Classification And Regression Trees Analyses Research Articles

Macular Telangiectasia Type 2: A Classification System Using MultiModal Imaging MacTel Project Report Number 10.

To develop a severity classification for macular telangiectasia type 2 (MacTel) disease using multimodal imaging. An algorithm was used on data from a prospective natural history study of MacTel for classification development. A total of 1733 participants enrolled in an international natural history study of MacTel. The Classification and Regression Trees (CART), a predictive nonparametric algorithm used in machine learning, analyzed the features of the multimodal imaging important for the development of a classification, including reading center gradings of the following digital images: stereoscopic color and red-free fundus photographs, fluorescein angiographic images, fundus autofluorescence images, and spectral-domain (SD)-OCT images. Regression models that used least square method created a decision tree using features of the ocular images into different categories of disease severity. The primary target of interest for the algorithm development by CART was the change in best-corrected visual acuity (BCVA) at baseline for the right and left eyes. These analyses using the algorithm were repeated for the BCVA obtained at the last study visit of the natural history study for the right and left eyes. The CART analyses demonstrated 3 important features from the multimodal imaging for the classification: OCT hyper-reflectivity, pigment, and ellipsoid zone loss. By combining these 3 features (as absent, present, noncentral involvement, and central involvement of the macula), a 7-step scale was created, ranging from excellent to poor visual acuity. At grade 0, 3 features are not present. At the most severe grade, pigment and exudative neovascularization are present. To further validate the classification, using the Generalized Estimating Equation regression models, analyses for the annual relative risk of progression over a period of 5 years for vision loss and for progression along the scale were performed. This analysis using the data from current imaging modalities in participants followed in the MacTel natural history study informed a classification for MacTel disease severity featuring variables from SD-OCT. This classification is designed to provide better communications to other clinicians, researchers, and patients. Proprietary or commercial disclosure may be found after the references.

Retrospective analysis of estrogenic endocrine disruption and land-use influences in the Chesapeake Bay watershed

The Chesapeake Bay is the largest estuary in the United States and its watershed includes river drainages in six states and the District of Columbia. Sportfishing is of major economic interest, however, the rivers within the watershed provide numerous other ecological, recreational, cultural and economic benefits, as well as serving as a drinking water source for millions of people. Consequently, major fish kills and the subsequent finding of estrogenic endocrine disruption (intersex or testicular oocytes and plasma vitellogenin in male fishes) raised public and management concerns. Studies have occurred at various sites within the Bay watershed to document the extent and severity of endocrine disruption, identify risk factors and document temporal and spatial variability. Data from these focal studies, which began in 2004, were used in CART (classification and regression trees) analyses to better identify land use associations and potential management practices that influence estrogenic endocrine disruption. These analyses emphasized the importance of scale (immediate versus upstream catchment) and the complex mixtures of stressors which can contribute to surface water estrogenicity and the associated adverse effects of exposure. Both agricultural (percent cultivated, pesticide application, phytoestrogen cover crops) and developed (population density, road density, impervious surface) land cover showed positive relationships to estrogenic indicators, while percent forest and shrubs generally had a negative association. The findings can serve as a baseline for assessing ongoing restoration and management practices.

Using Classification and Regression Trees (CART) and random forests to analyze attrition: Results from two simulations.

In this article, we describe a recent development in the analysis of attrition: using classification and regression trees (CART) and random forest methods to generate inverse sampling weights. These flexible machine learning techniques have the potential to capture complex nonlinear, interactive selection models, yet to our knowledge, their performance in the missing data analysis context has never been evaluated. To assess the potential benefits of these methods, we compare their performance with commonly employed multiple imputation and complete case techniques in 2 simulations. These initial results suggest that weights computed from pruned CART analyses performed well in terms of both bias and efficiency when compared with other methods. We discuss the implications of these findings for applied researchers.

Using decision trees to understand structure in missing data

ObjectivesDemonstrate the application of decision trees—classification and regression trees (CARTs), and their cousins, boosted regression trees (BRTs)—to understand structure in missing data.SettingData taken from employees at 3 different industrial sites...

Classification models for neurocognitive impairment in HIV infection based on demographic and clinical variables.

