Gradient-boosted Model Research Articles

MP64-01 PREDICTING RISK OF SIDE-SPECIFIC EXTRAPROSTATIC EXTENSION IN MEN WITH PROSTATE CANCER USING EXPLAINABLE ARTIFICIAL INTELLIGENCE

You have accessJournal of UrologyProstate Cancer: Localized: Surgical Therapy VI (MP64)1 Sep 2021MP64-01 PREDICTING RISK OF SIDE-SPECIFIC EXTRAPROSTATIC EXTENSION IN MEN WITH PROSTATE CANCER USING EXPLAINABLE ARTIFICIAL INTELLIGENCE Jethro Kwong, Adree Khondker, Christopher Tran, Emily Evans, Amna Ali, Munir Jamal, Thomas Short, Frank Papanikolaou, John Srigley, and Andrew Feifer Jethro KwongJethro Kwong More articles by this author , Adree KhondkerAdree Khondker More articles by this author , Christopher TranChristopher Tran More articles by this author , Emily EvansEmily Evans More articles by this author , Amna AliAmna Ali More articles by this author , Munir JamalMunir Jamal More articles by this author , Thomas ShortThomas Short More articles by this author , Frank PapanikolaouFrank Papanikolaou More articles by this author , John SrigleyJohn Srigley More articles by this author , and Andrew FeiferAndrew Feifer More articles by this author View All Author Informationhttps://doi.org/10.1097/JU.0000000000002104.01AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: Current machine learning (ML) models are limited by poor interpretability, precluding their routine use in planning nerve-sparing at radical prostatectomy (RP). We aimed to leverage explainable artificial intelligence techniques to provide accurate, interpretable, and personalized predictions for side-specific extraprostatic extension (ssEPE). METHODS: A retrospective sample of 900 prostatic lobes (450 patients) from RP specimens at our institution between 2010 and 2020, was used as the training cohort. Features (ie: variables) included patient demographics, clinical, sonographic, and site-specific data from transrectal ultrasound-guided prostate biopsy. The label (ie: outcome) of interest was the presence of ssEPE in the prostatectomy specimen. A ten-fold stratified cross-validation method was performed to train a gradient-boosted model, optimize hyperparameters, and for internal validation. Our model was further externally validated using a testing cohort of 122 lobes (61 patients) from RP specimens at a separate institution between 2016 and 2020. An existing model from the literature which has the highest performance for predicting ssEPE was selected as the baseline model for comparison. Discriminative capability was quantified by area under receiver-operating-characteristic (AUROC) and precision-recall curve (AUPRC). Clinical utility was determined by decision curve analysis. Shapley Additive exPlanations were used to interpret the ML model’s predictions. RESULTS: The incidence of ssEPE in the training and testing cohorts were 30.7 and 41.8%, respectively. Our model outperformed the baseline model with a mean AUROC of 0.81 vs 0.75 (p<0.01) and mean AUPRC of 0.69 vs 0.60, respectively, on cross-validation of the training cohort. Similarly, our model performed favourably on the testing cohort with an AUROC of 0.81 vs 0.76 (p=0.03) and AUPRC of 0.78 vs 0.72. On decision curve analysis, our model achieved a higher net benefit for threshold probabilities between 0.15 to 0.65. A web application incorporating our model was developed in which de-identified patient data can be inputted to generate an individualized ssEPE prostate map with annotated explanations to highlight which features had the greatest impact on model predictions (www.ssepe.ml). CONCLUSIONS: We have developed a user-friendly application that enables physicians without prior ML experience to assess ssEPE risk and understand the factors driving these predictions to aid surgical planning and patient counselling. Source of Funding: None © 2021 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 206Issue Supplement 3September 2021Page: e1110-e1110 Advertisement Copyright & Permissions© 2021 by American Urological Association Education and Research, Inc.MetricsAuthor Information Jethro Kwong More articles by this author Adree Khondker More articles by this author Christopher Tran More articles by this author Emily Evans More articles by this author Amna Ali More articles by this author Munir Jamal More articles by this author Thomas Short More articles by this author Frank Papanikolaou More articles by this author John Srigley More articles by this author Andrew Feifer More articles by this author Expand All Advertisement Loading ...

