Performance Of Random Forest Model Research Articles

Monitoring Variables Influence on Random Forest Models to Forecast Injuries in Short-Track Speed Skating.

Injuries limit the athletes' ability to participate fully in their training and competitive process. They are detrimental to performance, affecting the athletes psychologically while limiting physiological adaptations and long-term development. This study aims to present a framework for developing random forest classifier models, forecasting injuries in the upcoming 1 to 7 days, to assist the performance support staff in reducing injuries and maximizing performance within the Canadian National Female Short-Track Speed Skating Program. Forty different variables monitored daily over two seasons (2018–2019 and 2019–2020) were used to develop two sets of forecasting models. One includes only training load variables (TL), and a second (ALL) combines a wide array of monitored variables (neuromuscular function, heart rate variability, training load, psychological wellbeing, past injury type, and location). The sensitivity (ALL: 0.35 ± 0.19, TL: 0.23 ± 0.03), specificity (ALL: 0.81 ± 0.05, TL: 0.74 ± 0.03) and Matthews Correlation Coefficients (MCC) (ALL: 0.13 ± 0.05, TL: −0.02 ± 0.02) were computed. Paired T-test on the MCC revealed statistically significant (p < 0.01) and large positive effects (Cohen d > 1) for the ALL forecasting models' MCC over every forecasting window (1 to 7 days). These models were highly determined by the athletes' training completion, lower limb and trunk/lumbar injury history, as well as sFatigue, a training load marker. The TL forecasting models' MCC suggests they do not bring any added value to forecast injuries. Combining a wide array of monitored variables and quantifying the injury etiology conceptual components significantly improve the injury forecasting performance of random forest models. The ALL forecasting models' performances are promising, especially on one time windows of one or two days, with sensitivities and specificities being respectively above 0.5 and 0.7. They could add value to the decision-making process for the support staff in order to assist the Canadian National Female Team Short-Track Speed Skating program in reducing the number of incomplete training days, which could potentially increase performance. On longer forecasting time windows, ALL forecasting models' sensitivity and MCC decrease gradually. Further work is needed to determine if such models could be useful for forecasting injuries over three days or longer.

Terrain Variables Based Spatial Mapping of Soil Properties in a Watershed of Himalayan Landscape Using Random Forest Model

Soil information, including their spatial variability is vital for devising various soil and land management policies and strategies as well as assessing the environmental impact of different land use changes. Among the various soil forming factors, terrain plays a vital role in determining the spatial distribution of soil properties especially in hilly terrains due to its impact on soil processes as well as the redistribution of soil materials. The study was carried out for assessing the applicability and efficiency of various terrain parameters for prediction as well as spatial mapping of different soil properties using Random Forest (RF) modeling approach, in a mid-Himalayan watershed located in Tehri Garhwal district, Uttarakhand. Field data collection was done by collecting soil samples from 0-15cm depth (surface layer) at 68 different locations in the study area, following a transect survey method. Collected samples were pre-processed and analyzed in the laboratory for estimating various soil properties such as sand, silt and clay contents as well as soil nutrients such as Soil Organic Carbon (SOC) and nitrogen (N). Fifteen terrain parameters derived by digital terrain analysis of CartoDEM of 10m spatial resolution were used as predictor variables during RF modeling. Pearson correlation analysis revealed poor linear relationship of soil properties with various terrain parameters. Variable importance ranking revealed that compound topographic index, elevation, heat load index, integrated moisture index and mean curvature are the most important variables, explaining maximum variability in different properties. The performance of RF models for predicting various soil properties were evaluated based on the coefficient of determination (R2), Root Mean Square Error (RMSE) as well as Lin’s Concordance Correlation Coefficient (CCC). The models were found to perform well with R2 and CCC values of 0.47 and 0.50 for sand, 0.52 and 0.55 for silt, 0.44 and 0.57 for clay, 0.41 and 0.36 for organic carbon and 0.55 and 0.34 for nitrogen, respectively. The validated models were further employed for generating the spatial distribution maps of various soil properties. Such attempts combining various environmental variables with machine learning based models may aid us in mapping the spatial distribution of various soil properties in the hilly mountainous terrains with considerable accuracies.

