Random Forest Procedure Research Articles

Estimating monthly NO2, O3, and SO2 concentrations via an ensemble three-stage procedure with downscaled satellite remote sensing data and ground measurements

The gaseous air pollutants nitrogen dioxide (NO2), ozone (O3), and sulfur dioxide (SO2), in addition to particulate matter (PM2.5), are known to be associated with many adverse health effects. However, exposure estimates may not be available in rural or mountainous areas without monitoring stations. In this study, we retrieved satellite remote sensing data for NO2, O3, and SO2 from the Ozone Monitoring Instrument (OMI) L3 products. Together with ground measurements (air monitoring stations and meteorological and land use data), we estimated the monthly NO2, O3, and SO2 concentrations with a spatial resolution of 3×3 km across Taiwan from 2005 to 2019. A three-stage estimation procedure was utilized: in Stage 1, an ensemble generalized additive model (GAM) and machine learning method were used to determine the spatiotemporal variations; in Stage 2, the remote sensing data were downscaled; and in Stage 3, the downscaled concentrations were reused in the Stage 1 procedure for fine-tuning estimations. We obtained overall leave-one-out cross-validation (LOOCV) R2 values ranging from 0.927–0.950, 0.704–0.721, and 0.601–0.716, and root-mean-square-errors (RMSEs) ranging from 1.59–2.28, 3.81–4.18, and 0.67–1.32 ppb for NO2, O3, and SO2, respectively, via the random forest procedure. The annual NO2 and SO2 concentrations greatly improved from 2005–2019, especially in the western residential area of Taiwan. However, despite these improvements in air quality, the annual O3 concentrations tended to increase from 2015–2019. This might be due to the complex mixtures of precursors (e.g., NO2), atmospheric circulation, barriers of the Central Mountain Ridge, and increasing ground temperatures over the past decade. The proposed multistage estimation procedure performed well over the whole island with complex terrain and topography. The study outcomes may provide epidemiological information for long-term ambient exposure estimates and guidance for future administrative policies.

Developing the POTOMAC Model: A Novel Prediction Model to Study the Impact of Lymphopenia Kinetics on Survival Outcomes in Head and Neck Cancer Via an Ensemble Tree-Based Machine Learning Approach.

Lymphopenia is associated with poor survival outcomes in head and neck squamous cell carcinoma (HNSCC), yet there is no consensus on whether we should limit lymphopenia risks during treatment. To fully elucidate the prognostic role of baseline versus treatment-related lymphopenia, a robust analysis is necessary to investigate the relative importance of various lymphopenia metrics (LMs) in predicting survival outcomes. In this prospective cohort study, 363 patients were eligible for analysis (patients with newly diagnosed, nonmetastatic HNSCC treated with neck radiation with or without chemotherapy in 2015-2019). Data were acquired on 28 covariates: seven baseline, five disease, seven treatment, and nine LMs, including static and time-varying features for absolute lymphocyte count (ALC), neutrophil-to-lymphocyte ratio, and immature granulocytes (IGs). IGs were included, given their hypothesized role in inhibiting lymphocyte function. Overall, there were 4.0% missing data. Median follow-up was 2.9 years. We developed a model (POTOMAC) to predict survival outcomes using a random survival forest (RSF) procedure. RSF uses an ensemble approach to reduce the risk of overfitting and provides internal validation of the model using data that are not used in model development. The ability to predict survival risk was assessed using the AUC for the predicted risk score. POTOMAC predicted 2-year survival with AUCs at 0.78 for overall survival (primary end point) and 0.73 for progression-free survival (secondary end point). Top modifiable risk factors included radiation dose and max ALC decrease. Top baseline risk factors included age, Charlson Comorbidity Index, Karnofsky Performance Score, and baseline IGs. Top-ranking LMs had superior prognostic performance when compared with human papillomavirus status, chemotherapy type, and dose (up to 2, 8, and 65 times higher in variable importance score). POTOMAC provides important insights into potential approaches to reduce mortality in patients with HNSCC treated by chemoradiation but needs to be validated in future studies.