A predictive algorithm using clinical and laboratory parameters may assist in ruling out and in diagnosing MDS

We present a noninvasive Web-based app to help exclude or diagnose myelodysplastic syndrome (MDS), a bone marrow (BM) disorder with cytopenias and leukemic risk, diagnosed by BM examination. A sample of 502 MDS patients from the European MDS (EUMDS) registry (n > 2600) was combined with 502 controls (all BM proven). Gradient-boosted models (GBMs) were used to predict/exclude MDS using demographic, clinical, and laboratory variables. Area under the receiver operating characteristic curve (AUC), sensitivity, and specificity were used to evaluate the models, and performance was validated using 100 times fivefold cross-validation. Model stability was assessed by repeating its fit using different randomly chosen groups of 502 EUMDS cases. AUC was 0.96 (95% confidence interval, 0.95-0.97). MDS is predicted/excluded accurately in 86% of patients with unexplained anemia. A GBM score (range, 0-1) of less than 0.68 (GBM < 0.68) resulted in a negative predictive value of 0.94, that is, MDS was excluded. GBM ≥ 0.82 provided a positive predictive value of 0.88, that is, MDS. The diagnosis of the remaining patients (0.68 ≤ GBM < 0.82) is indeterminate. The discriminating variables: age, sex, hemoglobin, white blood cells, platelets, mean corpuscular volume, neutrophils, monocytes, glucose, and creatinine. A Web-based app was developed; physicians could use it to exclude or predict MDS noninvasively in most patients without a BM examination. Future work will add peripheral blood cytogenetics/genetics, EUMDS-based prospective validation, and prognostication.

Abstract 336: Gut microbiome predicts response to CDK4/6 inhibitor and immune check point inhibitor combination in patients with hormone receptor positive metastatic breast cancer

Abstract Background: CDK4/6 inhibitors (CDK4/6i) have been shown to modulate immune responses in the preclinical setting. Palbociclib is a front-line therapy for hormone receptor positive (HR+) metastatic breast cancer (MBC), and the combination with immune check point inhibitors (ICI) is increasingly being studied. An ongoing phase I trial was designed to test the safety and efficacy of palbociclib, pembrolizumab, and letrozole in postmenopausal women with HR+ HER2- MBC (NCT02778685). Stool microbiome was collected and analyzed. The aim of this study is to determine the association of gut microbiota and response using metagenome sequencing data through a machine learning model. Methods: Postmenopausal women with histologically proven stage IV HR+ HER2- MBC were enrolled. Stool samples were collected at baseline and during treatment for analysis using metagenome sequencing. Response per RECIST 1.1 were grouped into: responders (complete response or partial response) and non-responders (stable disease or progressive disease). Using metagenomic relative abundance data, a gradient-boosted tree model was developed with leave-one-patient-out cross validations to predict patient response at baseline. Kruskal-Wallis tests were used to assess the differences of the most important microbiota relative abundance (generated from the machine learning model) between responders and non-responders. Results: Forty-seven stool samples from 11 patients were collected at baseline and during treatment, and metagenome sequencing was performed. For predictive modeling, the validation Area Under the Receiver Operating Characteristic Curve (AUROC) and Area Under the Precision Recall Curve (AUPRC) were 0.71 and 0.83, respectively. Among the top 5 features from the model, patients with a larger relative abundance of Gemmiger formicillis have increased probability of responding to therapy (p&amp;lt;0.05 uncorrected), whereas an increase of Bacteroides vulgatus have reduced probability of responding to therapy (p&amp;lt;0.05 uncorrected). The relative abundance of Gemmiger formicillis at baseline has been reported in recent studies to be higher in patients with long term benefit (Aarnoutse et al. 2017). Thus, there is increasing evidence that gut microbiota plays a role in the pathogenesis of breast cancer and response of MBC patients to systemic cancer therapy. Citation Format: Chi Wah Wong, Susan E. Yost, Jin Sun Lee, Sarah K. Highlander, Yuan Yuan. Gut microbiome predicts response to CDK4/6 inhibitor and immune check point inhibitor combination in patients with hormone receptor positive metastatic breast cancer [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 336.

Can We Diagnose MDS without Bone Marrow Examination? a Proposed EUMDS-Based Non-Invasive Diagnostic Model