Exploring the Use of Waste Marble Powder in Concrete and Predicting Its Strength with Different Advanced Algorithms.

Recently, the high demand for marble stones has progressed in the construction industry, ultimately resulting in waste marble production. Thus, environmental degradation is unavoidable because of waste generated from quarry drilling, cutting, and blasting methods. Marble waste is produced in an enormous amount in the form of odd blocks and unwanted rock fragments. Absence of a systematic way to dispose of these marble waste massive mounds results in environmental pollution and landfills. To reduce this risk, an effort has been made for the incorporation of waste marble powder into concrete for sustainable construction. Different proportions of marble powder are considered as a partial substitute in concrete. A total of 40 mixes are prepared. The effectiveness of marble in concrete is assessed by comparing the compressive strength with the plain mix. Supervised machine learning algorithms, bagging (Bg), random forest (RF), AdaBoost (AdB), and decision tree (DT) are used in this study to forecast the compressive strength of waste marble powder concrete. The models’ performance is evaluated using correlation coefficient (R2), root mean square error, and mean absolute error and mean square error. The achieved performance is then validated by using the k-fold cross-validation technique. The RF model, having an R2 value of 0.97, has more accurate prediction results than Bg, AdB, and DT models. The higher R2 values and lesser error (RMSE, MAE, and MSE) values are the indicators for better performance of RF model among all individual and ensemble models. The implementation of machine learning techniques for predicting the mechanical properties of concrete would be a practical addition to the civil engineering domain by saving effort, resources, and time.

Research on stock trend prediction method based on optimized random forest

AbstractAs a complex hot problem in the financial field, stock trend forecasting uses a large amount of data and many related indicators; hence it is difficult to obtain sustainable and effective results only by relying on empirical analysis. Researchers in the field of machine learning have proved that random forest can form better judgements on this kind of problem, and it has an auxiliary role in the prediction of stock trend. This study uses historical trading data of four listed companies in the USA stock market, and the purpose of this study is to improve the performance of random forest model in medium‐ and long‐term stock trend prediction. This study applies the exponential smoothing method to process the initial data, calculates the relevant technical indicators as the characteristics to be selected, and proposes the D‐RF‐RS method to optimize random forest. As the random forest is an ensemble learning model and is closely related to decision tree, D‐RF‐RS method uses a decision tree to screen the importance of features, and obtains the effective strong feature set of the model as input. Then, the parameter combination of the model is optimized through random parameter search. The experimental results show that the average accuracy of random forest is increased by 0.17 after the above process optimization, which is 0.18 higher than the average accuracy of light gradient boosting machine model. Combined with the performance of the ROC curve and Precision–Recall curve, the stability of the model is also guaranteed, which further demonstrates the advantages of random forest in medium‐ and long‐term trend prediction of the stock market.

Grain Yield Estimation in Rice Breeding Using Phenological Data and Vegetation Indices Derived from UAV Images

The accurate estimation of grain yield in rice breeding is crucial for breeders to screen and select qualified cultivars. In this study, a low-cost unmanned aerial vehicle (UAV) platform mounted with an RGB camera was carried out to capture high-spatial resolution images of rice canopy in rice breeding. The random forest (RF) regression techniques were used to establish yield models by using (1) only color vegetation indices (VIs), (2) only phenological data, and (3) fusion of VIs and phenological data as inputs, respectively. Then, the performances of RF models were compared with the manual observation and CERES-Rice model. The results indicated that the RF model using VIs only performed poorly for estimating yield; the optimized RF model that combined the use of phenological data and color VIs performed much better, which demonstrated that the phenological data significantly improved the model performance. Furthermore, the yield estimation accuracy of 21 rice cultivars that were continuously planted over three years in the optimal RF model had no significant difference (p &gt; 0.05) with that of the CERES-Rice model. These findings demonstrate that the RF model, by combining phenological data and color Vis, is a potential and cost-effective way to estimate yield in rice breeding.

