Random Forest Prediction Model Research Articles

A multicenter random forest model for effective prognosis prediction in collaborative clinical research network.

The accuracy of a prognostic prediction model has become an essential aspect of the quality and reliability of the health-related decisions made by clinicians in modern medicine. Unfortunately, individual institutions often lack sufficient samples, which might not provide sufficient statistical power for models. One mitigation is to expand data collection from a single institution to multiple centers to collectively increase the sample size. However, sharing sensitive biomedical data for research involves complicated issues. Machine learning models such as random forests (RF), though they are commonly used and achieve good performances for prognostic prediction, usually suffer worse performance under multicenter privacy-preserving data mining scenarios compared to a centrally trained version. In this study, a multicenter random forest prognosis prediction model is proposed that enables federated clinical data mining from horizontally partitioned datasets. By using a novel data enhancement approach based on a differentially private generative adversarial network customized to clinical prognosis data, the proposed model is able to provide a multicenter RF model with performances on par with-or even better than-centrally trained RF but without the need to aggregate the raw data. Moreover, our model also incorporates an importance ranking step designed for feature selection without sharing patient-level information. The proposed model was evaluated on colorectal cancer datasets from the US and China. Two groups of datasets with different levels of heterogeneity within the collaborative research network were selected. First, we compare the performance of the distributed random forest model under different privacy parameters with different percentages of enhancement datasets and validate the effectiveness and plausibility of our approach. Then, we compare the discrimination and calibration ability of the proposed multicenter random forest with a centrally trained random forest model and other tree-based classifiers as well as some commonly used machine learning methods. The results show that the proposed model can provide better prediction performance in terms of discrimination and calibration ability than the centrally trained RF model or the other candidate models while following the privacy-preserving rules in both groups. Additionally, good discrimination and calibration ability are shown on the simplified model based on the feature importance ranking in the proposed approach. The proposed random forest model exhibits ideal prediction capability using multicenter clinical data and overcomes the performance limitation arising from privacy guarantees. It can also provide feature importance ranking across institutions without pooling the data at a central site. This study offers a practical solution for building a prognosis prediction model in the collaborative clinical research network and solves practical issues in real-world applications of medical artificial intelligence.

Prediction of the uniaxial compressive strength of rocks from simple index tests using a random forest predictive model

Prediction of the uniaxial compressive strength of rocks from simple index tests using a random forest predictive model

Ultra-short-term prediction method of photovoltaic electric field power based on ground-based cloud image segmentation

As a large number of photovoltaic power stations are built and put into operation, the total amount of photovoltaic power generation accounts for an increasing proportion of the total electricity. The inability to accurately predict solar energy output has brought great uncertainty to the grid. Therefore, predicting the future power of photovoltaic fields is of great significance. According to different time scales, predictions are divided into long-term, medium-term and ultra-short-term predictions. The main difficulty of ultra-short-term forecasting lies in the power fluctuations caused by sudden and drastic changes in environmental factors. The shading of clouds is directly related to the irradiance received on the surface of the photovoltaic panel, which has become the main factor affecting the fluctuation of photovoltaic power generation. Therefore, sky images captured by conventional cameras installed near solar panels can be used to analyze cloud characteristics and improve the accuracy of ultra-short-term predictions. This paper uses historical power information of photovoltaic power plants and cloud image data, combined with machine learning methods, to provide ultra-short-term predictions of the power generation of photovoltaic power plants. First, the random forest method is used to use historical power generation data to establish a single time series prediction model to predict ultra-short-term power generation. Compared with the continuous model, the root mean square (RMSE) error of prediction is reduced by 28.38%. Secondly, the Unet network is used to segment the cloud image, and the cloud amount information is analyzed and input into the random forest prediction model to obtain the bivariate prediction model. The experimental results prove that, based on the cloud amount information contained in the cloud chart, the bivariate prediction model has an 11.56% increase in prediction accuracy compared with the single time series prediction model, and an increase of 36.66% compared with the continuous model.

