Search NASA⌕ Search

SEARCH · Search NASA

Results for “Random forests”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 109 records · Page 6

Search for associated production of z and Higgs bosons in proton-antiproton collisions at 1.96 TeV

We present a search for associated production of Z and Higgs bosons in 4.2 fb -1 of $\bar{p}$p collisions at √s = 1.96 TeV, produced in RunII of the Tevatron and recorded by the D0 detector. The search is performed in events containing at least two muons and at least two jets. The ZH signal is distinguished from the expected backgrounds by means of multivariate classifiers known as random forests. Binned random forest output distributions are used in comparing the data to background-only and signal+background hypotheses. No excess is observed in the data, so we set upper limits on ZH production with a 95% confidence level.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗

Southwest Pacific tropical cyclone development classification utilizing machine learning and synoptic composites

This study evaluates the ability of machine learning algorithms to classify tropical depressions (TDs) and tropical storms (TSs) in the western region of the southwest Pacific Ocean (SWPO). Decision rules are generated to predict the environment required for a depression to fully develop into a mature storm, and the most influential predictors in the classification decision are ranked. TD and TS are discriminated based on a maximum sustained wind speed threshold (≥17 ms -1 ). Various aerosol, thermodynamic, and dynamic parameters are extracted closest to the initiation point of each non-developing and developing sample. The covariates associated with each labelled sample are used to train a decision tree and random forest model. Results using a testing dataset suggest the random forest approach more accurately distinguishes between non-developing and developing samples. The classification accuracy of the decision tree and random forest are 72% and 91%, respectively. Random forest outperformed the decision tree by providing higher accuracy in test data. The most important variables for binary classification are sea salt aerosol optical depth (AOD), 1,000 mb relative humidity, and sea surface temperature. AOD is a quantitative estimate of the aerosols presents in the air through the extinction of a ray of light as it passes through the atmosphere. Mean composite maps constructed in an unsupervised manner have been created for the most important variables identified by the random forest classifier during TD and TS events to highlight the difference in geophysical and aerosol variables' climatology during the two different classifications. This work will advance the risk management strategies for northeastern Australia and other SWPO basin islands to control their tropical cyclone related losses through prioritizing forecasting variables that are the strongest predictors of the strengthening of tropical depressions into tropical cyclones.

54 ENVIRONMENTAL SCIENCES↗

Applied Machine-Learning Models to Identify Spectral Sub-Types of M Dwarfs from Photometric Surveys

M dwarfs are the most abundant stars in the Solar Neighborhood and they are prime targets for searching for rocky planets in habitable zones. Consequently, a detailed characterization of these stars is in demand. The spectral sub-type is one of the parameters that is used for the characterization and it is traditionally derived from the observed spectra. However, obtaining the spectra of M dwarfs is expensive in terms of observation time and resources due to their intrinsic faintness. We study the performance of four machine-learning (ML) models—K-Nearest Neighbor (KNN), Random Forest (RF), Probabilistic Random Forest (PRF), and Multilayer Perceptron (MLP)—in identifying the spectral sub-types of M dwarfs at a grand scale by deploying broadband photometry in the optical and near-infrared. We trained the ML models by using the spectroscopically identified M dwarfs from the Sloan Digital Sky Survey (SDSS) Data Release (DR) 7, together with their photometric colors that were derived from the SDSS, Two-Micron All-Sky Survey, and Wide-field Infrared Survey Explorer. We found that the RF, PRF, and MLP give a comparable prediction accuracy, 74%, while the KNN provides slightly lower accuracy, 71%. We also found that these models can predict the spectral sub-type of M dwarfs with ~99% accuracy within ±1 sub-type. The five most useful features for the prediction are r - z, r - i, r - J, r - H , and g - z, and hence lacking data in all SDSS bands substantially reduces the prediction accuracy. However, we can achieve an accuracy of over 70% when the r and i magnitudes are available. Since the stars in this study are nearby (d ≲ 1300 pc for 95% of the stars), the dust extinction can reduce the prediction accuracy by only 3%. Finally, we used our optimized RF models to predict the spectral sub-types of M dwarfs from the Catalog of Cool Dwarf Targets for the Transiting Exoplanet Survey Satellite, and we provide the optimized RF models for public use.