Sentinel‐Based Inventory of Thermokarst Lakes and Ponds Across Permafrost Landscapes on the Qinghai‐Tibet Plateau

AbstractThermokarst lakes and ponds (hereafter referred to as thaw lakes) play an important role in the permafrost regions by regulating hydrology, ecology, and biogeochemistry. However, detailed quantitative information on thaw lake extent and distribution remains poorly resolved across the entire permafrost regions on the Qinghai‐Tibet Plateau (QTP). Here, we applied the random forest (RF) model and manual visual vectorization methods to extract thaw lake boundaries on the QTP based on Sentinel‐2 images. Accuracy assessment was comprehensively demonstrated regarding the inherent spatial resolution of imagery and RF model performance. The results showed that the accumulated uncertainty of the total thaw lake area was ±5.75 km2, and the mean accuracy (91.9%) from field‐measured boundaries of 132 thaw lakes supported the accuracy of this inventory. A total of ∼161,300 thaw lakes with sizes ranging from 500 m2 to 3 km2 were detected, with a total area of ∼2,825.45 ± 5.75 km2. Most thaw lakes were detected in the continuous permafrost type (94.1%) and within the elevations of 4,500–5,000 m (68.4%). The small thaw lakes (&lt;10,000 m2) predominated the total lake number (78.9%) but contributed to a small portion of the total lake area (12.7%). Spatial distributions of thaw lakes in terms of different climatic and environmental conditions were also comprehensively explored, including temperature, precipitation, ground thermal stability, active layer thickness, vegetation, soil properties, and underground ice content. This inventory is expected to be incorporated into Earth system models for a more comprehensive projection of the large‐scale biogeochemical feedback of thermokarst landforms on the QTP under continued global warming.

Analysis of merged whole blood transcriptomic datasets to identify circulating molecular biomarkers of feed efficiency in growing pigs

BackgroundImproving feed efficiency (FE) is an important goal due to its economic and environmental significance for farm animal production. The FE phenotype is complex and based on the measurements of the individual feed consumption and average daily gain during a test period, which is costly and time-consuming. The identification of reliable predictors of FE is a strategy to reduce phenotyping efforts.ResultsGene expression data of the whole blood from three independent experiments were combined and analyzed by machine learning algorithms to propose molecular biomarkers of FE traits in growing pigs. These datasets included Large White pigs from two lines divergently selected for residual feed intake (RFI), a measure of net FE, and in which individual feed conversion ratio (FCR) and blood microarray data were available. Merging the three datasets allowed considering FCR values (Mean = 2.85; Min = 1.92; Max = 5.00) for a total of n = 148 pigs, with a large range of body weight (15 to 115 kg) and different test period duration (2 to 9 weeks). Random forest (RF) and gradient tree boosting (GTB) were applied on the whole blood transcripts (26,687 annotated molecular probes) to identify the most important variables for binary classification on RFI groups and a quantitative prediction of FCR, respectively. The dataset was split into learning (n = 74) and validation sets (n = 74). With iterative steps for variable selection, about three hundred’s (328 to 391) molecular probes participating in various biological pathways, were identified as important predictors of RFI or FCR. With the GTB algorithm, simpler models were proposed combining 34 expressed unique genes to classify pigs into RFI groups (100% of success), and 25 expressed unique genes to predict FCR values (R2 = 0.80, RMSE = 8%). The accuracy performance of RF models was slightly lower in classification and markedly lower in regression.ConclusionFrom small subsets of genes expressed in the whole blood, it is possible to predict the binary class and the individual value of feed efficiency. These predictive models offer good perspectives to identify animals with higher feed efficiency in precision farming applications.