2813 Clinical Prediction of Unsuccessful Endometrial Ablation: Random Forest vs Logistic Regression

Study Objective To compare the performance of the prediction models of surgical re-intervention within 2 years after endometrial ablation (EA) by a multivariate random forest model vs the previously presented multivariate logistic regression model. Design Retrospective cohort study, minimal follow-up time of 2 years. Setting Data from Catharina Hospital, Eindhoven and Elkerliek Hospital, Helmond, both non-university teaching hospitals in the Netherlands, were used. Patients or Participants Pre-menopausal women (18+) who have had an EA for heavy menstrual bleeding between January 2004 & April 2013. A total number of 446 patients were eligible for analysis. Interventions Used ablation methods were Cavatherm (Veldana Medical SA, Morges, Switzerland), Gynecare Thermachoice (Ethicon, Sommerville, US.) and Thermablate EAS (Idoman, Ireland). Used interventions and other ablation techniques had the same outcomes according to previously published literature. Measurements and Main Results Data-analysis was done by using IBM SPSS statistics software version 21.0 (IBM Corp., Armonk, NY, USA). The random forest model was trained in MATLAB (2018b) using the TreeBagger function in the Statistics and Machine Learning Toolbox. The prediction model based on a multivariate logistic regression analysis had an AUC of 0.71. The machine learning model had an AUC of 0.63 and an AUC of 0.65 after hyperparameter optimization. Conclusion Based on the preliminary results, we can conclude that the random forest model in this case is not better than the logistic regression model to predict the outcome of surgical re-intervention within two years after EA. In summary, the performance of a random forest clinical prediction model is not necessarily superior to a logistic regression model. The performance of each model is influenced by the sample size, the number of predictors, hyperparameter tuning and the linearity of associations.

Establishment of a CT image radiomics-based prediction model for the differential diagnosis of silicosis and tuberculosis nodules

Objective: To establish a CT image radiomics-based prediction model for the differential diagnosis of silicosis and tuberculosis nodules. Methods: A total of 53 patients with silicosis and 89 patients with tuberculosis who underwent routine CT scans in Suzhou Fifth People's Hospital from January to August, 2018 were enrolled in this study. AK/ITK software was used to segment the images to obtain 139 silicosis lesions and 119 tuberculosis lesions. For each lesion image, 396 features were extracted, and feature dimension reduction was applied to select the most characteristic feature subset. Support vector machine (SVM) , feedforward back propagation neural network (FNN-BP) , and random forest (RF) were implemented using R software (Rstudio V1.1.463) , and the algorithm that achieved the largest area under of the receiver operating characteristic (ROC) curve (AUC) was selected as the final prediction model. Results: RF was the best prediction model for the differential diagnosis of silicosis and tuberculosis nodules, with an accuracy of 83.1%, a sensitivity of 0.76, a specificity of 0.9, and an AUC of 0.917 (95% confidence interval: 0.8431-0.9758) . RF had a significantly larger AUC than SVM and FNN-BP (P<0.05) . Conclusion: CT image-based RF prediction model can be used to differentially diagnose silicosis and tuberculosis nodules.

APPLICATION OF MACHINE LEARNING TOOLS FOR PREDICTING DETERMINANT FACTORS

Abstract: Machine learning is a technique of optimizing a performance criterion using example data and past experience. Data in machine learning plays a key role, and machine leaning tools are used to discover and learn knowledge from the datasets stored. The purpose of this research is to build a model that can predict the determinant factors for crop production status using machine learning techniques as a means of visualizing the data. In order to conduct this research supervised machine learning techniques were employed. For the purpose of this research, the datasets were collected from selected region agricultural offices. The data sets used for the training and testing of the predictive model is 10,000 instances with 41 regular attributes. As a result, for identifying the determinant factors Rapid Miner machine learning tool was used. In order to find the best predictive modeling technique different experiments were conducted using Random Forest, Decision tree, Naive Bays and ID3 predictive models. To validate the predictive performance of the selected models split and cross validation testing methods was used. As the findings of this research shows that, Random Forest and decision tree models were performed the highest accuracy and precision than others. Therefore, the Random Forest predictive modeling have been used to predict the determinate factors form small and large datasets.