79 ASTRONOMY AND ASTROPHYSICS↗

Label Assist: Personalized Travel Models for Longitudinal Data Collection

Understanding travel behavior is crucial to transportation decarbonization. OpenPATH is an open-source mobility platform which collects and analyzes human travel behavior at the individual level. The mobile application passively senses trips and prompts users to label them. However, users find the labeling process burdensome; less than half the trips are typically labeled, making much of the data unusable in aggregate analyses of mobility patterns. Prior work has addressed the response fatigue challenge through automated mode inference using sensor data, but sensors cannot capture all aspects of travel behavior. We explore an alternative approach in which we leverage prior user input to predict travel choices in novel trips. We first explore trip clustering methods and develop a novel two-step pipeline using DBSCAN and SVMs to extract realistic geospatial clusters. We then propose two strategies to predict trip labels: (i) clustering trips and extrapolating labels for similar trips, and (ii) random forest classification. The random forest approach is able to achieve - $70-80% accuracy (purpose: 72%, mode: 79%, replaced mode: 81%). These novel approaches to trip classification allow us to increase the rate of user labeling by suggesting predicted labels to be verified by the user. Unlabeled trips can also contribute to aggregate analyses, using label predictions and their associated confidences as a substitute. While there exist other travel survey apps with the ability to infer travel choices, to our knowledge, this is the first paper to describe such a supervised system and rigorously evaluate it.

ADVANCED PROPULSION SYSTEMS,ENERGY PLANNING, POLIC↗

Weather and Random Forest-based Load Profiling Approximation Models and its Transferability across Climate Zones

This study is to provide predictive understanding of the associations of various weather attributes with residential and commercial load profiles, for a variety of climate zones and seasons. In this work, machine learning (ML) approaches were used to identify and quantify the impacts of various weather attributes on residential and commercial electricity demand and its components across the western United States. Performance and transferability of the developed ML models were then evaluated across different temperate zones (e.g., southern, middle, and northern US) and across coastal, mid-continent, and wet zones, with inputs of weather condition data from the National Oceanic and Atmospheric Administration (NOAA) at representative weather stations. The predictive models were developed based on the ranked/screened factors using the regression tree (RT) and random forest (RF) approaches, for five different scenarios (seasons).

load composite, Random Forest, regression tree, lo↗

How Should Machine Learning Be Successfully Used for Wind Speed Vertical Extrapolation?

An accurate characterization of the wind resource available at hub-height is required for an efficient and bankable wind farm project. However, direct measurement of wind speed at the constantly increasing height of the hub of commercial wind turbines is oftentimes challenging and expensive, so that it is common practice to vertically extrapolate the wind resource from lower and more easily accessible levels. Conventional techniques for wind speed vertical extrapolation include the use of a power law and a logarithmic profile. While simple, the limits in accuracy of these methods have been shown in various studies. Recently, machine learning has been proposed as a new method to vertically extrapolate winds. All the published studies on the topic assess the performance of machine learning techniques in vertically extrapolating the wind resource at the same location where the algorithm has been trained. However, in real-world applications, the wind resource is measured at the instrument location, but it then needs to be extrapolated at hub height at the location of the wind turbines within the find farm. To be able to fully recommend the use of machine learning techniques over the simple power law and logarithmic law, the spatial variability of the performance improvements of the machine learning approaches needs to be assessed. Here, we propose a round-robin validation of a machine learning-based method for wind speed extrapolation. We use 20 months of observations at four locations spanning a 100 km wide region at the Southern Great Plains (SGP) atmospheric observatory, in north-central Oklahoma. At each location, we train a random forest to predict 30-min average wind speed at 143 m AGL. We use as input features lidar wind speed at 65 m AGL, time of day, sonic anemometer wind speed at 4 m AGL, turbulent kinetic energy, and Obukhov length. First, we perform a same-site comparison of the performance of the proposed random forest against the conventional techniques for wind speed extrapolation (namely power law and logarithmic profile, with widely accepted stability corrections). We find that the random forest outperforms the power law in vertically extrapolating wind speed in all the considered stability regimes, with a 33% reduction in MAE for stable conditions, and a 31% reduction in unstable conditions. Similar results are found when comparing predictions of extrapolated winds from the logarithmic profile and the random forest with the observed values. Next, we propose a round-robin validation, to use the random forest trained at each site to extrapolate wind speed at the remaining three sites. We find that the performance of the random forest approach degrades when the algorithm is tested at a site different than the training one. However, even under those circumstances, the machine learning-based approach still outperforms the conventional techniques for wind speed extrapolation, with, on average, a reduction in mean absolute error between 15 and 20% over the conventional methods, with the largest benefits obtained under stable conditions.