Background: Bone marrow examination (BME) is the gold standard for diagnosing myelodysplastic syndromes (MDS). Because it is invasive, many physicians and patients (pts) avoid BME, and the MDS Diagnosis is missed. Recently (Oster et al. Int MDS, Valencia, May 2017), we proposed a non-invasive MDS diagnostic model, using logistic regression (LoR) with a small group of patients (47 BME-proven MDS vs 63 controls with no MDS by BME). The variables gender, age, Hb, MCV, PLT and WBC were found to be the most important. Putting the variables in the formula (Fig 1) yielded a score Y, reflecting the probability of having MDS. The area under the curve (AUC) for the ROC curve was 0.748, indicating a good model, but one insufficient to provide a single Y, above which MDS could be established clearly. We therefore established two cut-off values: Pts with Y≥0.633 were found to have high likelihood for MDS, and those with Y≤0.288, a low likelihood of having MDS. Those pts (about 50%) where Y fell between the two cutoffs were indeterminate. In this study, the large European MDS Registry (EUMDS) database was used to validate and improve the model.Methods: The EUMDS registry contains data of 2200 MDS pts. A random sample (age≥65) of 131 MDS pts from the registry, and 115 controls with no MDS were added to the original dataset, providing 178 in each group. Two models, LoR with covariates selected using the Lasso method, and gradient-boosted (GB) were assessed. AUC was used to evaluate the models. Model stability was assessed by repeating the fit of the models using different randomly chosen groups of 131 EUMDS pts that replaced the first 131 EUMDS pt group (random augmentation). Another analysis was performed using all the EUMDS pts.Results: For the LoR model, MCV, PLT and WBC were selected, and the AUC was 0.75 (95% CI 0.70 - 0.80, Fig 2). For the GB model, the AUC was 0.87 (95% CI 0.84-0.91, Fig 2). The most influential variables were MCV and WBC. 50 repetitions of the random augmentation led to a mean AUC of 0.87 (IQR 0.86-0.88), attesting to the stability of the model. Using the entire 2200 pts EUMDS data set also provided consistent results.Conclusions: A classical LoR model using the large EUMDS dataset, confirmed our original finding that using simple parameters, MDS can be diagnosed or excluded in a large portion of pts without BME. Using the complex GB model, that not only combines the chosen variables, but also takes into account the interactions among them, the AUC approaches 90%. We will continue to improve and validate the model. In our original model using two cutoff points, we were able to establish or rule out MDS with high accuracy in approximately 50% of the patients. However, about half of the patients fell within the middle ‘indeterminate’ zone. With this improved AUC, it is likely that the ‘indeterminate’ zone will shrink and we will be able to establish or rule out MDS in an even greater proportion of patients without an invasive BME. Moreover, adding more controls will allow increasing the number of variables (e.g. folic acid, B12) that can have an additional impact on the model and its diagnostic capability. [Display omitted] DisclosuresFenaux:Celgene: Honoraria, Research Funding; Astex: Honoraria, Research Funding; Janssen: Honoraria, Research Funding; Amgen: Honoraria, Research Funding; Novartis: Honoraria, Research Funding; Amgen: Honoraria, Research Funding; Novartis: Honoraria, Research Funding; Janssen: Honoraria, Research Funding; Celgene: Honoraria, Research Funding; Astex: Honoraria, Research Funding. Sanz:Gamida Cell: Research Funding. Germing:Novartis: Honoraria, Research Funding; Janssen: Honoraria; Celgene: Honoraria, Research Funding. Almeida:Novartis: Consultancy; Celgene: Consultancy; Servier: Consultancy; Alexion: Honoraria; Bristol Meyer Squibb: Honoraria.

A novel approach for estimation of aboveground biomass of a carbon-rich mangrove site in India

Mangroves can play a crucial part in climate change mitigation policies due to their high carbon-storing capacity. However, the carbon sequestration potential of Indian mangroves generally remained unexplored to date. In this study, multi-temporal Sentinel-1 and 2 data-derived variables were used to estimate the AGB of a tropical carbon-rich mangrove forest of India. Ensemble prediction of multiple machine learning algorithms, including Random Forest (RF), Gradient Boosted Model (GBM), and Extreme Gradient Boosting (XGB), were used for AGB prediction. The multi-temporal dataset was used in two different ways to find the most suitable method of using them. The results of the analysis showed that the modeling field measured AGB with individual date data values results in estimates with root mean square errors (RMSE) ranging from 149.242 t/ha for XGB to 151.149 t/ha for the RF. Modeling AGB with the average and percentile metrics of the multi-temporal image stack improves the prediction accuracy of AGB, with RMSE ranging from 81.882 t/ha for the XGB to 74.493 t/ha for the RF. The AGB modeling using ensemble prediction showed further improvement in accuracy with an RMSE of 72.864 t/ha and normalized RMSE of 11.38%. In this study, the intra-seasonal variation of Sentinel-1 and 2 data for mangrove ecosystems was explored for the first time. The variations in remotely sensed variables could be attributed mainly to soil moisture availability and rainfall in the mangrove ecosystem. The efficiency of Sentinel-1 and 2 data-derived variables and ensemble prediction of machine learning models for Indian mangroves were also explored for the first time. The methodologies established in this study can be used in the future for accurate prediction and repeated monitoring of AGB for mangrove ecosystems.

Explainable prediction of survival using clinical, molecular, and radiomic profiles in recurrent high-grade glioma patients treated with CAR T-cell therapy.

