that cover large areas and have high degrees of variability. [ABSTRACT FROM AUTHOR]

We analyze the role of macroeconomic uncertainty in predicting synchronization in housing price movements across all the United States (US) states plus District of Columbia (DC). We first use a Bayesian dynamic factor model to decompose the house price movements into a national, four regional (Northeast, South, Midwest, and West), and state-specific factors. We then study the ability of macroeconomic uncertainty in forecasting the comovements in housing prices, by controlling for a wide-array of predictors, such as factors derived from a large macroeconomic dataset, oil shocks, and financial market-related uncertainties. To accommodate for multiple predictors and nonlinearities, we take a machine learning approach of random forests. Our results provide strong evidence of forecastability of the national house price factor based on the information content of macroeconomic uncertainties over and above the other predictors. This result also carries over, albeit by a varying degree, to the factors associated with the four census regions, and the overall house price growth of the US economy. Moreover, macroeconomic uncertainty is found to have predictive content for (stochastic) volatility of the national factor and aggregate US house price. Our results have important implications for policymakers and investors. [ABSTRACT FROM AUTHOR]

Artificial levees are anthropogenic structures designed to hydrologically disconnect rivers from floodplains. The extent of artificial levees in the contiguous United States (CONUS) is unknown. To better estimate the distribution of artificial levees, we tested several different geomorphic, land cover, and spatial variables developed from the National Elevation Dataset, the National Land Cover database, and the National Hydrology Dataset HR Plus. We used known levee locations from the National Levee Database as training data. We tested machine learning and general logistic models' ability to detect artificial levees in a 100‐year hydrogeomorphic floodplain of seven geographically diverse 8‐digit HUC basins. Random forest models outperformed other models in predicting the location of levees using variables representing geomorphic attributes, land cover, and distance from streams ranging in size between stream order one through six. To demonstrate the ability of our approach to detect unknown levees, we conducted a leave‐one‐out cross‐validation in the lower Mississippi Basin using approximately 1,100 artificial levees. This approach detected known levees constituting 94% of the total levee length in the basin. Scaling up to the CONUS, we applied a high performing (overall accuracy of 97%) random forest model using land cover and stream order variables. We detected 182,213 km of potential levees, mostly along streams of order 2–6 in the Mississippi and Missouri River Basins, indicating that the national levee database contains 20.4% of levee length. Potential levees and those documented in the national levee database modify 2% of the total length of streams in the contiguous United States. Plain Language Summary: There has been a lot of research exploring how humans have impacted river systems to include impacts from dams and roads at national and global scales. However, the study of artificial levees has been limited to local studies due to incomplete databases of levee location. Artificial levees are linear dirt mounds, built next to rivers to stop flooding of property built on floodplains. Because the location of every artificial levee is not known in the US, we do not know to what extent floodplains have been separated from their rivers. To ameliorate this, our study explores different methods to detect unknown levees and applies an algorithm that is 97% accurate to the contiguous U.S. Most levees are built within floodplains, so we limited our study to the 100‐year floodplain. We tested different types of algorithms in a limited study with variables explaining levee shape, human land cover, and distance from streams of different sizes. Then, we applied the most effective models to the contiguous U.S. We located over 182,000 km of potential levees, potential areas with artificial levees not identified in the NLD, most of which were built on smaller streams and concentrated in the Mississippi and Missouri River basins. Key Points: An incomplete database of artificial levees inhibits the study of artificial levee impacts to floodplain extent in the contiguous USADifferent methods of detecting artificial levees are tested in case study basins and in the contiguous USAThis study detects 182,000 km of potential levees, along smaller streams in the Mississippi and Missouri Basins [ABSTRACT FROM AUTHOR]

Machine learning (ML) interpretability has become increasingly crucial for identifying accurate and relevant structural relationships between spatial events and factors that explain them. Methodologically aspatial ML algorithms with an apparent high predictive power ignore non-stationary domain relationships in spatio-temporal data (e.g. dependence, heterogeneity), leading to incorrect interpretations and poor management decisions. This study addresses this critical methodological issue of 'interpretability' in ML-based modeling of structural relationships using the example of heterogeneous drivers of wildfires across the United States. Specifically, we present and evaluate a spatio-temporally interpretable random forest (iST-RF) that uses spatio-temporal sampling-based training and weighted prediction. Although the ultimate scientific objective is to derive interpretation in space-time, experiments show that iST-RF can improve predictive accuracy (76%) compared to the aspatial RF approach (70%) while enhancing interpretations of the trained model's spatio-temporal relevance for its ensemble prediction. This novel approach can help balance prediction and interpretation with fidelity in a spatial data science life cycle. However, challenges exist for predictive modeling when the dataset is very small because in such cases locally optimized sub-model's prediction performance can be suboptimal. With that caveat, our proposed approach is an ideal choice for identifying drivers of spatio-temporal events at country- or regional-scale studies. [ABSTRACT FROM AUTHOR]