Otolith shape analysis as a tool for stock identification of two commercially important marine fishes: Helicolenus dactylopterus and Merlucciusmerluccius

Stock identification studies are essential to understanding fish population structure and connectivity across wide geographical areas, and thus contribute to efficient fisheries management. The blackbelly rosefish, Helicolenus dactylopterus, and European hake, Merlucciusmerluccius, are two economically important marine fishes, but there are still gaps in knowledge regarding their present stock structure. Our objective was to assess the ability of otolith shape to define stock structure for the two species along the Northeastern Atlantic Ocean and the Mediterranean Sea, based on samples from eight and seven areas, for blackbelly rosefish and European hake, respectively. Shape analysis was obtained through Wavelet analysis. Canonical analysis of principal coordinates provided significant evidence for different population units with a clear separation between the Atlantic and Mediterranean populations for both species. However, random forest procedures indicated that the discrimination power varied with species and locations. For blackbelly rosefish, various Atlantic populations were more evident than for European hake. Overall, the usefulness of otolith shape to delineate stock structure of two species with distinct life history traits across a broad spatial region from the mid-Atlantic isles to the polar region, as well as the Mediterranean was demonstrated. Moving forward, it will be key to align our growing understanding of population structure with our increasing knowledge on species’ biological traits to ensure management units reflect population structure.

Patients with acute myocardial infarction and aortic valve sclerosis have an increased risk of recurrent myocardial infarction

Abstract Funding Acknowledgements Type of funding sources: Foundation. Main funding source(s): Gigi e Pupa Ferrari ONLUS. Background It has been shown that atherosclerosis risk factors have been associated with aortic valve sclerosis (AVSc) and that AVSc is associated with both all-cause and cardiovascular mortality. Thus, it is not surprising that atherosclerotic disorders (e.g., carotid or coronary atherosclerosis) and AVSc often coexist in the same subject, eventually leading to acute myocardial infarction (AMI). Patients with AMI are at increased risk of recurrent cardiovascular events with an average all-cause mortality rate at 5 years of 25%. Therefore, it is crucial to identify the precise phenotype(s) that characterize(s) patients who will experience adverse events after AMI and AVSc may serve as a marker of risk. Purpose To evaluate whether AVSc can be an independent prognostic predictor in patients with AMI or a biomarker reflecting their comorbidity burden. Methods We analysed 2388 AMI patients who were admitted at our cardiology center (2010-2019). All patients underwent echocardiographic evaluation for AVSc assessment. Long-term all-cause mortality was retrieved from a regional database system and medical records of cardiovascular hospitalizations were collected for composite events. We employed unsupervised topological data analysis (TDA) to stratify sub-groups of AMI patients with specific probabilities of recurrent AMI and evaluate the importance of AVSc in this setting. Results AVSc was detected in 1151 (48%) patients. We found a significant association of AVSc with adverse cardiovascular events even after full adjustment (HR 1.29, 95%CI 1.02-1.63; p=0.037). Since the composite outcome included cardiovascular conditions with different aetiologies, we performed a sub-analysis looking closely at each outcome. We found that recurrent AMI was associated with AVSc presence in AMI patients (HR 1.54, 95%CI: 1.16-2.05; p=0.003). Of note, in the fully adjusted model, AVSc was significantly associated with recurrent AMI in patients younger than 75 years of age (HR 1.59, 95%CI: 01.04-2.45; p=0.034). Finally, the TDA highlighted the presence of 8 clusters of patients and these clusters exhibit specific risks of recurrent AMI within 2 years after discharge. The evaluation of time-to-event curves allowed us to combine these clusters into three super-clusters (i.e., mild-, moderate-, and high-risk patients) with the average risk of event-free rate after 2 years from discharge close to 98%, 92%, and 83%, respectively. Of note, the random forest procedure showed that AVSc was the most relevant variable to discriminate the three classes of risk (i.e., super-clusters). Conclusions AVSc is frequently detected in AMI patients and it is strongly associated with long-term cardiovascular events, especially in those younger than 75 years old. AVSc has been identified to be the most relevant variable to identify patients at high risk of recurrent AMI.