Application of machine learning models in predicting early stone-free rate after flexible ureteroscopic lithotripsy for renal stones

To establish predictive models based on random forest and XGBoost machine learning algorithm and to investigate their value in predicting early stone-free rate (SFR) after flexible ureteroscopic lithotripsy (fURL) in patients with renal stones. The clinical data of 201 patients with renal stones who underwent fURL were retrospectively investigated. According to the stone-free standard, the patients were divided into stone-free group (SF group) and stone-residual group (SR group). We compared a number of factors including patient age, body mass index (BMI), stone number, stone volume, stone density and hydronephrosis between the two groups. For low calyceal calculi, renal anatomic parameters including infundibular angle (IPA), infundibular width (IW), infundibular length (IL) and pelvic calyceal height (PCH), would be measured. We brought above potential predictive factors into random forest and XGBoost machine learning algorithm respectively to develop two predictive models. The receiver operating characteristic curve (ROC curve) was established in order to test the predictive ability of the model. Clinical data of 71 patients were collected prospectively to validate the predictive models externally. In this study, 201 fURL operations were successfully completed. The one-phase early SFR was 61.2%. We built two predictive models based on random forest and XGBoost machine learning algorithm. The predictive variables' importance scores were obtained. The area under the ROC curve (AUROC) of the two predictive models for early stone clearance status prediction was 0.77. In the study, 71 test samples were used for external validation. The results showed that the total predictive accuracy, predictive specificity and predictive sensitivity of the random forest and XGBoost models were 75.7%, 82.6%, 60.0%, and 81.4%, 87.0%, 68.0%, respectively. The first four predictive variables in importance were stone volume, mean stone density, maximal stone density and BMI in both random forest and XGBoost predictive models. The predictive models based on random forest and XGBoost machine learning algorithm can predict postoperative early stone status after fURL for renal stones accurately, which will facilitate preoperative evaluation and clinical decision-making. Stone volume, mean stone density, maximal stone density and BMI may be the important predictive factors affecting early SFR after fURL for renal stones.

Genomic characterization of mumps viruses from a large-scale mumps outbreak in Arkansas, 2016

In 2016, a year-long large-scale mumps outbreak occurred in Arkansas among a highly-vaccinated population. A total of 2954 mumps cases were identified during this outbreak. The majority of cases (1676 (57%)) were school-aged children (5–17 years), 1536 (92%) of these children had completed the mumps vaccination schedule. To weigh the possibility that the mumps virus evaded vaccine-induced immunity in the affected Arkansas population, we established a pipeline for genomic characterization of the outbreak strains. Our pipeline produces whole-genome sequences along with phylogenetic analysis of the outbreak mumps virus strains. We collected buccal swab samples of patients who tested positive for the mumps virus during the 2016 Arkansas outbreak, and used the portable Oxford Nanopore Technology to sequence the extracted strains. Our pipeline identified the genotype of the Arkansas mumps strains as genotype G and presented a genome-based phylogenetic tree with superior resolution to a standard small hydrophobic (SH) gene-based tree. We phylogenetically compared the Arkansas whole-genome sequences to all publicly available mumps strains. While these analyses show that the Arkansas mumps strains are evolutionarily distinct from the vaccine strains, we observed no correlation between vaccination history and phylogenetic grouping. Furthermore, we predicted potential B-cell epitopes encoded by the Arkansas mumps strains using a random forest prediction model trained on antibody-antigen protein structures. Over half of the predicted epitopes of the Jeryl-Lynn vaccine strains in the Hemagglutinin-Neuraminidase (HN) surface glycoprotein (a major target of neutralizing antibodies) region are missing in the Arkansas mumps strains. In-silico analyses of potential epitopes may indicate that the Arkansas mumps strains display antigens with reduced immunogenicity, which may contribute to reduced vaccine effectiveness. However, our in-silico findings should be assessed by robust experiments such as cross neutralization assays. Metadata analysis showed that vaccination history had no effect on the evolution of the Arkansas mumps strains during this outbreak. We conclude that the driving force behind the spread of the mumps virus in the 2016 Arkansas outbreak remains undetermined.

Artificial Intelligence Prediction Model for the Cost and Mortality of Renal Replacement Therapy in Aged and Super-Aged Populations in Taiwan.