PREDICTION OF SYMPTOMS PROGRESSION FOR THE PATIENTS WITH KNEE OSTEOARTHRITIS BASED ON THE QUANTITATIVE STRUCTURAL FEATURES: DATA FROM THE FNIH OA BIOMARKERS CONSORTIUM

Cartilage repair can greatly alleviate the symptoms of the patients with knee osteoarthritis (KOA). However, some imaging results suggest that the patients with obvious cartilage repair may receive insignificant or even no improvement in their symptoms. This study aims to explore the possible reasons based on the structural feature of the knee joint and construct the models used to predict the progression of knee joint symptoms. 551 subjects from Osteoarthritis Biomarkers Consortium FNIH Project in the Osteoarthritis Initiative (OAI) were included and divided into training and test sets. A total of 153 structural features from five quantitative structural feature sets were included to access the structural characteristics of the knee joints. The Western Ontario and McMaster Universities (WOMAC) Osteoarthritis Index was used to evaluate the symptoms of the knee joints. A three-step feature selection method were used to screen the structural features. Finally, Naive Bayes (NB), logistic regression (LR), [Formula: see text]-nearest neighbor (KNN), support vector machine (SVM) and random forest (RF) models were constructed based on the selected features, and then compared using the receiver operating characteristic (ROC) curve. The distribution in the demographics and WOMAC symptoms scores of the participants was consistent in the training and test sets. Two demographic features and several structural features were selected using the three-step feature selection method. Among the constructed models, the models used for the progression prediction of pain, stiffness and total scores were better than that of physical function. The performance of RF model was the best while SVM model was the second best, and the performance of the remaining three models in predicting the progression of knee symptoms is indistinguishable. Structural feature-based models for the prediction of knee joint symptoms’ progression were constructed and compared. The constructed model showed good feasibility and accuracy, and may assist clinicians to predict the occurrence or progression of the knee joints symptoms in the evaluation and prognosis of cartilage repair.

POS-237 PREDICTING PROGRESSION TO END-STAGE KIDNEY DISEASE IN A CHRONIC KIDNEY DISEASE POPULATION: A COMPARISON BETWEEN A RANDOM FOREST MODEL AND THE KIDNEY FAILURE RISK EQUATION

Accurate detection and prediction of chronic kidney disease (CKD) trajectory are key to the fundamental clinical questions of timing and selection of effective therapies and management plans. Current prediction tools, such as the Kidney Failure Risk Equation (KFRE), are largely based on traditional regression techniques which assume a linear relationship between variables. Machine learning may offer a superior alternative when used to model complicated, nonlinear interactions between variables, as is likely to be the case with the precipitants and drivers of CKD progression. We hypothesized that a random forest (RF) classifier would have equivalent, if not superior, predictive performance to the 4-variable KFRE and explored the impact of serum creatinine and primary renal disease on prediction accuracy. A 2-year risk of end-stage kidney disease (ESKD) was calculated using the 4-variable KFRE in a cohort of patients with CKD stages 3-5 and compared to a RF model using the same four variables (age, gender, eGFR and urine albumin creatinine ratio). Four more RF models were developed using a combination of these as well as serum creatinine and primary renal disease. Performance of the KFRE and RF models was assessed by area under a receiver operating (AUC ROC) curve. Sensitivity and specificity were calculated for a KFRE test threshold of 10% and feature importance was evaluated for each RF model. Of 1365 patients included in the analysis, 208 progressed to ESKD in the 2 year follow-up period. The AUC ROC for KFRE was 0.95 (95% confidence interval (CI), 0.93 – 0.96) and for the RF model using the same 4 variables 0.97. The remaining four RF models had similar performance (AUC ROC 0.97 – 0.98). Specificity for all models and KFRE was excellent (91 – 94%) while sensitivity was lower for the KFRE (80%) and highest for the RF model in which eGFR was substituted by serum creatinine (92%). In the RF models, eGFR and serum creatinine had the largest effect on risk prediction while gender had the smallest. The performance of RF models in our population was non-inferior, had greater sensitivity, equivalent predictive weighting of creatinine and eGFR and little gain in performance with the addition of primary renal disease compared to KRFE in predicting 2-year risk of progression to ESKD. Random forest models should be considered as a potential tool to predict CKD progression and here forth be validated in larger and more diverse populations.

Addressing Measurement Error in Random Forests Using Quantitative Bias Analysis.