Selection of climate variables by statistical-computational modeling in forest seedlings treated with growth biopromoter

Plant production requires improvement in nursery-field techniques to increase yield and economic gains. Manipulating the interaction between biotic and abiotic factors in the production of forest seedlings can improve productivity. In this context, we examined and selected the climatic variables that influence the yield of forest seedlings treated with different dosages of growth biopromoter Trichoderma sp. Climatic data were obtained from the NASA Power platform and correlated with growth data from forest seedlings treated with Trichoderma sp. The growth variables used were base diameter, height, and north-south and east-west canopy diameters. Climatic variables were selected by multiple regression models via forward-backward stepwise regression, and artificial intelligence via random forest (RF) method. The climatic variables that most influenced the growth of plants treated with Trichoderma sp. were the incidence of shortwave insolation on the horizontal surface, longwave thermal infrared radiative flux, relative air humidity, temperature, wind speed, and precipitation. Wind speed influenced only seedling height, while temperature, precipitation, and radiative flux interfered with all growth variables. Both the multiple regression and RF methods similarly describe the interaction of fungus × plant × environment. Most of the climatic variables detected via multiple regression coincided with the RF procedure. Future studies on the application of Trichoderma sp. as a biopromoter of growth should consider the identified climatic variables to optimize the production of forest seedlings in nurseries and in the field.

Screening for PTSD and TBI in Veterans using Routine Clinical Laboratory Blood Tests

Post-traumatic stress disorder (PTSD) is a mental disorder diagnosed by clinical interviews, self-report measures and neuropsychological testing. Traumatic brain injury (TBI) can have neuropsychiatric symptoms similar to PTSD. Diagnosing PTSD and TBI is challenging and more so for providers lacking specialized training facing time pressures in primary care and other general medical settings. Diagnosis relies heavily on patient self-report and patients frequently under-report or over-report their symptoms due to stigma or seeking compensation. We aimed to create objective diagnostic screening tests utilizing Clinical Laboratory Improvement Amendments (CLIA) blood tests available in most clinical settings. CLIA blood test results were ascertained in 475 male veterans with and without PTSD and TBI following warzone exposure in Iraq or Afghanistan. Using random forest (RF) methods, four classification models were derived to predict PTSD and TBI status. CLIA features were selected utilizing a stepwise forward variable selection RF procedure. The AUC, accuracy, sensitivity, and specificity were 0.730, 0.706, 0.659, and 0.715, respectively for differentiating PTSD and healthy controls (HC), 0.704, 0.677, 0.671, and 0.681 for TBI vs. HC, 0.739, 0.742, 0.635, and 0.766 for PTSD comorbid with TBI vs HC, and 0.726, 0.723, 0.636, and 0.747 for PTSD vs. TBI. Comorbid alcohol abuse, major depressive disorder, and BMI are not confounders in these RF models. Markers of glucose metabolism and inflammation are among the most significant CLIA features in our models. Routine CLIA blood tests have the potential for discriminating PTSD and TBI cases from healthy controls and from each other. These findings hold promise for the development of accessible and low-cost biomarker tests as screening measures for PTSD and TBI in primary care and specialty settings.

Mapping areas prone to piping using random forest with key explanatory variables

In previous studies, areas prone to soil piping were mapped on the catchment scale by integrating data-mining algorithms and geographical databases. However, this left the question of which input data layers are most influential on soil pipe distribution in semiarid regions, and how they can be exploited for accurate predictions of areas prone to piping. Here, a random forest (RF) procedure was applied to classify areas prone to piping using input data layers selected based on knowledge gained through previous theoretical modeling and empirical measurements. The input data layers were: potential incoming solar radiation; distance from the closest streambank; streambank slope gradient; topographic wetness index; flow accumulation; and vegetation cover as a proxy for the shading effect. An extensive dataset of field measurements (N = 774) was used to train and validate the RF classification procedure. Our results indicated that: (i) RF—based on carefully selected influential input data layers—can be used to map areas prone to soil piping at the catchment scale with high accuracy; (ii) the effect of potential solar radiation and—to a lesser extent—tree shading is noteworthy for soil cracking and subsequent piping development; and (iii) soil pipes were found at short distances (<30 m) from steep bank slopes due to the effect of tension cracks in the streambanks. These observations, over a wide region, further establish the effect of soil cracking, and drying and wetting processes, on soil pipe development in semiarid regions. Future studies applying piping susceptibility mapping at the catchment scale may benefit from using these input layers.