Monte Carlo↗

AUTOMATED TRANSIENT IDENTIFICATION IN THE DARK ENERGY SURVEY

We describe an algorithm for identifying point-source transients and moving objects on reference-subtracted optical images containing artifacts of processing and instrumentation. The algorithm makes use of the supervised machine learning technique known as Random Forest. We present results from its use in the Dark Energy Survey Supernova program (DES-SN), where it was trained using a sample of 898,963 signal and background events generated by the transient detection pipeline. After reprocessing the data collected during the first DES-SN observing season (2013 September through 2014 February) using the algorithm, the number of transient candidates eligible for human scanning decreased by a factor of 13.4, while only 1.0% of the artificial Type Ia supernovae (SNe) injected into search images to monitor survey efficiency were lost, most of which were very faint events. Here we characterize the algorithm’s performance in detail, and we discuss how it can inform pipeline design decisions for future time-domain imaging surveys, such as the Large Synoptic Survey Telescope and the Zwicky Transient Facility. An implementation of the algorithm and the training data used in this paper are available at at http://portal.nersc.gov/project/dessn/autoscan.

79 ASTRONOMY AND ASTROPHYSICS↗

Automatic Classification of Biological Targets in a Tidal Channel Using a Multibeam Sonar

Multibeam sonars are widely used for environmental monitoring of fauna at marine renewable energy sites. However, they can rapidly accrue vast volumes of data, which poses a challenge for data processing. Here, using data from a deployment in a tidal channel with peak currents of 1–2 m s –1 , we demonstrate the data-reduction benefits of real-time automatic classification of targets detected and tracked in multibeam sonar data. First, we evaluate classification capabilities for three machine learning algorithms: random forests, support vector machines, and k-nearest neighbors. For each algorithm, a hill-climbing search optimizes a set of hand-engineered attributes that describe tracked targets. Here, the random forest algorithm is found to be most effective—in postprocessing, discriminating between biological and nonbiological targets with a recall rate of 0.97 and a precision of 0.60. In addition, 89% of biological targets are correctly classified as either seals, diving birds, fish schools, or small targets. Model dependence on the volume of training data is evaluated. Second, a real-time implementation of the model is shown to distinguish between biological targets and nonbiological targets with nearly the same performance as in postprocessing. From this, we make general recommendations for implementing real-time classification of biological targets in multibeam sonar data and the transferability of trained models.

16 TIDAL AND WAVE POWER↗

Interpreting High-resolution Spectroscopy of Exoplanets using Cross-correlations and Supervised Machine Learning

We present a new method for performing atmospheric retrieval on ground-based, high-resolution data of exoplanets. Our method combines cross-correlation functions with a random forest, a supervised machine-learning technique, to overcome challenges associated with high-resolution data. A series of cross-correlation functions are concatenated to give a “CCF-sequence” for each model atmosphere, which reduces the dimensionality by a factor of ∼100. The random forest, trained on our grid of ∼65,000 models, provides a likelihood-free method of retrieval. The precomputed grid spans 31 values of both temperature and metallicity, and incorporates a realistic noise model. We apply our method to HARPS-N observations of the ultra-hot Jupiter KELT-9b and obtain a metallicity consistent with solar (logM = − 0.2 ± 0.2). Our retrieved transit chord temperature (T=6000{sub −200}{sup +0}K) is unreliable as strong ion lines lie outside of the extent of the training set, which we interpret as being indicative of missing physics in our atmospheric model. We compare our method to traditional nested sampling, as well as other machine-learning techniques, such as Bayesian neural networks. We demonstrate that the likelihood-free aspect of the random forest makes it more robust than nested sampling to different error distributions, and that the Bayesian neural network we tested is unable to reproduce complex posteriors. We also address the claim in Cobb et al. 2019 that our random forest retrieval technique can be overconfident but incorrect. We show that this is an artifact of the training set, rather than of the machine-learning method, and that the posteriors agree with those obtained using nested sampling.