Although variables are often measured with error, the impact of measurement error on machine-learning predictions is seldom quantified. The purpose of this study was to assess the impact of measurement error on the performance of random-forest models and variable importance. First, we assessed the impact of misclassification (i.e., measurement error of categorical variables) of predictors on random-forest model performance (e.g., accuracy, sensitivity) and variable importance (mean decrease in accuracy) using data from the National Comorbidity Survey Replication (2001-2003). Second, we created simulated data sets in which we knew the true model performance and variable importance measures and could verify that quantitative bias analysis was recovering the truth in misclassified versions of the data sets. Our findings showed that measurement error in the data used to construct random forests can distort model performance and variable importance measures and that bias analysis can recover the correct results. This study highlights the utility of applying quantitative bias analysis in machine learning to quantify the impact of measurement error on study results.

Optimisation and evaluation of the random forest model in the efficacy prediction of chemoradiotherapy for advanced cervical cancer based on radiomics signature from high-resolution T2 weighted\xa0images

PurposeOur objective was to establish a random forest model and to evaluate its predictive capability of the treatment effect of neoadjuvant chemotherapy–radiation therapy.MethodsThis retrospective study included 82 patients with locally advanced cervical cancer who underwent scanning from March 2013 to May 2018. The random forest model was established and optimised based on the open source toolkit scikit-learn. Byoptimising of the number of decision trees in the random forest, the criteria for selecting the final partition index and the minimum number of samples partitioned by each node, the performance of random forest in the prediction of the treatment effect of neoadjuvant chemotherapy–radiation therapy on advanced cervical cancer (> IIb) was evaluated.ResultsThe number of decision trees in the random forests influenced the model performance. When the number of decision trees was set to 10, 25, 40, 55, 70, 85 and 100, the performance of random forest model exhibited an increasing trend first and then a decreasing one. The criteria for the selection of final partition index showed significant effects on the generation of decision trees. The Gini index demonstrated a better effect compared with information gain index. The area under the receiver operating curve for Gini index attained a value of 0.917.ConclusionThe random forest model showed potential in predicting the treatment effect of neoadjuvant chemotherapy–radiation therapy based on high-resolution T2WIs for advanced cervical cancer (> IIb).

Development and internal validation of risk prediction model of metabolic syndrome in oil workers

BackgroundThe prevalence of metabolic syndrome continues to rise sharply worldwide, seriously threatening people’s health. The optimal model can be used to identify people at high risk of metabolic syndrome as early as possible, to predict their risk, and to persuade them to change their adverse lifestyle so as to slow down and reduce the incidence of metabolic syndrome.MethodsDesign existing circumstances research. A total of 1468 workers from an oil company who participated in occupational health physical examination from April 2017 to October 2018 were included in this study. We established the Logistic regression model, the random forest model and the convolutional neural network model, and compared the prediction performance of the models according to the F1 score, sensitivity, accuracy and other indicators of the three models.ResultsThe results showed that the accuracy of the three models was 82.49,95.98 and 92.03%, the sensitivity was 87.94,95.52 and 90.59%, the specificity was 74.54, 96.65 and 94.14%, the F1 score was 0.86,0.97 and 0.93, and the area under ROC curve was 0.88,0.96 and 0.92, respectively. The Brier score of the three models was 0.15, 0.08 and 0.12, Observed-expected ratio was 0.83, 0.97 and 1.13, and the Integrated Calibration Index was 0.075,0.073 and 0.074, respectively, and explained how the random forest model was used for individual disease risk score.ConclusionsThe study showed that the prediction performance of random forest model is better than other models, and the model has higher application value, which can better predict the risk of metabolic syndrome in oil workers, and provide corresponding theoretical basis for the health management of oil workers.

Predicting tensile-shear strength of nugget using M5P model tree and random forest: An analysis

Predicting the outcomes of a weld based on few metals with respect to its process parameters is a trivial phenomenon. However, the prediction requires complex mathematical formulation when the number of metals grows. The exponential rise in testing data of welding components in recent time, have increased the data inconsistency and complexity by manifolds. Further, the multi-physical characteristic of welding data adds to its chaotic nature. This makes manual or simulation-based extraction of useful information from welded data extremely challenging. Developing predictive models for tensile-shear strength of Resistance Spot Welding (RSW) is highly latency-bound. The recent success of machine learning approaches in variety of fields gives us motivation to address this issue. In this paper, we proposed a machine learning model inspired from random forest (RF) which predicts the tensile-shear strength of nugget from its input parameters and large number of metals. We trained the prediction model using data from 435 spot-welding cases and compared its performance with widely used M5P model tree. For all cases, RF-based prediction model outperforms the M5P model in terms of accuracy. Four different feature extraction techniques namely manual feature selection, correlation attribute eval., classification attribute eval., and reliefF attribute eval. were investigated to improve the performance of random forest model. From these methods, when model is very complex i.e. higher training size, classification attribute eval. provides greater accuracy with RMSETest of 0.5442. Moreover, no overfitting and underfitting was observed in this prediction.