P331 Risk Factors and Prediction Model for Fecal Incontinence in Patients with Crohn’s Disease: A Multicenter Study

Abstract Background Fecal incontinence (FI) is a common complaint that seriously affects the quality of life in patients with Crohn’s disease (CD). We aimed to identify risk factors related to FI and construct a risk prediction model for FI in patients with CD. Methods Four hundred and sixty-one patients diagnosed with CD between June 2016 and April 2021 in Jiangsu Province Hospital of Chinese Medicine were retrospectively enrolled in this study and randomly divided into the development (n=368) and internal validation cohort (n=93). FI-related risk factors were selected from the development cohort using the random forest procedure and included in a logistic regression model from which the prediction model was elaborated. The discrimination, calibration and clinical benefit of the model were evaluated by examining the area under the receiver operating characteristic curves, calibration curves and decision curve analysis in internal validation and external validation (using 225 patients from four tertiary hospitals), respectively. Results Four independent variables were selected and included in the logistic regression model: body mass index, history of non-fistulizing perianal lesions surgery, the number of loose stools in the last week and perianal disease activity index. A nomogram was developed to facilitate risk score calculation. The model showed good discrimination ability with AUC was 0.798 and 0.780 in the internal and external cohorts, respectively. The calibration curves demonstrated good agreement with the model using the Hosmer-Lemeshow test in both cohorts (internal validation, P = 0.562; external validation, P = 0.383). DCA confirmed the clinical validity of the predictive model. Conclusion This study recognized four risk factors related to the prevalence of FI and developed a new model to effectively predict risk scores of FI in CD patients, helping to provide early risk stratification and timely intervention.

Machine-learning-based prediction of the effectiveness of the delivered dose by exhale-gated radiotherapy for locally advanced lung cancer: The additional value of geometric over dosimetric parameters alone.

This study aimed to assess interfraction stability of the delivered dose distribution by exhale-gated volumetric modulated arc therapy (VMAT) or intensity-modulated arc therapy (IMAT) for lung cancer and to determine dominant prognostic dosimetric and geometric factors. Clinical target volume (CTVPlan) from the planning CT was deformed to the exhale-gated daily CBCT scans to determine CTVi, treated by the respective dose fraction. The equivalent uniform dose of the CTVi was determined by the power law (gEUDi) and cell survival model (EUDiSF) as effectiveness measure for the delivered dose distribution. The following prognostic factors were analyzed: (I) minimum dose within the CTVi (Dmin_i), (II) Hausdorff distance (HDDi) between CTVi and CTVPlan, (III) doses and deformations at the point in CTVPlan at which the global minimum dose over all fractions per patient occurs (PDmin_global_i), and (IV) deformations at the point over all CTVi margins per patient with the largest Hausdorff distance (HDPworst). Prognostic value and generalizability of the prognostic factors were examined using cross-validated random forest or multilayer perceptron neural network (MLP) classifiers. Dose accumulation was performed using back deformation of the dose distribution from CTVi to CTVPlan. Altogether, 218 dose fractions (10 patients) were evaluated. There was a significant interpatient heterogeneity between the distributions of the normalized gEUDi values (p<0.0001, Kruskal-Wallis tests). Accumulated gEUD over all fractions per patient was 1.004-1.023 times of the prescribed dose. Accumulation led to tolerance of ~20% of fractions with gEUDi <93% of the prescribed dose. Normalized Dmin >60% was associated with predicted gEUD values above 95%. Dmin had the highest importance for predicting the gEUD over all analyzed prognostic parameters by out-of-bag loss reduction using the random forest procedure. Cross-validated random forest classifier based on Dmin as the sole input had the largest Pearson correlation coefficient (R=0.897) in comparison to classifiers using additional input variables. The neural network performed better than the random forest classifier, and the gEUD values predicted by the MLP classifier with Dmin as the sole input were correlated with the gEUD values characterized by R=0.933 (95% CI, 0.913-0.948). The performance of the full MLP model with all geometric input parameters was slightly better (R=0.952) than that based on Dmin (p=0.0034, Z-test). Accumulated dose distributions over the treatment series were robust against interfraction CTV deformations using exhale gating and online image guidance. Dmin was the most important parameter for gEUD prediction for a single fraction. All other parameters did not lead to a markedly improved generalizable prediction. Dosimetric information, especially location and value of Dmin within the CTV i , are vital information for image-guided radiation treatment.