79 ASTRONOMY AND ASTROPHYSICS↗

SPICE: discovery of phenotype-determining component interplays

Background: A latent behavior of a biological cell is complex. Deriving the underlying simplicity, or the fundamental rules governing this behavior has been the Holy Grail of systems biology. Data-driven prediction of the system components and their component interplays that are responsible for the target system’s phenotype is a key and challenging step in this endeavor. Results: The proposed approach, which we call System Phenotype-related Interplaying Components Enumerator (SPICE), iteratively enumerates statistically significant system components that are hypothesized (1) to play an important role in defining the specificity of the target system’s phenotype(s); (2) to exhibit a functionally coherent behavior, namely, act in a coordinated manner to perform the phenotype-specific function; and (3) to improve the predictive skill of the system’s phenotype(s) when used collectively in the ensemble of predictive models. SPICE can be applied to both instance-based data and network-based data. When validated, SPICE effectively identified system components related to three target phenotypes: biohydrogen production, motility, and cancer. Manual results curation agreed with the known phenotype-related system components reported in literature. Additionally, using the identified system components as discriminatory features improved the prediction accuracy by 10% on the phenotype-classification task when compared to a number of state-of-the-art methods applied to eight benchmark microarray data sets. Conclusion: We formulate a problem—enumeration of phenotype-determining system component interplays—and propose an effective methodology (SPICE) to address this problem. SPICE improved identification of cancer-related groups of genes from various microarray data sets and detected groups of genes associated with microbial biohydrogen production and motility, many of which were reported in literature. SPICE also improved the predictive skill of the system’s phenotype determination compared to individual classifiers and/or other ensemble methods, such as bagging, boosting, random forest, nearest shrunken centroid, and random forest variable selection method.

59 BASIC BIOLOGICAL SCIENCES↗

PAVC Gridded 20m Alaska NGEE Tier3 PFTs v1.0

These 20-meter spatial resolution gridded products provide per-pixel fractional cover (%) of Next Generation Ecosystem Experiments (NGEE) Arctic Plant Functional Types (PFTs) Tier 3 across Alaska, north of the boreal treeline. The products were developed for the NGEE Arctic project, which is improving Arctic vegetation representation and parameterization of the E3SM Land Model. This dataset includes 8 files containing fractional cover for NGEE Tier 3 PFTs (https://data.ess-dive.lbl.gov/view/doi:10.15485/2529470): (1) bryophytes; (2) lichens; (3) non-vascular plants, i.e., the sum of lichens and bryophytes; (4) deciduous shrubs, (5) evergreen shrubs, (6) forbs, (7) graminoids, and a non-PFT class, (8) litter. Each pixel contains the percent cover (expressed as a fraction of total ground cover) that was predicted by random-forest regression models. The random-forest models were trained on cover data collected at 978 plots from 2010 to 2021, of which are archived in the Pan-Arctic Vegetation Cover (PAVC) database (https://data.ess-dive.lbl.gov/datasets/doi:10.15485/2483557). The plot cover was linked to 20-meter spatial resolution, satellite-derived predictor variables: Sentinel-2 spectra and Sentinel-1 polarizations averaged over the 2019 growing season, as well as topographical features derived from ArcticDEM. Then, spatio-temporally anomalous plot data that introduced large variability to the regression outcomes were dropped using the Cook’s distance outlier detection method, and the models were re-created using high-quality plots and their associated satellite derived explanatory variables per each PFT. The correlations between plot-observed and satellite-derived fractional cover for all PFTs were well correlated (R2 = 0.69–0.95 and 0.5 for litter) and had low RMSE bias (0.02–0.11). This research was performed as a part of the NGEE Arctic project. The NGEE Arctic project was a research effort to reduce uncertainty in Earth System Models by developing a predictive understanding of carbon-rich Arctic ecosystems and feedbacks to climate. NGEE Arctic was supported by the Department of Energy's Office of Biological and Environmental Research.