[The Geostationary Lightning Mapper (GLM) detects total lightning continuously from space, and does not distinguish intra‐cloud (IC) from cloud‐to‐ground (CG) lightning. This research focuses on differentiating CG and IC lightning detected by GLM using a random forests (RF) model. GLM flash and group characteristics are implemented into the RF model and used to predict flash type. The most important flash characteristic for distinguishing flash type is the maximum group area. Other features with high feature importance include time‐of‐day, elongation, propagation, footprint, slope, maximum distance between groups and events, and mean energy. Skill scores showcase the model's ability to distinguish flash type with moderate skill, with 81% probability of detection, 71% percent correct, 36% false alarm rate, 36% false alarm ratio, and 56% critical success index. These scores improve further when study area is limited to CONUS and the Southeastern United States. These results can be used to aid in future climatological analysis of flash type.] Plain Language Summary: The Geostationary Lightning Mapper (GLM) is a satellite‐based lightning sensor that allows for continuous detection of light emitted from the cloud tops due to lightning. GLM does not distinguish if the lightning is connecting to ground (CG) or remaining in the cloud (IC). In order to distinguish CG and IC flashes, this research uses a machine learning method called random forests (RF). The RF model attempts to classify lightning flashes based on their size, duration, and intensity. The most important flash characteristics for distinguishing flash type are the features related to the areal size of the lightning and the time of day the lightning occurs. Overall, moderate success is shown when attempting to divide total lightning into CG and IC over the 2018 period. This information can be used by researchers to improve the use of GLM in the study of different storm types as well as aiding in wildfire forecasting. Key Points: Although there is substantial overlap in the distributions for cloud‐to‐ground (CG) and intra‐cloud (IC) flash characteristics, the mean values differMost important features for flash distinction are maximum group area, time‐of‐day, elongation, propagation, footprint, slope, and max group/event distanceResults show skill in the random forests model's ability to distinguish CG and IC flashes in the Geostationary Lightning Mapper data [ABSTRACT FROM AUTHOR]

Approximately seven years of daily initializations from the convection-allowing National Severe Storms Laboratory Weather Research and Forecasting Model are used as inputs to train random forest (RF) machine learning models to probabilistically predict instances of excessive rainfall. Unlike other hazards, excessive rainfall does not have an accepted definition, so multiple definitions of excessive rainfall and flash flooding—including flash flood reports and 24-h average recurrence intervals (ARIs)—are used to explore RF configuration forecast sensitivities. RF forecasts are analogous to operational Weather Prediction Center (WPC) day-1 Excessive Rainfall Outlooks (EROs) and their resolution, reliability, and skill are strongly influenced by rainfall definitions and how inputs are assembled for training. Models trained with 1-yr ARI exceedances defined by the Stage-IV (ST4) precipitation analysis perform poorly in the northern Great Plains and Southwest United States, in part due to a high bias in the number of training events in these regions. Increasing the ARI threshold to 2 years or removing ST4 data from training, optimizing forecast skill geographically, and spatially averaging meteorological inputs for training generally results in improved CONUS-wide RF forecast skill. Both EROs and RF forecasts have seasonal skill—–poor forecasts in the late fall and winter and skillful forecasts in the summer and early fall. However, the EROs are consistently and significantly better than their RF counterparts, regardless of RF configuration, particularly in the summer months. The results suggest careful consideration should be made when developing ML-based probabilistic precipitation forecasts with convection-allowing model inputs, and further development is necessary to consider these forecast products for operational implementation. Significance Statement: Machine learning (ML) models can deduce statistical relationships between a set of predictors and meteorological events. In this work, ML models are developed to predict excessive rainfall events. Since excessive rainfall is difficult to uniformly define across the United States, multiple ML models are built from a variety of rainfall datasets with predictors gathered from output of a high-resolution numerical weather prediction model and forecasts are made from each model. Forecasts made from these models are highly sensitive to both the definitions of excessive rainfall (e.g., 100 mm of rain in a day may cause flooding in a usually dry area, but not in a wet area) and the predictors used. Forecast skill can increase when excessive rainfall events are rarer and when predictors synthesize the surrounding environment rather than characterize specific geographical points. ML-based models have great potential for excessive rainfall prediction, but careful attention to the configuration of these models is required. [ABSTRACT FROM AUTHOR]