Enhanced machine learning based feature subset through FFS enabled classification for cervical cancer diagnosis

A dataset that has massive features and imbalanced classes may be challenging for obtaining adequate accuracy in classification approaches of Machine Learning (ML). The purpose of this research is to find the optimal feature subset for cervical cancer diagnosis with efficient classification approach by estimating the performance of various Machine Learning predictive models. Filter-based feature selection techniques of Relief and Information Gain are applied in this study to calculate the rank for each feature that can be applied to order and select highest scoring features for feature selection. An optimal feature subset is generated with wrapper approach through Recursive Feature Elimination which uses a Random Forest procedure and Genetic Algorithm has been employed based on evolutionary principle. The predictive models are established with 10fold cross validation using prevalent classification algorithms like Random Forest, C5.0, K-Nearest Neighbour and Naïve Bayes. The results showed an enhancement in the average performance of these classifiers concurrently and the classification error for these classifiers decreases substantially. The experiments also exhibited that by employing this approach an optimal and reduced feature subset is desirable for the enrichment of classification accuracy with a lower computational cost. The features generated by fused approach of Relief and Genetic algorithm methods were able to predict the results in an efficient manner, hence an optimal feature subset has been nominated through this procedure. Maximum number of classifiers have shown good results in terms of performance outcomes. In addition, Random Forest method has shown advanced accuracy rate with an improved percentage of sensitivity and specificity results. Also, this work established that the best and optimal feature subset selection through Fused Feature Selection (FFS) approach could reduce the complexity of the predictive model.

A stacked deep learning approach to cyber-attacks detection in industrial systems: application to power system and gas pipeline systems.

Presently, Supervisory Control and Data Acquisition (SCADA) systems are broadly adopted in remote monitoring large-scale production systems and modern power grids. However, SCADA systems are continuously exposed to various heterogeneous cyberattacks, making the detection task using the conventional intrusion detection systems (IDSs) very challenging. Furthermore, conventional security solutions, such as firewalls, and antivirus software, are not appropriate for fully protecting SCADA systems because they have distinct specifications. Thus, accurately detecting cyber-attacks in critical SCADA systems is undoubtedly indispensable to enhance their resilience, ensure safe operations, and avoid costly maintenance. The overarching goal of this paper is to detect malicious intrusions that already detoured traditional IDS and firewalls. In this paper, a stacked deep learning method is introduced to identify malicious attacks targeting SCADA systems. Specifically, we investigate the feasibility of a deep learning approach for intrusion detection in SCADA systems. Real data sets from two laboratory-scale SCADA systems, a two-line three-bus power transmission system and a gas pipeline are used to evaluate the proposed method’s performance. The results of this investigation show the satisfying detection performance of the proposed stacked deep learning approach. This study also showed that the proposed approach outperformed the standalone deep learning models and the state-of-the-art algorithms, including Nearest neighbor, Random forests, Naive Bayes, Adaboost, Support Vector Machine, and oneR. Besides detecting the malicious attacks, we also investigate the feature importance of the cyber-attacks detection process using the Random Forest procedure, which helps design more parsimonious models.

A Data-Driven Transcriptional Taxonomy of Adipogenic Chemicals to Identify White and Brite Adipogens.

Background:Chemicals in disparate structural classes activate specific subsets of the transcriptional programs of peroxisome proliferator-activated receptor- () to generate adipocytes with distinct phenotypes.Objectives:Our objectives were to a) establish a novel classification method to predict ligands and modifying chemicals; and b) create a taxonomy to group chemicals on the basis of their effects on transcriptome and downstream metabolic functions. We tested the hypothesis that environmental adipogens highly ranked by the taxonomy, but segregated from therapeutic ligands, would induce white but not brite adipogenesis.Methods:3T3-L1 cells were differentiated in the presence of 76 chemicals (negative controls, nuclear receptor ligands known to influence adipocyte biology, potential environmental ligands). Differentiation was assessed by measuring lipid accumulation. mRNA expression was determined by RNA-sequencing (RNA-Seq) and validated by reverse transcription–quantitative polymerase chain reaction. A novel classification model was developed using an amended random forest procedure. A subset of environmental contaminants identified as strong agonists were analyzed by their effects on lipid handling, mitochondrial biogenesis, and cellular respiration in 3T3-L1 cells and human preadipocytes.Results:We used lipid accumulation and RNA-Seq data to develop a classification system that a) identified agonists; and b) sorted chemicals into likely white or brite adipogens. Expression of Cidec was the most efficacious indicator of strong activation. 3T3-L1 cells treated with two known environmental ligands, tetrabromobisphenol A and triphenyl phosphate, which sorted distinctly from therapeutic ligands, had higher expression of white adipocyte genes but no difference in Pgc1a and Ucp1 expression, and higher fatty acid uptake but not mitochondrial biogenesis. Moreover, cells treated with two chemicals identified as highly ranked agonists, tonalide and quinoxyfen, induced white adipogenesis without the concomitant health-promoting characteristics of brite adipocytes in mouse and human preadipocytes.Discussion:A novel classification procedure accurately identified environmental chemicals as ligands distinct from known -activating therapeutics.Conclusion:The computational and experimental framework has general applicability to the classification of as-yet uncharacterized chemicals. https://doi.org/10.1289/EHP6886