Background: Prognosis of the aged population requiring maintenance dialysis has been reportedly poor. We aimed to develop prediction models for one-year cost and one-year mortality in aged individuals requiring dialysis to assist decision-making for deciding whether aged people should receive dialysis or not. Methods: We used data from the National Health Insurance Research Database (NHIRD). We identified patients first enrolled in the NHIRD from 2000–2011 for end-stage renal disease (ESRD) who underwent regular dialysis. A total of 48,153 Patients with ESRD aged ≥65 years with complete age and sex information were included in the ESRD cohort. The total medical cost per patient (measured in US dollars) within one year after ESRD diagnosis was our study’s main outcome variable. We were also concerned with mortality as another outcome. In this study, we compared the performance of the random forest prediction model and of the artificial neural network prediction model for predicting patient cost and mortality. Results: In the cost regression model, the random forest model outperforms the artificial neural network according to the mean squared error and mean absolute error. In the mortality classification model, the receiver operating characteristic (ROC) curves of both models were significantly better than the null hypothesis area of 0.5, and random forest model outperformed the artificial neural network. Random forest model outperforms the artificial neural network models achieved similar performance in the test set across all data. Conclusions: Applying artificial intelligence modeling could help to provide reliable information about one-year outcomes following dialysis in the aged and super-aged populations; those with cancer, alcohol-related disease, stroke, chronic obstructive pulmonary disease (COPD), previous hip fracture, osteoporosis, dementia, and previous respiratory failure had higher medical costs and a high mortality rate.

Synergy of ICESat-2 and Landsat for Mapping Forest Aboveground Biomass with Deep Learning

Spatially continuous estimates of forest aboveground biomass (AGB) are essential to supporting the sustainable management of forest ecosystems and providing invaluable information for quantifying and monitoring terrestrial carbon stocks. The launch of the Ice, Cloud, and land Elevation Satellite-2 (ICESat-2) on September 15th, 2018 offers an unparalleled opportunity to assess AGB at large scales using along-track samples that will be provided during its three-year mission. The main goal of this study was to investigate deep learning (DL) neural networks for mapping AGB with ICESat-2, using simulated photon-counting lidar (PCL)-estimated AGB for daytime, nighttime, and no noise scenarios, Landsat imagery, canopy cover, and land cover maps. The study was carried out in Sam Houston National Forest located in south-east Texas, using a simulated PCL-estimated AGB along two years of planned ICESat-2 profiles. The primary tasks were to investigate and determine neural network architecture, examine the hyper-parameter settings, and subsequently generate wall-to-wall AGB maps. A first set of models were developed using vegetation indices calculated from single-date Landsat imagery, canopy cover, and land cover, and a second set of models were generated using metrics from one year of Landsat imagery with canopy cover and land cover maps. To compare the effectiveness of final models, comparisons with Random Forests (RF) models were made. The deep neural network (DNN) models achieved R2 values of 0.42, 0.49, and 0.50 for the daytime, nighttime, and no noise scenarios respectively. With the extended dataset containing metrics calculated from Landsat images acquired on different dates, substantial improvements in model performance for all data scenarios were noted. The R2 values increased to 0.64, 0.66, and 0.67 for the daytime, nighttime, and no noise scenarios. Comparisons with Random forest (RF) prediction models highlighted similar results, with the same R2 and root mean square error (RMSE) range (15–16 Mg/ha) for daytime and nighttime scenarios. Findings suggest that there is potential for mapping AGB using a combinatory approach with ICESat-2 and Landsat-derived products with DL.

Prediction Model for Road Traffic Accident Based on Random Forest

More than 1.3 million people worldwide die in road traffic accidents every year, and the number of traffic accident deaths in China is increasing by 10% every year. Traffic safety has become an important direction for the development of road transportation systems. At present, research in progress of road traffic accidents is mainly from accidents numbers and the degree of accident damage. Firstly, this paper proposes a random forest prediction model for the prediction of the number of road traffic accidents, which applied to the actual to verify the validity of the model by using the R software to solve it. 1. Predictive

Prescribed fire affects the concentration and aromaticity of soluble soil organic matter in forest soils