Foliar fungicide usage in soybeans in the north-central United States increased steadily over the past two decades. An agronomically-interpretable machine learning framework was used to understand the importance of foliar fungicides relative to other factors associated with realized soybean yields, as reported by growers surveyed from 2014 to 2016. A database of 2738 spatially referenced fields (of which 30% had been sprayed with foliar fungicides) was fit to a random forest model explaining soybean yield. Latitude (a proxy for unmeasured agronomic factors) and sowing date were the two most important factors associated with yield. Foliar fungicides ranked 7th out of 20 factors in terms of relative importance. Pairwise interactions between latitude, sowing date and foliar fungicide use indicated more yield benefit to using foliar fungicides in late-planted fields and in lower latitudes. There was a greater yield response to foliar fungicides in higher-yield environments, but less than a 100 kg/ha yield penalty for not using foliar fungicides in such environments. Except in a few production environments, yield gains due to foliar fungicides sufficiently offset the associated costs of the intervention when soybean prices are near-to-above average but do not negate the importance of disease scouting and fungicide resistance management. [ABSTRACT FROM AUTHOR]

Background: Cardiac dysrhythmias (CD) affect millions of Americans in the United States (US), and are associated with considerable morbidity and mortality. New strategies to combat this growing problem are urgently needed. Objectives: Predicting CD using electronic health record (EHR) data would allow for earlier diagnosis and treatment of the condition, thus improving overall cardiovascular outcomes. The Guideline Advantage (TGA) is an American Heart Association ambulatory quality clinical data registry of EHR data representing 70 clinics distributed throughout the US, and has been used to monitor outpatient prevention and disease management outcome measures across populations and for longitudinal research on the impact of preventative care. Methods: For this study, we represented all time-series cardiovascular health (CVH) measures and the corresponding data collection time points for each patient by numerical embedding vectors. We then employed a deep learning technique–long-short term memory (LSTM) model–to predict CD from the vector of time-series CVH measures by 5-fold cross validation and compared the performance of this model to the results of deep neural networks, logistic regression, random forest, and Naïve Bayes models. Results: We demonstrated that the LSTM model outperformed other traditional machine learning models and achieved the best prediction performance as measured by the average area under the receiver operator curve (AUROC): 0.76 for LSTM, 0.71 for deep neural networks, 0.66 for logistic regression, 0.67 for random forest, and 0.59 for Naïve Bayes. The most influential feature from the LSTM model were blood pressure. Conclusions: These findings may be used to prevent CD in the outpatient setting by encouraging appropriate surveillance and management of CVH. [ABSTRACT FROM AUTHOR]

Excessive rainfall is difficult to forecast, and there is a need for tools to aid Weather Prediction Center (WPC) forecasters when generating Excessive Rainfall Outlooks (EROs), which are issued for the contiguous United States at lead times of 1–3 days. To address this need, a probabilistic forecast system for excessive rainfall, known as the Colorado State University Machine Learning Probabilities (CSU-MLP) system, was developed based on ensemble reforecasts, precipitation observations, and machine-learning algorithms, specifically random forests. The CSU-MLP forecasts were designed to emulate the EROs, with the goal being a tool that forecasters can use as a "first guess" in the ERO forecast process. Resulting from close collaboration between CSU and WPC and evaluation at the Flash Flood and Intense Rainfall Experiment, iterative improvements were made to the forecast system and it was transitioned into operational use at WPC. Quantitative evaluation shows that the CSU-MLP forecasts are skillful and reliable, and they are now being used as a part of the WPC forecast process. This project represents an example of a successful research-to-operations transition, and highlights the potential for machine learning and other postprocessing techniques to improve operational predictions. [ABSTRACT FROM AUTHOR]