Booking Prediction Models for Peer-to-peer Accommodation Listings using Logistics Regression, Decision Tree, K-Nearest Neighbor, and Random Forest Classifiers

Background: Literature in the peer-to-peer accommodation has put a substantial focus on accommodation listings' price determinants. Developing prediction models related to the demand for accommodation listings is vital in revenue management because accurate price and demand forecasts will help determine the best revenue management responses.Objective: This study aims to develop prediction models to determine the booking likelihood of accommodation listings.Methods: Using an Airbnb dataset, we developed four machine learning models, namely Logistics Regression, Decision Tree, K-Nearest Neighbor (KNN), and Random Forest Classifiers. We assessed the models using the AUC-ROC score and the model development time by using the ten-fold three-way split and the ten-fold cross-validation procedures.Results: In terms of average AUC-ROC score, the Random Forest Classifiers outperformed other evaluated models. In three-ways split procedure, it had a 15.03% higher AUC-ROC score than Decision Tree, 2.93 % higher than KNN, and 2.38% higher than Logistics Regression. In the cross-validation procedure, it has a 26,99% higher AUC-ROC score than Decision Tree, 4.41 % higher than KNN, and 3.31% higher than Logistics Regression. It should be noted that the Decision Tree model has the lowest AUC-ROC score, but it has the smallest model development time.Conclusion: The performance of random forest models in predicting booking likelihood of accommodation listings is the most superior. The model can be used by peer-to-peer accommodation owners to improve their revenue management responses.

Combined Clinical and Radiomic Features Model for Prediction of Pain Relief in Palliative Radiotherapy for Spine Metastases

Pain relief is a major concern in the radiation therapy for symptomatic bone metastases. There are previous studies about investigating the association between pain relief and clinical features. This retrospective study enrolled a total of 40 patients with painful spine metastasis who received palliative radiation therapy from 2018 to 2019. Exclusion criteria were: (1) NRS 0-1; (2) reirradiation; (3) treated with SBRT; (4) metal artifacts close to target spine and (5) other extraspinal metastases apart from target spine in radiation field. Pain response defined by international consensus criteria. The 5 clinical predictive factors referred from previous study included NRS, primary tumor sites, performance status (PS), and added biologically effective dose (BED10) in this analysis. As for radiomic features, 321 features were extracted from three specified models by CT scans, which was categorizes as spinal canal model (SC model), spine model (S model), spine and surrounding tissues model (SS model), respectively. Recursive Feature Elimination and the importance score which provided by random forest method were applied to features for feature selection. A random forest model was designed to use the six selected features. The predictive performance of Random Forest models built on six radiomic features was compared with two other models, such as clinical features model, and a combined model with three clinical and three radiomic features. Internal validation was performed using ‘leave-one-out’ cross-validation method. Area under the curve (AUC) was used as predictive performance metrics. Overall 29 patients (72.5%) showed a response. The three radiomic features were extracted from SS (MajorAxisLength, RunLengthNonUniformity). The random forest models built on clinical features, radiomic features, and combination features achieved AUC of 0.47 (95%-CI 0.27-0.68), 0.70 (95%-CI 0.51-0.88), 0.77(95%-CI 0.60-0.93). We evaluate our CT-based radiomic model for prediction of response in palliative radiation therapy to spinal metastasis, and found that the combination features model seemed to be useful compared to other models.