104 Background: High-grade glioma (HGG) is an aggressive heterogeneous primary CNS neoplasm with high recurrence rate and poor survival. Multiple ongoing clinical trials are leveraging targeted molecular and immunologic therapeutics (e.g., pembrolizumab, Chimeric Antigen Receptor [CAR] T-cell therapy) in effort to improve survival. Explainable predictive models have shown value in identifying biomarkers predictive of treatment response as well as informing prognosis. In this study, we developed an explainable machine learning model leveraging clinical, molecular and radiomic (imaging) features to predict overall survival in patients suffering from HGG treated with CAR T-cell therapy. Methods: In this IRB-approved phase 1 clinical trial, 60 patients (39 males, median age = 49) suffering from HGG underwent surgical resection and CAR T-cell therapy1. All patients underwent baseline MRI scans prior to both surgical resection and CAR T-cell administration in the resection cavity. Using contrast-enhanced T1-weighted MRIs, we segmented the enhancing tumor (ET) and generated radiomic features. For predictive modeling, we incorporated the following features: Age, gender, race, ethnicity, histology, tumor grade (WHO), IL-13 receptor alpha 2 (IL-13Rα2) expression (H score), unifocal or multifocal lesions, tumor location (lobe), shape-based radiomics (tumor volume, surface area, and sphericity). We utilized gradient-boosted tree models to classify whether survival is above or below 180 days with two-loop nested cross-validation. For the inner validation loop, we optimized the model with hyper-parameter tuning. For the outer validation loop, we tested the optimal model on the hold-out data and the predictions were used as survival scores (0 – 1). Larger scores imply better predicted survival. For prediction explanations, we adopted the Shapley additive explanation (SHAP) framework. Results: The outer validation loop Area Under the Receiver Operating Characteristic Curve and Area under the Precision-Recall Curve were 0.76 and 0.81, respectively. Among the top five most important features calculated from SHAP; patients with larger tumor surface area, tumor volume and age have reduced survival scores while patients with larger IL-13Rα2 and tumor sphericity have increased survival scores. We stratified the patients into two distinct prognostic sub-groups (30 patients each group) using the survival scores obtained from the outer loop, with a log-rank test p &lt; 0.01. Conclusions: In patients with HGG treated with CAR T-cell therapy, we found that tumor surface area/volume and age are inversely related to survival while increased IL-13Rα2 expression and tumor sphericity were positive predictor of survival. Our model can potentially be used to optimize clinical trial enrollment through more precise patient screening and treatment planning.

The performance of some machine learning approaches and a rich context model in student answer prediction

Web-based learning systems with adaptive capabilities to personalize content are becoming nowadays a trend in order to offer interactive learning materials to cope with a wide diversity of students attending online education. Learners’ interaction and study practice (quizzing, reading, exams) can be analyzed in order to get some insights into the student’s learning style, study schedule, knowledge, and performance. Quizzing might be used to help to create individualized/personalized spaced repetition algorithm in order to improve long-term retention of knowledge and provide efficient learning in online learning platforms. Current spaced repetition algorithms have pre-defined repetition rules and parameters that might not be a good fit for students’ different learning styles in online platforms. This study uses different machine learning models and a rich context model to analyze quizzing and reading records from e-learning platform called Hypocampus in order to get some insights into the relevant features to predict learning outcome (quiz answers). By knowing the answer correctness, a learning system might be able to recommend personalized repetitive schedule for questions with maximizing long-term memory retention. Study results show that question difficulty level and incorrectly answered previous questions are useful features to predict the correctness of student’s answer. The gradient-boosted tree and XGBoost models are best in predicting the correctness of the student’s answer before answering a quiz. Additionally, some non-linear relationship was found between the reading learning material behavior in the platform and quiz performance that brings added value to the accuracy for all used models.

Abstract PS5-41: Machine learning model of gut microbiota predicts neratinib induced diarrhea in patients with breast cancer

Abstract Background: Neratinib is a potent small molecule tyrosine kinase inhibitor (TKI) of human epidermal growth factor receptors (HER1,2,4). One of the major side effects of neratinib is diarrhea. The human gut contains a dense microbiome ecosystem that is essential in maintaining a healthy host physiology, and its disruption may lead to increased risk of toxicities from cancer therapy. In this study, we aimed to develop a machine learning model based on analysis of gut microbiota data to predict neratinib-induced diarrhea. Methods: Patients were enrolled in a phase II trial evaluating safety and tolerability of neratinib in older adults with HER2+ breast cancer (NCT02673398). Neratinib was administered as single agent, 240 mg oral daily in a 28-day cycle. Stool samples were collected at baseline and during treatment for 16S rRNA gene sequencing. Using microbial relative abundance data, we developed gradient-boosted tree models with two nested loops of cross validations to classify whether diarrhea would occur or not after treatment onset. For the inner validation loop, we used ten-fold cross validation to determine the optimal model from hyper-parameters including regularization. For the outer validation loop, we utilized a leave-one-patient-out cross validation to test this model on the hold-out patient’s baseline data and the predictions were used for model assessment. Results: A total of 11 patients and 50 longitudinal stool samples were collected. The median age was 66 years. 73% developed grade ≥ 1 diarrhea attributed to neratinib. Shannon diversity index of gut microbiome was not associated with diarrhea. For predictive modeling, the outer validation loop Area Under the Receiver Operating Characteristic Curve (AUROC) and Area Under the Precision Recall Curve (AUPRC) were 0.92 and 0.97, respectively. The two most important taxa predictive of protection from diarrhea were Ruminiclostridium 9, and Bacteroides sp. HPS0048. We found that patients with a larger relative abundance of Ruminiclostridium 9 and Bacteroides sp. HPS0048 have reduced risk of neratinib-related diarrhea. Conclusions: The machine learning model can identify breast cancer patients at risk of diarrhea prior to neratinib use. Future studies are required to validate this finding. Citation Format: Chi Wah Wong, Susan E. Yost, Jin Sun Lee, John D. Gillece, Megan Folkerts, Lauren Reining, Sarah K. Highlander, Zahra Eftekhari, Joanne Mortimer, Yuan Yuan. Machine learning model of gut microbiota predicts neratinib induced diarrhea in patients with breast cancer [abstract]. In: Proceedings of the 2020 San Antonio Breast Cancer Virtual Symposium; 2020 Dec 8-11; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2021;81(4 Suppl):Abstract nr PS5-41.