Sonographic Features of Chronic Kidney Disease in Agricultural Community in Sri Lanka

Objectives: The aim of this study was to use ultrasound-based kidney morphological features to classify chronic kidney disease (CKD) in an agricultural community in Sri Lanka where there is a high prevalence of CKD with unknown etiology. Materials and Methods: A cohort of CKD patients (n = 50) and healthy subjects (n = 26) underwent B-mode renal ultrasound. CKD patients were further categorized as those clinically diagnosed with diabetes mellitus, hypertension, and other known causes (n = 30) and those of unknown etiology (n = 20). Following kidney morphological features were calculated: Length (LEN), width (WDTH), cortical thickness, volume (VOL), and shape index. Results: CKD kidneys of both groups were significantly smaller than the healthy kidneys (P &lt; 0.001). Based on a random forest procedure, the top three influential features that distinguished CKD kidneys from healthy kidneys were: VOL normalized to waist circumference (CKD = 0.6 ± 0.2 cm2, healthy = 0.9 ± 0.2 cm2), VOL normalized to body surface area (CKD = 36 ± 9 cm3/m2, healthy = 52 ± 13 cm3/m2), and WDTH (CKD = 3.6 ± 0.5 cm, healthy = 4.3 ± 0.6 cm). Patients with CKD of unknown etiology had higher kidney LEN and VOL normalized to height (HGHT) (LEN/HGHT = 0.58 ± 0.05 cm/m, VOL/HGHT = 0.40 ± 0.09 cm3/m, P &lt; 0.05) compared to those of the known etiology group (LEN/HGHT = 0.51 ± 0.09 cm/m, VOL/HGHT = 0.30 ± 0.10 cm3/m). Conclusion: The study shows that ultrasound-based kidney volume can distinguish healthy versus diseased kidneys as well as CKD of known versus unknown etiology. Normalizing for height is required when comparing diseased groups.

X-Ray Spectra and Multiwavelength Machine Learning Classification for Likely Counterparts to Fermi 3FGL Unassociated Sources

Abstract We conduct X-ray spectral fits on 184 likely counterparts to Fermi-LAT 3FGL unassociated sources. Characterization and classification of these sources allows for more complete population studies of the high-energy sky. Most of these X-ray spectra are well fit by an absorbed power-law model, as expected for a population dominated by blazars and pulsars. A small subset of seven X-ray sources have spectra unlike the power law expected from a blazar or pulsar and may be linked to coincident stars or background emission. We develop a multiwavelength machine learning classifier to categorize unassociated sources into pulsars and blazars using gamma-ray and X-ray observations. Training a random forest (RF) procedure with known pulsars and blazars, we achieve a cross-validated classification accuracy of 98.6%. Applying the RF routine to the unassociated sources returned 126 likely blazar candidates (defined as P bzr ≥ 90%) and five likely pulsar candidates (P bzr ≤ 10%). Our new X-ray spectral analysis does not drastically alter the RF classifications of these sources compared to previous works, but it builds a more robust classification scheme and highlights the importance of X-ray spectral fitting. Our procedure can be further expanded with UV, visual, or radio spectral parameters or by measuring flux variability.

Subsampling from features in large regression to find “winning features”

AbstractFeature (or variable) selection from a large number of p features continuously challenges data science, especially for ever‐enlarging data and in discovering scientifically important features in a regression setting. For example, to develop valid drug targets for ovarian cancer, we must control the false‐discovery rate (FDR) of a selection procedure. The popular approach to feature selection in large‐p regression uses a penalized likelihood or a shrinkage estimation, such as a LASSO, SCAD, Elastic Net, or MCP procedure. We present a different approach called the Subsampling Winner algorithm (SWA), which subsamples from p features. The idea of SWA is analogous to selecting US national merit scholars' that selects semifinalists based on student's performance in tests done at local schools (a.k.a. subsample analyses), and then determine the finalists (a.k.a. winning features) from the semifinalists. Due to its subsampling nature, SWA can scale to data of any dimension. SWA also has the best‐controlled FDR compared to the penalized and Random Forest procedures while having a competitive true‐feature discovery rate. Our application of SWA to an ovarian cancer data revealed functionally important genes and pathways.