Abstract P185: Using Electronic Health Records To Identify Patients With Apparent Treatment Resistant Hypertension

Introduction: Apparent treatment resistant hypertension (aTRH) affects 10-20% of hypertensive adults and increases risk of cardiovascular events and mortality. Fewer than half of these patients have true resistant hypertension. The majority experience pseudo-resistant hypertension due to inadequate medication adherence, white coat hypertension, and secondary causes of hypertension. We hypothesize that electronic health records can be leveraged to identify aTRH patients who would benefit from targeted counseling, medication reconciliation, and screening for secondary causes of hypertension. Methods: We studied electronic health record (EHR) data from 395 hypertensive adults in our primary care population who received longitudinal care between 2007 and 2017. Patients who met the 2008 AHA definition of resistant hypertension by chart review were considered to have aTRH. We also included 100 patients identified by heuristics targeting secondary hypertension. We extracted from the EHR demographics, vitals, laboratory results, diagnosis codes, and medications. Results outside of physiologic range were excluded and median imputation was used to handle missing data. Random forest model performance was assessed by 5-fold cross validation. Model discrimination was evaluated at an estimated positive predictive value of 75%. Results: The prevalence of aTRH in our randomly selected and full cohorts was 20.3% (n=295) and 25.8% (n=395), respectively. In cross-validation, the random forest model demonstrated a median sensitivity of 65% (IQR: 60% - 65%) and a median AUROC of 0.92 (IQR: 0.90 - 0.92). The most influential variables were related to the prescription of three or more hypertension medications; number of days on diuretics, angiotensin-converting enzyme inhibitors, or angiotensin II receptor blockers; systolic blood pressure measurements; and hypertension or diabetes diagnosis codes. Conclusion: EHR data can be used to accurately identify patients with aTRH. We expect the implementation of a clinical decision support system leveraging such models could lead to the improved care for aTRH patients.

Estimating PM2.5 concentrations in Yangtze River Delta region of China using random forest model and the Top-of-Atmosphere reflectance

Previous studies that have used remote sensing data to estimate the PM2.5 concentrations mainly focused on the retrieval of aerosol optical depth (AOD) with moderate-to-low spatial resolution. However, the complex process of retrieving AOD from satellite Top-of-Atmosphere (TOA) reflectance always generates the missingness of AOD values due to the limitation of AOD retrieval algorithms. This study validated the possibility of using satellite TOA reflectance for estimating PM2.5 concentrations, rather than using conventional AOD products retrieved from remote sensing imageries. Given that the TOA-PM2.5 relationship cannot be accurately expressed by simple linear correlation, we developed a random forest model that integrated satellite TOA reflectance from Moderate Resolution Imaging Spectroradiometer (MODIS) Level 1B product, meteorological fields and land-use variables to estimate the ground-level PM2.5 concentrations. The highly-polluted Yangtze River Delta (YRD) region of eastern China was employed as our study case. The results showed that our model performed well with a site-based and a time-based CV R2 of 0.92 and 0.88, respectively. The derived annual and seasonal distributions of PM2.5 concentrations exhibited high PM2.5 values in northern YRD region (i.e., Jiangsu province) and relatively low values in southern region (i.e., Zhejiang province), which shared a similar distribution trend with the observed PM2.5 concentrations. This study demonstrated the outstanding performance of random forest model using satellite TOA reflectance, and also provided an effective method for remotely sensed PM2.5 estimation in regions where AOD retrievals are unavailable.

Predicting soil texture from smartphone-captured digital images and an application

The rapid and non-invasive prediction of soil sand, silt, and clay is becoming increasingly attractive given the laborious nature of traditional soil textural analysis. This study proposed a novel and cheap setup comprising a smartphone, a custom-made dark chamber, and a smartphone application for predicting soil texture of the dried, ground, and sieved samples. The image acquisition system was used to capture triplicate images from 90 mineral soil samples, representing a wide textural variability from sand to clay. Local features, color features, and texture features were extracted from the cropped images and subsequently used in different combinations to predict laboratory-measured clay, silt, and sand via random forest (RF) and convolutional neural network (CNN) algorithms. Results indicated high prediction accuracy for clay (R2 = 0.97–0.98) and sand (R2 = 0.96–0.98) and moderate prediction accuracy for silt (R2 = 0.62–0.75) using both algorithms. Color features outperformed all other image-extracted features and showed the maximum influence on RF model performance. The better performance of the color features can be attributed to the color features of mineral matter and soil organic matter (SOM). An Android-based smartphone application based on the calibrated CNN model was able to predict and return soil textural values. These results exhibited the potential of the proposed system as a proximal sensor for rapid, cost-effective, and eco-friendly soil textural analysis using computer-vision and deep learning. More research is warranted to augment the setup design, develop a standalone mobile application, and measure the impacts of soil moisture and high SOM on the model prediction performance to extend the approach for on-site prediction of soil texture.