A Hidden Markov Tree Model for Flood Extent Mapping in Heavily Vegetated Areas based on High Resolution Aerial Imagery and DEM: A Case Study on Hurricane Matthew Floods

ABSTRACT Flood extent mapping plays a crucial role in disaster management and national water forecasting. In recent years, high-resolution optical imagery becomes increasingly available with the deployment of numerous small satellites and drones. However, analysing such imagery data to extract flood extent poses unique challenges due to the rich noise and shadows, obstacles (e.g. tree canopies, clouds), and spectral confusion between pixel classes (flood, dry) due to spatial heterogeneity. Existing machine-learning techniques often focus on spectral and spatial features from raster images without fully incorporating the geographic terrain within classification models. In contrast, we recently proposed a novel machine-learning model called geographical hidden Markov tree (HMT) that integrates spectral features of pixels and topographic constraint from Digital Elevation Model (DEM) data (i.e. water flow directions) in a holistic manner. This paper evaluates the model through case studies on high-resolution aerial imagery from National Oceanic and Atmospheric Administration (NOAA) National Geodetic Survey (NGS) together with DEM. Three scenes are selected in heavily vegetated floodplains near the cities of Grimesland and Kinston in North Carolina during Hurricane Matthew floods in 2016. Results show that the proposed HMT model outperforms several state of the art machine-learning algorithms (e.g. random forests, gradient-boosted model) by an improvement of F-score (the harmonic mean of the user’s accuracy and producer’s accuracy) from around 70% to 80% to over 95% on our datasets.

Defining the relative contribution of health care environmental components to patient outcomes in the model of 30-day readmission after coronary artery bypass graft (CABG)

BackgroundPatient outcomes following health care interventions may be dependent on a variety of factors: patient, surgeon, hospital, information technology, and temporal, cultural, and socioeconomic factors, among others. In this study, we characterize the relative contribution of each of these factors using a model of 30-day readmission following coronary artery bypass graft. MethodsThe Healthcare Cost and Utilization Project, the American Hospital Association Annual Health Survey Databases, the Healthcare Information and Management Systems Society, and the Distressed Communities Index from 2010 to 2013 were linked for Florida, Iowa, Massachusetts, Maryland, New York, and Washington. Logistic regression, random forest, decision tree, gradient boosting, k-nearest-neighbors classification, and XGBoost tree models were implemented. Modeling results were compared on the basis of predictive accuracy, sensitivity, specificity, and area under the curve. Decision tree performed best and was selected for further analysis. A gradient-boosted model was used to quantify factor contribution. ResultsThe model had 45,352 patients, 54,096 admissions, and a 16.2% 30-day readmission rate after coronary artery bypass graft. The top 10 predictors were disposition at discharge, number of chronic conditions, total procedures, median household income, adults without high school diplomas, primary payer method, Agency for Healthcare Research and Quality comorbidity: renal failure, patient location (urban-rural), admission type, and age categories. The top 3 socioeconomic predictors were estimated state median household income, adults without high school diplomas, and patient location (urban versus rural designation). The relative contribution of patient/temporal, socioeconomic, hospital information technology, and hospital factors to readmission is 83.45%, 5.71%, 6.34%, and 4.31%, respectively. ConclusionIn this model, the contribution of socioeconomic factors is substantive but lags significantly behind patient/temporal factors. With ever increasing availability of data, identification of contributors to patient outcomes within the overall health care macroenvironment will allow prioritization of interventions.

Development and validation of a model to predict survival in colorectal cancer using a gradient-boosted machine

ObjectiveThe success of treatment planning relies critically on our ability to predict the potential benefit of a therapy. In colorectal cancer (CRC), several nomograms are available to predict different outcomes...