Comparison of Machine Learning Algorithms for Wildland-Urban Interface Fuelbreak Planning Integrating ALS and UAV-Borne LiDAR Data and Multispectral Images

Controlling vegetation fuels around human settlements is a crucial strategy for reducing fire severity in forests, buildings and infrastructure, as well as protecting human lives. Each country has its own regulations in this respect, but they all have in common that by reducing fuel load, we in turn reduce the intensity and severity of the fire. The use of Unmanned Aerial Vehicles (UAV)-acquired data combined with other passive and active remote sensing data has the greatest performance to planning Wildland-Urban Interface (WUI) fuelbreak through machine learning algorithms. Nine remote sensing data sources (active and passive) and four supervised classification algorithms (Random Forest, Linear and Radial Support Vector Machine and Artificial Neural Networks) were tested to classify five fuel-area types. We used very high-density Light Detection and Ranging (LiDAR) data acquired by UAV (154 returns·m−2 and ortho-mosaic of 5-cm pixel), multispectral data from the satellites Pleiades-1B and Sentinel-2, and low-density LiDAR data acquired by Airborne Laser Scanning (ALS) (0.5 returns·m−2, ortho-mosaic of 25 cm pixels). Through the Variable Selection Using Random Forest (VSURF) procedure, a pre-selection of final variables was carried out to train the model. The four algorithms were compared, and it was concluded that the differences among them in overall accuracy (OA) on training datasets were negligible. Although the highest accuracy in the training step was obtained in SVML (OA=94.46%) and in testing in ANN (OA=91.91%), Random Forest was considered to be the most reliable algorithm, since it produced more consistent predictions due to the smaller differences between training and testing performance. Using a combination of Sentinel-2 and the two LiDAR data (UAV and ALS), Random Forest obtained an OA of 90.66% in training and of 91.80% in testing datasets. The differences in accuracy between the data sources used are much greater than between algorithms. LiDAR growth metrics calculated using point clouds in different dates and multispectral information from different seasons of the year are the most important variables in the classification. Our results support the essential role of UAVs in fuelbreak planning and management and thus, in the prevention of forest fires.

Factors Affecting Sentinel Node Metastasis in Thin (T1) Cutaneous Melanomas: Development and External Validation of a Predictive Nomogram.

Thin melanomas (T1; ≤ 1 mm) constitute 70% of newly diagnosed cutaneous melanomas. Regional node metastasis determined by sentinel node biopsy (SNB) is an important prognostic factor for T1 melanoma. However, current melanoma guidelines do not provide clear indications on when to perform SNB in T1 disease and stress an individualized approach to SNB that considers all clinicopathologic risk factors. We aimed to identify determinants of sentinel node (SN) status for incorporation into an externally validated nomogram to better select patients with T1 disease for SNB. The development cohort comprised 3,666 patients with T1 disease consecutively treated at the Istituto Nazionale Tumori (Milan, Italy) between 2001 and 2018; 4,227 patients with T1 disease treated at 13 other European centers over the same period formed the validation cohort. A random forest procedure was applied to the development data set to select characteristics associated with SN status for inclusion in a multiple binary logistic model from which a nomogram was elaborated. Decision curve analyses assessed the clinical utility of the nomogram. Of patients in the development cohort, 1,635 underwent SNB; 108 patients (6.6%) were SN positive. By univariable analysis, age, growth phase, Breslow thickness, ulceration, mitotic rate, regression, and lymphovascular invasion were significantly associated with SN status. The random forest procedure selected 6 variables (not growth phase) for inclusion in the logistic model and nomogram. The nomogram proved well calibrated and had good discriminative ability in both cohorts. Decision curve analyses revealed the superior net benefit of the nomogram compared with each individual variable included in it as well as with variables suggested by current guidelines. We propose the nomogram as a decision aid in all patients with T1 melanoma being considered for SNB.

A Machine Learning Approach for High-Dimensional Time-to-Event Prediction With Application to Immunogenicity of Biotherapies in the ABIRISK Cohort.