54 ENVIRONMENTAL SCIENCES↗

Strong correlation between stress drop and peak ground acceleration for recent M1–4 earthquakes in the San Francisco Bay Area

Theoretical and observational studies suggest that between-event variability in the median ground motions of larger ( M≥5 ) earthquakes is controlled primarily by the dynamic properties of the earthquake source, such as Brune-type stress drop. Analogous results remain equivocal for smaller events due to the lack of comprehensive and overlapping ground-motion and source-parameter datasets in this regime. Here in this paper, we investigate the relationship between peak ground acceleration (PGA) and dynamic stress drop for a new dataset of 5297 earthquakes that occurred in the San Francisco Bay area from 2002 through 2016. For each event, we measure PGA on horizontal-component channels of stations within 100 km and estimate stress drop from P-wave spectra recorded on vertical-component channels of the same stations. We then develop a nonparametric ground-motion prediction equation (GMPE) applicable for the moderate (M 1–4) earthquakes in our study region, using a mixed-effects generalization of the Random Forest algorithm. We use the Random Forest GMPE to model the joint influence of magnitude, distance, and near-site effects on observed PGA. We observe a strong correlation between dynamic stress drop and the residual PGA of each event, with the events with higher-than-expected PGA associated with higher values of stress drop. The strength of this correlation increases as a function of magnitude but remains significant even for smaller magnitude events with corner frequencies that approach the observable bandwidth of the acceleration records. Mainshock events are characterized by systematically higher stress drop and PGA than aftershocks of equivalent magnitude. Coherent local variations in the distribution of dynamic stress drop provide observational constraints to support the future development of nonergodic GMPEs that account for variations in median stress drop at different source locations.

58 GEOSCIENCES↗

Learning From User Behavior: A Survey-Assist Algorithm for Longitudinal Mobility Data Collection

GPS-based travel surveys are widely used in mobility studies to gather crucial qualitative data, like purpose, transportation mode and replaced mode. However, survey response still poses a burden to users, especially in long-term mobility studies, leading to response fatigue. We explore a survey-assist strategy to ease this burden by a novel, user-level modeling approach that leverages past responses from each user to predict responses for new trips, without relying on external data sources like GIS data. We investigate three main algorithms for predicting responses: (i) clustering trips and extrapolating responses for similar trips, (ii) using random forest classification, and (iii) clustering that uses a hybrid algorithm to determine spatial structure, which is then fed as input to a classic random forest classifier. The clustering approach can flexibly predict responses for even complex qualitative survey questions; it achieved F-scores of 65%. The random forest pipeline uses architecture that restricts it to predicting three predetermined survey questions: trip purpose, mode, and replaced mode. However, it achieved F-scores of 78%. While the survey-assist approach has been implemented by several proprietary systems, to our knowledge, this is the first exploration in the academic literature. It follows that this is also the first rigorous evaluation of multiple algorithms that can implement the approach. The evaluation uses a large scale, publicly available, longitudinal dataset consisting of ~ 92k trips from 235 users over a period of roughly one and a half years. With this approach, travel surveys can be pre-filled with the predicted responses for each trip, thus streamlining the survey process for users. Combined with an active learning system that requests user input on low-confidence predictions, models can be updated and improved over time to better support the long-term collection of longitudinal qualitative data.

clustering↗

Machine learning assisted phase and size-controlled synthesis of iron oxide particles