Abstract 866: Radiomic prediction of survival in recurrent high-grade glioma patients treated with CAR T-cell therapy

Abstract Introduction: High-grade glioma (HGG) is the most common subtype of primary brain tumors with high recurrence rate and poor survival. The emergence of targeted molecular and cellular therapies (e.g., pembrolizumab, Chimeric Antigen Receptor [CAR] T-cell therapy) are potentially promising in improving overall survival. Due to increased intratumor heterogeneity and inhomogeneous treatment response, there is an unmet need of imaging biomarkers predictive of treatment response and survival. Radiomics using machine learning methods have shown promise in predicting treatment response in various solid tumors, including HGG. In this study, we compare the survival prediction performance using machine learning models with different radiomic features individually derived from T1- and T2-weighted MR images in patients suffering from HGG treated with CAR-T cell therapy. Methods: In this IRB-approved phase 1 clinical trial, 61 patients (39 males, median age = 49) suffering from recurrent HGG underwent surgical resection and CAR T-cell therapy1. All patients underwent baseline MRI scans prior to both surgical resection and CAR T-cell administration in the resection cavity. For patients with a complete set of T1- and T2-weighted MRIs (n = 50), we generated segmentations in a semi-automated manner, labeling with each tumorous voxel as either contrast-enhanced tumor (ET), non-enhancing tumor (NET) and edema. From each tumor label, we extracted shape-based, texture-based and image-filtered radiomic features2. We utilized gradient-boosted tree models (lightGBM) to classify whether survival is above or below group median (188 days) by using two nested loops of 10-fold cross validations each. For the inner validation loop, we determined the optimal model from hyper-parameters including regularization. For the outer validation loop, we tested this model on the hold-out data and the predictions were used as radiomic risk scores. Results: For each of the ET, NET, and edema tumor ROIs, we extracted 1313 radiomic features for predictive modeling. The outer validation loop Area Under the Receiver Operating Characteristic Curve (AUC) for ET, NET and edema were 0.55, 0.70, and 0.46, respectively, suggesting that radiomic features calculated from NET voxels are the most predictive of survival compared to features from ET and edema voxels. We also stratified the patients into two distinct prognostic sub-groups (25 patients each group) using the NET radiomic risk scores obtained from the outer validation loop, with a log-rank test p-value of 0.01. Conclusions: In patients suffering from recurrent HGG who were treated with CAR T-cell therapy, we found that radiomic features derived from NET voxels are predictive of survival while the other two tumorous voxel types (ET and edema) are not. Further work is needed to incorporate clinical and molecular features that also may be predictive of survival. 1N Engl J Med 2016; 375:2561-2569. 2Cancer Research, 77(21), e104-e107 Citation Format: Chi Wah Wong, Sohaib Naim, Vincent La, Seth Michael Hilliard, Eemon Tizpa, Rashi Ranjan, Hannah Jade Young, Kimberly Jane Bonjoc, Aleksandr Filippov, Saman Tabassum Khan, Christine Brown, Behnam Badie, Ammar Ahmed Chaudhry. Radiomic prediction of survival in recurrent high-grade glioma patients treated with CAR T-cell therapy [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 866.

On experimenting large dataset for visualization using distributed learning and tree plotting techniques

Visualization as one major field making up data science has played significant roles in data exploration. With visualization at the center of every data analysis and application, exploratory analysis has proved the basis for which data analyst comparatively implement what-if scenario before and after processing. Interesting patterns generated from models visualized is very helpful in fast decision-making, model tuning and optimization. Although, conventional methods such as histogram, pie chart, box plot and bar graph are in most occasions not adequate enough to effectively convey the interesting pattern to be mined in large dataset. This paper therefore, presents a tree-plot approach which make use of an In-memory node mechanism from the h2o package to place the large dataset in memory. A Gradient Boosted Model (GBM) from same was implemented as the underlying learning algorithm to build the tree model, while the modeled trees were plotted using plotting techniques in data.tree. Execution process time, AUC, MSE and RMSE results obtained provide basis for evaluating how well the data was trained and for visualizing the modeled tree. It further substantial how a learner algorithm could work with a plotting method with less computational cost.

Risk Factors for Heart Failure with Preserved or Reduced Ejection Fraction Among Medicare Beneficiaries: Application of Competing Risks Analysis and Gradient Boosted Model.

BackgroundThe differential impact of various demographic characteristics and comorbid conditions on development of heart failure (HF) with preserved (pEF) and reduced ejection fraction (rEF) is not well studied among the elderly.MethodsUsing Medicare claims data linked to electronic health records, we conducted an observational cohort study of individuals ≥65 years of age without HF. A Cox proportional hazards model accounting for competing risk of HFrEF and HFpEF incidence was constructed. A gradient-boosted model (GBM) assessed the relative influence (RI) of each predictor in the development of HFrEF and HFpEF.ResultsAmong 138,388 included individuals, 9701 developed HF (incidence rate = 20.9 per 1000 person-years). Males were more likely to develop HFrEF than HFpEF (HR = 2.07, 95% CI: 1.81–2.37 vs. 1.11, 95% CI: 1.02–1.20, P for heterogeneity <0.01). Atrial fibrillation and pulmonary hypertension had stronger associations with the risk of HFpEF (HR = 2.02, 95% CI: 1.80–2.26 and 1.66, 95% CI: 1.23–2.22) while cardiomyopathy and myocardial infarction were more strongly associated with HFrEF (HR = 4.37, 95% CI: 3.21–5.97 and 1.94, 95% CI: 1.23–3.07). Age was the strongest predictor across all HF subtypes with RI from GBM >35%. Atrial fibrillation was the most influential comorbidity for the development of HFpEF (RI = 8.4%) while cardiomyopathy was the most influential comorbidity for the development of HFrEF (RI = 20.7%).ConclusionThese findings of heterogeneous relationships between several important risk factors and heart failure types underline the potential differences in the etiology of HFpEF and HFrEF.