The predictive performances of random forest models with limited sample size and different species traits

The random forest (RF) model is a powerful machine learning technique that has been increasingly used for species distribution modeling (SDM) by ecologists and fisheries scientists given various threats to marine habitats and biodiversity. However, the observations for model training are often constrained by limited surveys and financial resources. Under these circumstances, identifying the appropriate sample size for modeling is important for successful predictions. In addition, species with different biological characteristics present various challenges for SDM, which needs to be considered when evaluating model performance. We built and evaluated RF models for 21 marine demersal species using catch data and environmental variables collected during a bottom trawl survey in the coastal waters of Shandong Peninsula, China. The predictive performances of the RF models were evaluated for eight sample sizes using cross validation, in which a range of 10–80 sample sites were used to train the model. The resulting predictive performance was examined for a range of biological and behavioral traits. For most species, the predictive performance of the RF model was substantially improved when the sample size increased from 10 to 30 sites, but less improvement was evident with larger datasets. An ANOVA identified significant influences of migratory behavior, lifespan, body size, feeding mode and prevalence on the model predictability, whereas the effects of trophic level and taxon were insignificant, as were the interactions between the sample size and species traits. The abundance distributions could be better predicted for benthivores, and species with short migratory distances, short lifespans, and small body sizes, and for each species trait, the variation in the relative predictive performances of the trait categorical groups was generally consistent among sample sizes and performance metrics. Our study may contribute to an improved understanding of successful SDM and provide guidance for the application of RF models to predict the abundance distributions of fish species.

Some practical aspects of predicting texture data in digital soil mapping

Soil texture is the most well-known composition in soil science. When separate components of the texture (sand, silt, and clay) are predicted independently in digital soil mapping (DSM), there is no guarantee that the separate estimates will sum to 100%. Log-ratio transformations before DSM modelling are alternatives to guarantee a constant sum of the estimates. Little is known about the effect of non-summing to 100% and transformations of particle-size fractions (PSFs) when DSM products were used to predict soil functional properties using pedotransfer functions (PTFs). Therefore, this study was conducted to investigate the effect of different soil texture modelling methods on the estimation of available soil water capacity (AWC) and the total amount of irrigation water (TIW) required for wheat production on a 4600 ha area in Khuzestan province, southwestern Iran. Specifically, this study aimed i) to assess the performance of random forest models (RF) to predict untransformed (UT) and transformed PSFs using environmental covariates; ii) to study the effects of three widely used log-ratio transformations including additive, centroid and isometric log-ratio transformations (alr, clr, and ilr respectively) on the estimations of AWC and TIW. A total of 150 soil samples were collected from the surface layers (0–30 cm) based on the conditioned Latin hypercube sampling (cLHS) procedure. Results indicated that, in terms of root mean square error (RMSE), RF provided similar accuracies in predicting PSFs for both untransformed and transformed data. However, transformation resulted in biased estimates. In addition, RF prediction based on untransformed data resulted in more correctly soil texture classes allocation when compared to transformed data. The spatial distribution of the sum of the predicted untransformed fractions indicated only small parts of the area conformed to the 100% sum. Almost the same accuracies for estimates of AWC were obtained when both untransformed and transformed predicted texture components were used as the inputs to PTFs. Data transformation can result in biased estimates of AWC. The findings indicated no significant difference between transformation methods in predicting AWC and TIW. The general patterns of the spatial distribution of the predicted AWCs across the whole area were the same for transformed and untransformed data (except for clr transformation).