Predicting immunogenicity for biotherapies using patient and drug-related factors represents nowadays a challenging issue. With the growing ability to collect massive amount of data, machine learning algorithms can provide efficient predictive tools. From the bio-clinical data collected in the multi-cohort of autoimmune diseases treated with biotherapies from the ABIRISK consortium, we evaluated the predictive power of a custom-built random survival forest for predicting the occurrence of anti-drug antibodies. This procedure takes into account the existence of a population composed of immune-reactive and immune-tolerant subjects as well as the existence of a tiny expected proportion of relevant predictive variables. The practical application to the ABIRISK cohort shows that this approach provides a good predictive accuracy that outperforms the classical survival random forest procedure. Moreover, the individual predicted probabilities allow to separate high and low risk group of patients. To our best knowledge, this is the first study to evaluate the use of machine learning procedures to predict biotherapy immunogenicity based on bioclinical information. It seems that such approach may have potential to provide useful information for the clinical practice of stratifying patients before receiving a biotherapy.

Identification of genetic features associated with fine particulate matter (PM2.5) modulated DNA damage using improved random forest analysis.

Recent studies have found multiple single nucleotide variants (SNVs) associated with DNA damage. However, previous association analysis may ignore the potential interaction effects between SNVs. Therefore, we used an improved random forest (RF) analysis to identify the SNVs related to personal DNA damage in exon-focused genome-wide association study (GWAS). A total of 301 subjects from three independent centers (Zhuhai, Wuhan, and Tianjin) were retained for analysis. An improved RF procedure was used to systematically screen key SNVs associated with DNA damage. Furthermore, we used genetic risk score (GRS) and mediation analysis to investigate the integrative effect and potential mechanism of these genetic variants on DNA damage. Besides, gene set enrichment analysis was conducted to identify the pathways enriched by key SNVs using the Data-driven Expression Prioritized Integration for Complex Traits (DEPICT). Finally, a set of 24 SNVs with the lowest mean square errors (MSE) were identified by improved RF analysis. Both weighted and unweighted GRSs were associated with increased DNA damage levels (Pweight<0.001 and Punweight<0.001). Gene set enrichment analysis indicated that these loci were significantly enriched in several biological features associated with DNA damage. These findings suggested the role of SNVs in modifying DNA damage levels. It may be convincing that this improved RF analysis can effectively identify SNVs associated with DNA damage levels.

Heterogeneity of executive function revealed by a functional random forest approach across ADHD and ASD.

Those with autism spectrum disorder (ASD) and/or attention-deficit-hyperactivity disorder (ADHD) exhibit symptoms of hyperactivity and inattention, causing significant hardships for families and society. A potential mechanism involved in these conditions is atypical executive function (EF). Inconsistent findings highlight that EF features may be shared or distinct across ADHD and ASD. With ADHD and ASD each also being heterogeneous, we hypothesized that there may be nested subgroups across disorders with shared or unique underlying mechanisms. Participants (N=130) included adolescents aged 7-16 with ASD (n=64) and ADHD (n=66). Typically developing (TD) participants (n=28) were included for a comparative secondary sub-group analysis. Parents completed the K-SADS and youth completed an extended battery of executive and other cognitive measures. A two stage hybrid machine learning tool called functional random forest (FRF) was applied as a classification approach and then subsequently to subgroup identification. We input 43 EF variables to the classification step, a supervised random forest procedure in which the features estimated either hyperactive or inattentive ADHD symptoms per model. The FRF then produced proximity matrices and identified optimal subgroups via the infomap algorithm (a type of community detection derived from graph theory). Resting state functional connectivity MRI (rs-fMRI) was used to evaluate the neurobiological validity of the resulting subgroups. Both hyperactive (Mean absolute error (MAE)=0.72, Null model MAE=0.8826, (t(58)=-4.9, p<.001) and inattentive (MAE=0.7, Null model MAE=0.85, t(58)=-4.4, p<.001) symptoms were predicted better than chance by the EF features selected. Subgroup identification was robust (Hyperactive: Q=0.2356, p<.001; Inattentive: Q=0.2350, p<.001). Two subgroups representing severe and mild symptomology were identified for each symptom domain. Neuroimaging data revealed that the subgroups and TD participants significantly differed within and between multiple functional brain networks, but no consistent "severity" patterns of over or under connectivity were observed between subgroups and TD. The FRF estimated hyperactive/inattentive symptoms and identified 2 distinct subgroups per model, revealing distinct neurocognitive profiles of Severe and Mild EF performance per model. Differences in functional connectivity between subgroups did not appear to follow a severity pattern based on symptom expression, suggesting a more complex mechanistic interaction that cannot be attributed to symptom presentation alone.