Prescribed burning is used widely across Australia to reduce fuel load and associated wildfire hazard. However, prescribed burning can influence carbon (C) storage in affected ecosystems, potentially influencing C cycling. One important component of the C cycle is soluble C, which is available to microorganisms and therefore a key driver of nutrient cycling. However, studies of the effects of fire on soluble C in natural ecosystems are few. In this study, UV–Vis spectra of water extracts of soils from an Australian Eucalyptus forest were used to investigate soluble soil C characteristics to 30 cm depth after autumn burning every three or ten years for three decades contrasted with long non-burnt controls. A random forest prediction model was fit to the UV–VIS spectra and the wavelengths important to predicting soluble C investigated using conditional inference trees. The main absorbance of the UV–Vis spectra was in the aromatic region (~280 nm) indicating chemically complex soluble organic matter in both burnt and non-burnt soils. Water extractable organic carbon decreased significantly with depth and was reduced by 3-yearly burning in surface soils. Furthermore, burning, irrespective of treatment, led to significant shifts in the locations of soil spectra peaks, which were consistent with both lower quantities of non-aromatic substances and less substituted aromatic substances in the burnt than non-burnt sites. These results indicate that repeated prescribed fire over decades reduced the quantity of soluble C, and enhanced its aromaticity, potentially reducing biological availability and therefore nutrient cycling in the soils.

Analysis of Alternative Fuel Vehicle (AFV) Adoption Utilizing Different Machine Learning Methods: A Case Study of 2017 NHTS

Alternative fuel vehicles (AFVs) are considered as the one of policies towards the sustainable transportation with low fossil fuel consumption and greenhouse gas emission. However, the demand for AFV is still not clear, due to the complex interaction between influencing factors. This paper aimed to investigate the AFV adoption behavior in order to estimate the AFV penetration in specific regions. 2017 National Household Travel Survey (NHTS), providing the detailed information of AFV users, was utilized to conduct the factor analysis and establish the prediction model. Three groups of variables, including vehicle-related, household-related and individual-related variables, were considered. Additionally, the AFV users are only 4.2% of all the respondents, which is a typical imbalanced distribution. Thus, this paper also introduced the methods to deal with the imbalanced dataset for AFV users and conventional vehicle (CV) users, which was rarely investigated in existing studies. To construct the prediction model, five machine learning methods, Logistic Regression (LR), Naive Bayes (NB), Random Forest (RF), Support Vector Machine (SVM) and Decision Tree (DT), were employed and compared. The performance analysis indicates that the RF model has the best prediction capability among the models. For validation, the RF prediction model was used to develop a AFV penetration map for U.S. by state. This study is believed to have a wide application in government policy making and vehicle manufacturing.

A New Compressional Wave Speed Inversion Method Based on Granularity Parameters

How to improve the prediction accuracy of compressional wave speed has always been one of the basic research subjects in geoacoustics study field. Due to the stability of granularity, whether in the laboratory or in the seabed environment, the regression relationship between compressional wave speed and granularity is an important sound speed inversion method. Machine Learning (ML) provides a new solution for more efficient sound speed prediction systems. In this study, two ML algorithm, Random forest (RF) and Support Vector Regression (SVR), combined with nine granularity parameters (mean grain size, median grain size, skewness, kurtosis, sorting coefficient, gravel, sand, silt, and clay content respectively.) to analysis the effect of granularity property on sound speed. As a result, the sound speed-granularity predictive models were established, and the sound speed accuracy obtained based on the predictive models are higher than that of the regression equations, and the RF model has a higher accuracy than the SVR model. Based on the RF predictive model, the feature selection was conducted and the results show that the most influential parameter of granularity is mean grain size. Furthermore, the RF model can also predict the sound speed with high precision in the absence of partial parameters, which can be a useful tool for ocean engineering and seismic inversion.

Combining multimodal imaging and treatment features improves machine learning-based prognostic assessment in patients with glioblastoma multiforme.

BackgroundFor Glioblastoma (GBM), various prognostic nomograms have been proposed. This study aims to evaluate machine learning models to predict patients' overall survival (OS) and progression‐free survival (PFS) on the basis of clinical, pathological, semantic MRI‐based, and FET‐PET/CT‐derived information. Finally, the value of adding treatment features was evaluated.MethodsOne hundred and eighty‐nine patients were retrospectively analyzed. We assessed clinical, pathological, and treatment information. The VASARI set of semantic imaging features was determined on MRIs. Metabolic information was retained from preoperative FET‐PET/CT images. We generated multiple random survival forest prediction models on a patient training set and performed internal validation. Single feature class models were created including "clinical," "pathological," "MRI‐based," and "FET‐PET/CT‐based" models, as well as combinations. Treatment features were combined with all other features.ResultsOf all single feature class models, the MRI‐based model had the highest prediction performance on the validation set for OS (C‐index: 0.61 [95% confidence interval: 0.51‐0.72]) and PFS (C‐index: 0.61 [0.50‐0.72]). The combination of all features did increase performance above all single feature class models up to C‐indices of 0.70 (0.59‐0.84) and 0.68 (0.57‐0.78) for OS and PFS, respectively. Adding treatment information further increased prognostic performance up to C‐indices of 0.73 (0.62‐0.84) and 0.71 (0.60‐0.81) on the validation set for OS and PFS, respectively, allowing significant stratification of patient groups for OS.ConclusionsMRI‐based features were the most relevant feature class for prognostic assessment. Combining clinical, pathological, and imaging information increased predictive power for OS and PFS. A further increase was achieved by adding treatment features.