The objective of the current study was to build predictive models for suicidal ideation in a sample of children aged 9–10 using features previously implicated in risk among older adolescent and adult populations. This case-control analysis utilized baseline data from the Adolescent Brain and Cognitive Development (ABCD) Study, collected from 21 research sites across the United States (N = 11,369). Several regression and ensemble learning models were compared on their ability to classify individuals with suicidal ideation and/or attempt from healthy controls, as assessed by the Kiddie Schedule for Affective Disorders and Schizophrenia–Present and Lifetime Version. When comparing control participants (mean age: 9.92±0.62 years; 4944 girls [49%]) to participants with suicidal ideation (mean age: 9.89±0.63 years; 451 girls [40%]), both logistic regression with feature selection and elastic net without feature selection predicted suicidal ideation with an AUC of 0.70 (CI 95%: 0.70–0.71). The random forest with feature selection trained to predict suicidal ideation predicted a holdout set of children with a history of suicidal ideation and attempt (mean age: 9.96±0.62 years; 79 girls [41%]) from controls with an AUC of 0.77 (CI 95%: 0.76–0.77). Important features from these models included feelings of loneliness and worthlessness, impulsivity, prodromal psychosis symptoms, and behavioral problems. This investigation provided an unprecedented opportunity to identify suicide risk in youth. The use of machine learning to examine a large number of predictors spanning a variety of domains provides novel insight into transdiagnostic factors important for risk classification. [ABSTRACT FROM AUTHOR]

A 30 × 30m-resolution gridded dataset of forest plot identifiers was developed for the conterminous United States (CONUS) using a random forests machine-learning imputation approach. Forest plots from the US Forest Service Forest Inventory and Analysis program (FIA) were imputed to gridded c2014 landscape data provided by the LANDFIRE project using topographic, biophysical, and disturbance variables. The output consisted of a raster map of plot identifiers. From the plot identifiers, users of the dataset can link to a number of tree- and plot-level attributes stored in the accompanying tables and in the publicly available FIA DataMart, and then produce maps of any of these attributes, including number of trees per acre, tree species, and forest type. Of 67,141 FIA plots available, 62,758 of these (93.5%) were utilized at least once in imputation to 2,841,601,981 forested pixels in CONUS. Continuous high-resolution forest structure data at a national scale will be invaluable for analyzing carbon dynamics, habitat distributions, and fire effects. Measurement(s) forested area Technology Type(s) machine learning Factor Type(s) biophysical variables • geographic location • forest type • forest height • forest percent cover Sample Characteristic - Environment vegetation layer Sample Characteristic - Location United States of America Machine-accessible metadata file describing the reported data: https://doi.org/10.6084/m9.figshare.13285904 [ABSTRACT FROM AUTHOR]

Key Points: Groundwater withdrawals are not actively monitored in most places of the world at a scale necessary to implement sustainable solutionsVarious multitemporal remote sensing data are integrated into a machine learning framework to effectively predict groundwater withdrawalsThe results over the High Plains Aquifer, Kansas, USA, show that this approach is applicable to similar regions having sparse in situ data Effective monitoring of groundwater withdrawals is necessary to help mitigate the negative impacts of aquifer depletion. In this study, we develop a holistic approach that combines water balance components with a machine learning model to estimate groundwater withdrawals. We use both multitemporal satellite and modeled data from sensors that measure different components of the water balance and land use at varying spatial and temporal resolutions. These remote sensing products include evapotranspiration, precipitation, and land cover. Due to the inherent complexity of integrating these data sets and subsequently relating them to groundwater withdrawals using physical models, we apply random forests—a state of the art machine learning algorithm—to overcome such limitations. Here, we predict groundwater withdrawals per unit area over a highly monitored portion of the High Plains aquifer in the central United States at 5 km resolution for the Years 2002–2019. Our modeled withdrawals had high accuracy on both training and testing data sets (R2 ≈ 0.99 and R2 ≈ 0.93, respectively) during leave‐one‐out (year) cross validation with low mean absolute error (MAE) ≈ 4.31 mm and root‐mean‐square error (RMSE) ≈ 13.50 mm for the year 2014. Moreover, we found that even for the extreme drought year of 2012, we have a satisfactory test score (R2 ≈ 0.84) with MAE ≈ 9.72 mm and RMSE ≈ 24.17 mm. Therefore, the proposed machine learning approach should be applicable to similar regions for proactive water management practices. Plain Language Summary: Groundwater is an essential component of global water resources and is the largest source of Earth's liquid freshwater. It is extensively used for drinking water and to support global food production. Consequently, groundwater consumption has significantly increased owing to the pressing demands for water, food, and energy primarily driven by the increasing global population. Despite its critical role in the water‐food‐energy nexus, very few regions in the United States or elsewhere actively monitor their groundwater withdrawals (extraction or pumping) for implementing sustainable water management solutions. We develop a novel approach combining water balance components and land use measured using openly available remote sensing products with a machine learning model to estimate groundwater withdrawals. This framework automatically learns the interrelationships among these variables and groundwater withdrawals. Our study area is a portion of the High Plains Aquifer in Kansas (central United States) where overpumping has caused substantial groundwater storage loss. Also, a large amount of groundwater pumping data are available for validation. The results indicate good accuracy even for extreme drought years. Thus, this approach should be applicable to similar regions having sparsely or moderately available groundwater pumping data, enabling water managers to proactively implement sustainable solutions addressing water security issues. [ABSTRACT FROM AUTHOR]