Synthesis of iron oxides with specific phases and particle sizes is a crucial challenge in various fields, including materials science, energy storage, biomedical applications, environmental science, and earth science. However, despite significant advances in this area, much of the current palette of particle outcomes has been based on time-consuming trial-and-error exploration of synthesis conditions. The present study was designed to explore a very different approach to 1) predict the outcome of synthesis from specified reaction parameters based on using machine learning (ML) techniques, and 2) correlate sets of parameters to obtain products with desired outcomes by a newly designed recommendation algorithm. To achieve this, four ML algorithms were tested, namely random forest, logistic regression, support vector machine, and k-nearest neighbor. Among the models, random forest outperformed the others, attaining 96% and 81% accuracy when predicting the phase and size of iron oxide particles in the test dataset. Surprisingly, the permutation feature importance analysis revealed that volume, which may strongly relate to pressure, was one of the important features, along with precursor concentration, pH, temperature, and time, influencing the phase and size of iron oxide particles during synthesis. To verify the robustness of the random forest models, prediction and experimental results were compared based on 24 randomly generated methods in additive and non-additive systems not included in the datasets. The predictions of product phase and particle size from the models agreed well with the experimental results. Furthermore, a searching and ranking algorithm was developed to recommend potential synthesis parameters for obtaining iron oxide products with the desired phase and particle size from previous studies in the dataset. Furthermore, this study lays the foundation for a closed-loop approach in materials synthesis and preparation, beginning with suggesting potential reaction parameters from the dataset and predicting potential outcomes, followed by conducting experiments and analyses, and ultimately enriching the dataset.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗

Accuracy of predictions made by machine learned models for biocrude yields obtained from hydrothermal liquefaction of organic wastes

Hydrothermal liquefaction (HTL) has potential for converting abundant wet organic wastes into renewable fuels. Because HTL consists of a complex reaction network, deterministic, physics-based prediction of its biocrude yield is prohibitively difficult. Data-driven methods provide an alternative to the physics-based approach; however, rigorous testing must be performed to ensure the accuracy of predictions made by data-driven methods. To this end, a data set was assembled consisting of 570 data points appearing in the open literature. The data set was divided into training, validation, and test sub-sets and used for evaluating different machine learning regression approaches to predict biocrude yield. Among the tested algorithms, Random Forest and eXtreme Gradient Boosting (XGBoost) predicted biocrude yields in a test set that had not been used for training with the greatest accuracy, with root mean square errors (RMSE) of 8.34 and 8.57, respectively. Further refinement of the Random Forest model reduced its RMSE to 8.07. In comparison, predictions of a series of literature models resulted in RMSE ranging from 9.16 in the most accurate case to 27.6 in the least accurate; most literature models yielded RMSE values > 10. Using biocrude yield predictions from the most accurate Random Forest model and a probabilistic economic analysis found that the model accuracy is sufficient to prioritize allocation of resources based on projected minimum fuel selling price. In our report the models and analysis represent a major advance in the ability to use readily available data to predict biocrude yields on new feedstocks that have not previously been studied.

42 ENGINEERING↗

Nonlocal atlas-guided multi-channel forest learning for human brain labeling

Purpose: It is important for many quantitative brain studies to label meaningful anatomical regions in MR brain images. However, due to high complexity of brain structures and ambiguous boundaries between different anatomical regions, the anatomical labeling of MR brain images is still quite a challenging task. In many existing label fusion methods, appearance information is widely used. However, since local anatomy in the human brain is often complex, the appearance information alone is limited in characterizing each image point, especially for identifying the same anatomical structure across different subjects. Recent progress in computer vision suggests that the context features can be very useful in identifying an object from a complex scene. In light of this, the authors propose a novel learning-based label fusion method by using both low-level appearance features (computed from the target image) and high-level context features (computed from warped atlases or tentative labeling maps of the target image). Methods: In particular, the authors employ a multi-channel random forest to learn the nonlinear relationship between these hybrid features and target labels (i.e., corresponding to certain anatomical structures). Specifically, at each of the iterations, the random forest will output tentative labeling maps of the target image, from which the authors compute spatial label context features and then use in combination with original appearance features of the target image to refine the labeling. Moreover, to accommodate the high inter-subject variations, the authors further extend their learning-based label fusion to a multi-atlas scenario, i.e., they train a random forest for each atlas and then obtain the final labeling result according to the consensus of results from all atlases. Results: The authors have comprehensively evaluated their method on both public LONI-LBPA40 and IXI datasets. To quantitatively evaluate the labeling accuracy, the authors use the dice similarity coefficient to measure the overlap degree. Their method achieves average overlaps of 82.56% on 54 regions of interest (ROIs) and 79.78% on 80 ROIs, respectively, which significantly outperform the baseline method (random forests), with the average overlaps of 72.48% on 54 ROIs and 72.09% on 80 ROIs, respectively. Conclusions: The proposed methods have achieved the highest labeling accuracy, compared to several state-of-the-art methods in the literature.