Predicting Effectiveness of Generate-and-Validate Patch Generation Systems Using Random Forest

One way to improve practicability of automatic program repair (APR) techniques is to build prediction models which can predict whether an application of a APR technique on a bug is effective or not. Existing prediction models have some limitations. First, the prediction models are built with hand crafted features which usually fail to capture the semantic characteristics of program repair task. Second, the performance of the prediction models is only evaluated on Genprog, a genetic-programming based APR technique. This paper develops prediction models, i.e., random forest prediction models for SPR, another kind of generate- and-validate APR technique, which can distinguish ineffective repair instances from effective repair instances. Rather than handcrafted features, we use features automatically learned by deep belief network (DBN) to train the prediction models. The empirical results show that compared to the baseline models, that is, all effective models, our proposed models can at least improve the F1 by 9% and AUC(area under the receiver operating characteristics curve) by 19%. At the same time, the prediction model using learned features at least outperforms the one using hand-crafted features in terms of F1 by 11%

1726. A Random Forest Prediction Model Accurately Identifies Periods at Increased Risk for Positive Legionella Cultures in a Hospital Water Distribution System

BackgroundHospitals devote considerable resources to water distribution system (WS) surveillance and remediation for Legionella in an effort to reduce risk of transmitting Legionnaires disease (LD). There are no models that accurately predict periods of greatest risk for Legionella culture positivity (cx +) within a WS. Our goal was to build and validate a model based on weather and water parameters that predicted Legionella cx+ in our hospital WS.MethodsOne liter water samples from fixtures at 2 campuses were cultured for Legionella on BCYE plates with cysteine as part of infection prevention protocols. Logistic regression (LR) and random forest (RF) models included daily hospital WS measurements and Pittsburgh FAA weather observation station data. Training and validation used 2014–2015 and 2016–2017 data, respectively. Models predicted a first +cx within 14 day windows.ResultsCxs were defined as + by loop-day, if any cx from within a unique WS loop was + for Legionella on a given day. Of the 7,272 water samples, 5,304 were collected from 16 buildings on 2 campuses in which ≥1 cx + was obtained. A total of 1,262 WS loop-days were collected over 339 unique days from these buildings. Details on training and validation data sets appear in figure. Overall, water was Legionella cx + on 3% of loop-days. Models predicted positivity if risk was >6%. The LR model comprised of independent predictors of cx + had sensitivity/specificity of 44%/80% (AUC: 0.715; misclassification error: 0.21), and PPV/NPV of 9%/97% in the validation data set. The RF model comprised of the same predictors had sensitivity/specificity of 100%/98% (AUC: 1.0; misclassification error: 0.02), and PPV/NPV of 67%/100%. The most important RF variables in the validation data set were WS temperature and minimum pH over the 7 days prior to cx.ConclusionAn RF model using water and weather data was validated as an accurate predictor of new Legionella cx+ within a hospital WS. Most importantly, NPV for the model was 100%, meaning that no positive Legionella cxs were recovered during periods identified as low-risk. The RF model is a powerful tool for most efficiently directing resources to Legionella surveillance and LD prevention. Disclosures All authors: No reported disclosures.

Prediction of Incident Delirium Using a Random Forest classifier.