Assessing the Applicability of Random Forest, Stochastic Gradient Boosted Model, and Extreme Learning Machine Methods to the Quantitative Precipitation Estimation of the Radar Data: A Case Study to Gwangdeoksan Radar, South Korea, in 2018

Machine learning algorithms should be tested for use in quantitative precipitation estimation models of rain radar data in South Korea because such an application can provide a more accurate estimate of rainfall than the conventional ZR relationship-based model. The applicability of random forest, stochastic gradient boosted model, and extreme learning machine methods to quantitative precipitation estimation models was investigated using case studies with polarization radar data from Gwangdeoksan radar station. Various combinations of input variable sets were tested, and results showed that machine learning algorithms can be applied to build the quantitative precipitation estimation model of the polarization radar data in South Korea. The machine learning-based quantitative precipitation estimation models led to better performances than ZR relationship-based models, particularly for heavy rainfall events. The extreme learning machine is considered the best of the algorithms used based on evaluation criteria.

Gradient Boosted Models for Enhancing Fatigue Cracking Prediction in Mechanistic-Empirical Pavement Design Guide

AbstractThis study developed a gradient boosted model (GBM) to enhance the fatigue cracking predictive performance of transfer functions in the mechanistic-empirical pavement design guide (MEPDG). ...

MDS Diagnosis: Many Patients May Not Require Bone Marrow Examination

Background: Myelodysplastic syndromes (MDS) are diagnosed with a bone marrow examination (BME), an invasive procedure that patients (pts) would rather avoid. Earlier (Oster et al, Leuk Lymph 2018) we developed a logistic regression (LoR) model to diagnose MDS by incorporating 6 variables (age, gender, Hb, WBC, PLT, MCV) into a formula (Figure 1A). We improved the model using data from 178 MDS pts (47 from Tel Aviv, 131 from the EUMDS registry), and 178 controls (ASH 2017). Here we significantly improve the model using a much larger dataset (501 pts; 501 controls) and additional variables.Methods: The EUMDS registry contains data on 2600 BME-proven MDS pts. A random sample of 501 MDS pts from the registry was combined with 501 controls with no MDS (ruled out with BME). Gradient-boosted models (GBM) were used to predict having or not having MDS, using the variables age, gender, Hb, WBC, PLT, MCV, neutrophils, monocytes, glucose, and creatinine (Figure 1B). Area under the ROC curve (AUC), sensitivity and specificity were used to evaluate the models, and model performance was validated by using 100 times 5-fold cross-validation. Model stability was also assessed by repeating the fit of the models using different randomly chosen groups of 501 EUMDS pts as cases.Results: The AUC was 0.97 (95% CI 0.96-0.98, Figure 2), compared with an AUC of 0.87 (0.84-0.91) achieved previously. Under cross-validation, AUC was 89%. Maximizing the sum of sensitivity and specificity led to sensitivity of 88% and specificity of 95%. This means we can calculate a threshold “GBM score,” assigning a subject to “MDS” or “no MDS” status with a specificity of 95% and a sensitivity of 88%. Alternatively, we can set two GBM score thresholds G1 & G2, where a GBM score > G2 provides 95% specificity and a score < G1 provides 95% sensitivity. A score between these two cutoffs gives an indeterminate probability of disease. Only 24% of our MDS patients and 15% of our control patients fall into this indeterminate region, compared with about 50% in our earlier model. The most influential variables were MCV, creatinine and neutrophils. Repeated random choice of cases from the EUMDS registry led to stable results.Conclusions: Using easily accessible parameters, MDS can be diagnosed or excluded non-invasively with high accuracy in a substantially large portion of patients. While the Logistic Regression model (Figure 1A) can be used with a relatively simple formula, the Gradient Boosted Model (Figure 1B) is more complex. The GBM combines the variables and the interactions among them, achieving an AUC that represents an excellent predictive ability and a considerable improvement over the previous model. Adding peripheral blood cytogenetic/genetic information could further improve non-invasive MDS diagnosis, and obviate the need for bone marrow examination in many patients. We continue to improve and validate the model. An on-line calculator/app for use in a clinical setting is being developed and will be presented. [Display omitted] DisclosuresSmith:Jazz Pharmaceuticals: Research Funding; Johnson & Johnson: Research Funding; Novartis: Research Funding; Gilead Sciences: Consultancy. Fenaux:Celgene: Honoraria, Research Funding; Otsuka: Honoraria, Research Funding; Jazz: Honoraria, Research Funding; Janssen: Honoraria, Research Funding. Stauder:Novartis: Honoraria, Membership on an entity's Board of Directors or advisory committees; Teva: Research Funding; Celgene: Honoraria, Membership on an entity's Board of Directors or advisory committees. Germing:Novartis: Honoraria, Research Funding; Celgene: Honoraria, Research Funding; Janssen: Honoraria.