61 RADIATION PROTECTION AND DOSIMETRY↗

Statistical-learning strategies generate only modestly performing predictive models for urinary symptoms following external beam radiotherapy of the prostate: A comparison of conventional and machine-learning methods

Purpose: Given the paucity of available data concerning radiotherapy-induced urinary toxicity, it is important to ensure derivation of the most robust models with superior predictive performance. This work explores multiple statistical-learning strategies for prediction of urinary symptoms following external beam radiotherapy of the prostate. Methods: The performance of logistic regression, elastic-net, support-vector machine, random forest, neural network, and multivariate adaptive regression splines (MARS) to predict urinary symptoms was analyzed using data from 754 participants accrued by TROG03.04-RADAR. Predictive features included dose-surface data, comorbidities, and medication-intake. Four symptoms were analyzed: dysuria, haematuria, incontinence, and frequency, each with three definitions (grade ≥ 1, grade ≥ 2 and longitudinal) with event rate between 2.3% and 76.1%. Repeated cross-validations producing matched models were implemented. A synthetic minority oversampling technique was utilized in endpoints with rare events. Parameter optimization was performed on the training data. Area under the receiver operating characteristic curve (AUROC) was used to compare performance using sample size to detect differences of ≥0.05 at the 95% confidence level. Results: Logistic regression, elastic-net, random forest, MARS, and support-vector machine were the highest-performing statistical-learning strategies in 3, 3, 3, 2, and 1 endpoints, respectively. Logistic regression, MARS, elastic-net, random forest, neural network, and support-vector machine were the best, or were not significantly worse than the best, in 7, 7, 5, 5, 3, and 1 endpoints. The best-performing statistical model was for dysuria grade ≥ 1 with AUROC ± standard deviation of 0.649 ± 0.074 using MARS. For longitudinal frequency and dysuria grade ≥ 1, all strategies produced AUROC>0.6 while all haematuria endpoints and longitudinal incontinence models produced AUROC<0.6. Conclusions: Logistic regression and MARS were most likely to be the best-performing strategy for the prediction of urinary symptoms with elastic-net and random forest producing competitive results. The predictive power of the models was modest and endpoint-dependent. New features, including spatial dose maps, may be necessary to achieve better models.

60 APPLIED LIFE SCIENCES↗

Wind power curve modeling in complex terrain using statistical models

The simplest power curves model wind power only as a function of the wind speed at the turbine hub height. While the latter is an essential predictor of power output, wind speed information in other parts of the vertical profile, as well as additional atmospheric variables, are also important determinants of power. The goal of this work was to determine the gain in predictive ability afforded by adding wind speed information at other heights, as well as other atmospheric variables, to the power prediction model. Using data from a wind farm with a moderately complex terrain in the Altamont Pass region in California, we trained three statistical models—a neural network, a random forest and a Gaussian process model—to predict power output from various sets of aforementioned predictors. The comparison of these predictions to the observed power data revealed that considerable improvements in prediction accuracy can be achieved both through the addition of predictors other than the hub-height wind speed and the use of statistical models. To our knowledge, the use of the Gaussian process model in this context is new, in contrast to neural networks and random forests. The advantage of this model over the other two models is that it provides a much more natural way to estimate the uncertainty associated with its predictions. In addition, this study presents one of the most comprehensive analyses to date of the relative importance of various wind power curve inputs.

Bulaevskaya, V.↗