Delirium is a serious medical complication associated with poor outcomes. Given the complexity of the syndrome, prevention and early detection are critical in mitigating its effects. We used Confusion Assessment Method (CAM) screening and Electronic Health Record (EHR) data for 64,038 inpatient visits to train and test a model predicting delirium arising in hospital. Incident delirium was defined as the first instance of a positive CAM occurring at least 48h into a hospital stay. A Random Forest machine learning algorithm was used with demographic data, comorbidities, medications, procedures, and physiological measures. The data set was randomly partitioned 80% / 20% for training and validating the predictive model, respectively. Of the 51,240 patients in the training set, 2774 (5.4%) experienced delirium during their hospital stay; and of the 12,798 patients in the validation set, 701 (5.5%) experienced delirium. Under-sampling of the delirium negative population was used to address the class imbalance. The Random Forest predictive model yielded an area under the receiver operating characteristic curve (ROC AUC) of 0.909 (95% CI 0.898 to 0.921). Important variables in the model included previously identified predisposing and precipitating risk factors. This machine learning approach displayed a high degree of accuracy and has the potential to provide a clinically useful predictive model for earlier intervention in those patients at greatest risk of developing delirium.

Surficial and Deep Earth Material Prediction from Geochemical Compositions

Prediction of true classes of surficial and deep earth materials using multivariate spatial data is a common challenge for geoscience modelers. Most geological processes leave a footprint that can be explored by geochemical data analysis. These footprints are normally complex statistical and spatial patterns buried deep in the high-dimensional compositional space. This paper proposes a spatial predictive model for classification of surficial and deep earth materials derived from the geochemical composition of surface regolith. The model is based on a combination of geostatistical simulation and machine learning approaches. A random forest predictive model is trained, and features are ranked based on their contribution to the predictive model. To generate potential and uncertainty maps, compositional data are simulated at unsampled locations via a chain of transformations (isometric log-ratio transformation followed by the flow anamorphosis) and geostatistical simulation. The simulated results are subsequently back-transformed to the original compositional space. The trained predictive model is used to estimate the probability of classes for simulated compositions. The proposed approach is illustrated through two case studies. In the first case study, the major crustal blocks of the Australian continent are predicted from the surface regolith geochemistry of the National Geochemical Survey of Australia project. The aim of the second case study is to discover the superficial deposits (peat) from the regional-scale soil geochemical data of the Tellus Project. The accuracy of the results in these two case studies confirms the usefulness of the proposed method for geological class prediction and geological process discovery.

Quantitative screening of serum protein biomarkers by reverse phase protein arrays.

Screening biomarkers in serum samples for different diseases has always been of great interest because it presents an early, reliable, and, most importantly, noninvasive means of diagnosis and prognosis. Reverse phase protein arrays (RPPAs) are a high-throughput platform that can measure single or limited sets of proteins from thousands of patients' samples in parallel. They have been widely used for detection of signaling molecules involved in diseases, especially cancers, and related regulation pathways in cell lysates. However, this approach has been difficult to adapt to serum samples. Previously, we developed a sensitive method called the enhanced protein array to quantitatively measure serum protein levels from large numbers of patient samples. Here, we further refine the technology on several fronts: 1. simplifying the experimental procedure; 2. optimizing multiple parameters to make the assay more robust, including the support matrix, signal reporting method, background control, and antibody validation; and 3. establishing a method for more accurate quantification. Using this technology, we quantitatively measured the expression levels of 10 proteins: alpha-fetoprotein (AFP), beta 2 microglobulin (B2M), Carcinoma Antigen 15-3(CA15-3), Carcinoembryonic antigen (CEA), golgi protein 73 (GP73), Growth differentiation factor 15 (GDF15), Human Epididymis Protein 4 (HE4), Insulin Like Growth Factor Binding Protein 2 (IGFBP2), osteopontin (OPN) and Beta-type platelet-derived growth factor receptor (PDGFRB) from serum samples of 132 hepatocellular carcinoma (HCC) patients and 78 healthy volunteers. We found that 6 protein expression levels are significantly increased in HCC patients. Statistical and bioinformatical analysis has revealed decent accuracy rates of individual proteins, ranging from 0.617 (B2M) to 0.908 (AFP) as diagnostic biomarkers to distinguish HCC from healthy controls. The combination of these 6 proteins as a specific HCC signature yielded a higher accuracy of 0.923 using linear discriminant analysis (LDA), logistic regression (LR), random forest (RF) and support vector machine (SVM) predictive model analyses. Our work reveals promise for using reverse phase protein arrays for biomarker discovery and validation in serum samples.