Method Based on Floating Car Data and Gradient-Boosted Decision Tree Classification for the Detection of Auxiliary Through Lanes at Intersections

The rapid detection of information on continuously changing intersection auxiliary through lane is a major task of lane-level navigation data updates. However, existing lane number detection methods possess long update cycles and high computational costs. Therefore, this study proposes a novel method based on floating car data (FCD) for the detection of auxiliary through lane changes at road intersections. First, roads near intersections are divided into three sections and the spatial distribution characteristics of the FCD of each section are analyzed. Second, the FCD is preprocessed to obtain a standardized FCD dataset by removing redundant data through an improved amplitude-limiting average filtering method. Third, a basic classifier for the number of lanes is constructed. Fourth, the final number of lanes of the road section is determined by combining the basic classifier and the gradient-boosted decision tree model. Finally, the presence of an auxiliary through lane at the intersection is determined in accordance with the change in the number of intersection lanes. The method was tested using data for a road in Wuchang District, Wuhan City. Experimental results show that this method can rapidly obtain auxiliary through lane information from the FCD and is superior to other classification methods.

Model-specification uncertainty in future area burned by wildfires in Canada

Climate change will drive significant changes in annual area burned (burning rates) in the boreal forest although the trends, which are highly variable among studies, which may be caused by model specifications. In order to investigate this, we used 100 models predicting burning rates that are based on two predictor datasets (annual or 30-year averages) and five statistical algorithms (generalised linear model (GLM), random forest, gradient-boosted model (GBM), regression trees, multivariate adaptive regression splines (MARS)) to build a consensus model projecting future burning rates in boreal Canada with three global climate models (GCMs) (CanESM2, HadGEM and MIROC) and three anthropogenic climate forcing scenarios (RCP 2.6, RCP 4.5 and RCP 8.5). Results of the ensemble models were then used to quantify and map the uncertainty created by model specifications. The consensus model projects strong increase (&gt;4-fold by 2080s) in burning rates, particularly under high climate-forcing scenarios. Even with very high goodness-of-fit in the consensus model, the model-specification uncertainty for future periods (&gt;200%) could still be much higher than that of different GCMs and RCP scenarios. When tallied, we show that the total uncertainty could greatly hinder our ability to detect significant trends in burning rates for much of Canada at the end of the 21st century.

P018: A prospective diagnostic support tool for the differentiation of abdominal pain in the adult emergency department population

Introduction: The chaotic environment of the emergency department has a deleterious effect on clinical judgement. The diagnosis of abdominal pathology is difficult to differentiate. There are also many diagnoses that could be considered abdominal in nature, exacerbating the task of diagnosing these patients. We propose a novel machine-learning method, Hierarchical Structured Models (HSMs), to provide an adjunct to clinician judgement, that provides a ranking of the probabilities of a patient having each of 39 abdominal pathologies, using only variables at the triage stage of emergency department care, and compare its performance to several machine-learning methods. Methods: This was a retrospective analysis of 25,861 patients that presented with one of 39 ICD-9 abdominal pathologies. 90% of the data was used to build and fine-tune the model, and 10% was used for testing. Predictors included age, gender, triage vitals and presenting complaint. All variables were solely collected from the Emergency Department Information System (EDIS). A decision tree structure was built using hierarchical clustering algorithms, and then a support vector machine (SVM) was fit at each node. To optimize the parameters for each node, a grid-search method was used to maximize ten-fold classification accuracy. The output of the decision tree was the probability of a particular presentation having each of the 39 diagnoses. This output was translated to a ranking of the relative likelihood of each of the diagnoses as a suggestion system for the treating physician. The accuracy of the system on the test set was compared to conventional machine-learning methods: pair-wise SVMs, gradient boosted models (GBM), neural networks (NN) and k-nearest neighbours (KNN). Results: The HSM ranked the correct diagnosis first 51.0% of the time, and ranked the correct diagnosis within the top three ranks 67.6% of the time. The most accurate model was GBMs (52.3%), and the least was neural networks (50.4%). Conclusion: The HSM approach using only variables available electronically at triage successfully ranked the correct diagnosis 51.0% of the time, and within the top three 67.6% of the time. Future research will focus on the inclusion of clinically lab results and radiology reports that are available electronically to improve HSM accuracy, and supplement physician diagnosis.