Rising admissions from emergency departments (EDs) to hospitals are a primary concern for many healthcare systems. The issue of how to differentiate urgent admissions from non-urgent or even elective admissions is crucial. We aim to develop a model for classifying inpatient admissions based on a patient's primary diagnosis as either emergency care or elective care and predicting urgency as a numerical value. We use supervised machine learning techniques and train the model with physician-expert judgments. Our model is accurate (96%) and has a high area under the ROC curve (>.99). We provide the first comprehensive classification and urgency categorization for inpatient emergency and elective care. This model assigns urgency values to every relevant diagnosis in the ICD catalog, and these values are easily applicable to existing hospital data. Our findings may provide a basis for policy makers to create incentives for hospitals to reduce the number of inappropriate ED admissions. [ABSTRACT FROM AUTHOR]

The abundance of type and quantity of available data in the healthcare field has led many to utilize machine learning approaches to keep up with this influx of data. Data pertaining to COVID-19 is an area of recent interest. The widespread influence of the virus across the United States creates an obvious need to identify groups of individuals that are at an increased risk of mortality from the virus. We propose a so-called clustered random forest approach to predict COVID-19 patient mortality. We use this approach to examine the hidden heterogeneity of patient frailty by examining demographic information for COVID-19 patients. We find that our clustered random forest approach attains predictive performance comparable to other published methods. We also find that follow-up analysis with neural network modeling and k-means clustering provide insight into the type and magnitude of mortality risks associated with COVID-19. [ABSTRACT FROM AUTHOR]

Slope failure probabilistic models generated using random forest (RF) machine learning (ML), manually interpreted incident points, and light detection and ranging (LiDAR) digital terrain variables are assessed for predicting and generalizing to new geographic extents. Specifically, models for four Major Land Resource Areas (MLRAs) in the state of West Virginia in the United States (US) were created. All region-specific models were then used to predict withheld validation data within all four MLRAs. For all validation datasets, the model trained using data from the same MLRA provided the highest reported overall accuracy (OA), Kappa statistic, F1 Score, area under the receiver operating characteristic curve (AUC ROC), and area under the precision-recall curve (AUC PR). However, the model from the same MLRA as the validation dataset did not always provide the highest precision, recall, and/or specificity, suggesting that models extrapolated to new geographic extents tend to either overpredict or underpredict the land area of slope failure occurrence whereas they offer a better balance between omission and commission error within the region in which they were trained. This study highlights the value of developing region-specific inventories, models, and high resolution and detailed digital elevation data, since models may not generalize well to new geographic extents, potentially resulting from spatial heterogeneity in landscape and/or slope failure characteristics. [ABSTRACT FROM AUTHOR]

A task for environmental health research is to produce complete pollution exposure maps despite limited monitoring data. Satellite-derived aerosol optical depth (AOD) is frequently used as a predictor in various models to improve PM2.5 estimation, despite significant gaps in coverage. We analyze PM2.5 and AOD from July 2011 in the contiguous United States. We examine two methods to aid in gap-filling AOD: (1) lattice kriging, a spatial statistical method adapted to handle large amounts data, and (2) random forest, a tree-based machine learning method. First, we evaluate each model's performance in the spatial prediction of AOD, and we additionally consider ensemble methods for combining the predictors. In order to accurately assess the predictive performance of these methods, we construct spatially clustered holdouts to mimic the observed patterns of missing data. Finally, we assess whether gap-filling AOD through one of the proposed ensemble methods can improve prediction of PM2.5 in a random forest model. Our results suggest that ensemble methods of combining lattice kriging and random forest can improve AOD gap-filling. Based on summary metrics of performance, PM2.5 predictions based on random forest models were largely similar regardless of the inclusion of gap-filled AOD, but there was some variability in daily model predictions. [ABSTRACT FROM AUTHOR]