ObjectiveWe used demographic and clinical data to design practical classification models for prediction of neurocognitive impairment (NCI) in people with HIV infection.MethodsThe study population comprised 331 HIV-infected patients with available demographic, clinical, and neurocognitive data collected using a comprehensive battery of neuropsychological tests. Classification and regression trees (CART) were developed to obtain detailed and reliable models to predict NCI. Following a practical clinical approach, NCI was considered the main variable for study outcomes, and analyses were performed separately in treatment-naïve and treatment-experienced patients.ResultsThe study sample comprised 52 treatment-naïve and 279 experienced patients. In the first group, the variables identified as better predictors of NCI were CD4 cell count and age (correct classification [CC]: 79.6%, 3 final nodes). In treatment-experienced patients, the variables most closely related to NCI were years of education, nadir CD4 cell count, central nervous system penetration-effectiveness score, age, employment status, and confounding comorbidities (CC: 82.1%, 7 final nodes). In patients with an undetectable viral load and no comorbidities, we obtained a fairly accurate model in which the main variables were nadir CD4 cell count, current CD4 cell count, time on current treatment, and past highest viral load (CC: 88%, 6 final nodes).ConclusionPractical classification models to predict NCI in HIV infection can be obtained using demographic and clinical variables. An approach based on CART analyses may facilitate screening for HIV-associated neurocognitive disorders and complement clinical information about risk and protective factors for NCI in HIV-infected patients.

Meniscus body position in the normal knee and its change over 4 years

used logistic regression models to evaluate if the presence of JSN or worsening of JSN (JSN progression), as dichotomous outcomes, were associated with BML volumewhile controlling for sex, body mass index, and age. We also used classification and regression trees (CART) to assess whether baseline BML volume or BML volume change was more influential in predicting JSN progression. To confirm the primary results we used data from the clinical trial (n 1⁄4 103) and performed logistic regression models to examine the association between manually measured BML size change (classified as BML regression, no change, or BML progression) and the outcome of cartilage thickness derived from manual cartilage segmentation (stratified into tertiles). Results: Among 362 knees with BML and JSN data, 240 (66%) had JSN at baseline including 25 knees with severe JSN (grade 3) that were excluded from analyses of JSN progression. Larger baseline BML volume was associated with the presence of baseline JSN (odds ratio [OR] 1⁄4 1.5; 95% confidence interval [CI]1⁄4 1.2 1.8) and JSN progression in the same TF compartment (OR 1⁄4 1.3; 95% CI 1⁄4 1.1 1.5). There was a trend that BML regression and BML progression may be associated with greater odds of JSN progression relative to knees with no or minimal changes in BML volume (Table). Sensitivity analyses, excluding knees with more severe lateral JSN, revealed that BML regression in the medial TF compartment was associated with greater odds of medial JSN progression (Table). In CART analyses baseline BML volume was the variable best able to discriminate between those who did and did not have JSN progression. Finally, confirmatory analyses demonstrated that knees with BML regression (OR 1⁄4 2.7; 95% CI 1⁄4 1.0 7.1) or BML progression (OR 1⁄4 2.0; 95% CI 1⁄4 0.9 4.6) had greater odds of cartilage thickness loss. Conclusions: Baseline BML size is an important assessment in knee OA but a reduction of BML size, as measured on traditional MRI, may not be associated with reduced odds of structural changes.

Using Classification Trees for Predicting National Bridge Inventory Condition Ratings

In bridge management practice, bridge condition is the fundamental information needed to allow decision makers to make well-informed decisions regarding preservation, rehabilitation, or replacement of a bridge or network of bridges. The National Bridge Inventory (NBI) condition ratings, collected since the early 1970s, and the Commonly Recognized (CoRe) element condition data, collected since the early 1990s, are two major sources of bridge condition data in the United States. General NBI condition ratings are utilized for performance assessment, performance reporting, resource allocation, and selection of bridge projects by all levels of government. Since the early 1990s, the bridge management community has been interested in an algorithm to predict the categorical NBI condition rating classes from the more quantitative and detailed CoRe element condition data. An algorithm with sufficient predictive accuracy would make only CoRe element inspections necessary and would provide time and resource savings. This paper presents a new methodology for this purpose, using classification and regression trees (CARTs). The CART analyses were conducted with the bridge condition data provided by three state transportation agencies, using data from 2006 to 2010. The statistical results point to a more accurate prediction method than the previous algorithms described in the literature.

Genetics of emergent suicidality during antidepressive treatment—Data from a naturalistic study on a large sample of inpatients with a major depressive episode

Factors contributing to treatment-emergent suicidal ideation (TESI) using antidepressants have been in the focus of recent research strategies. We investigated previously established clinical predictors of TESI and combined these with several polymorphisms of candidate genes in patients with major depressive disorder. Common polymorphisms involved in the tryptophan hydroxylase 1 (TPH1) and 2 (TPH2), serotonin transporter, monoamine oxidase A (MAOA) and brain-derived neurotrophic factor (BDNF) were investigated in a naturalistic inpatient study of the German research network on depression. We compared patients showing TESI with non-TESI suicidal patients and with non-suicidal patients using univariate tests to detect relevant factors, which were further tested in logistic regression and CART (Classification and Regression Trees) analyses. Of the 269 patients, TESI occurred in 22 patients (17 female), 117 patients were defined as non-TESI suicidal patients, and 130 patients were classified as non-suicidal. When comparing cases with both control groups we found the TPH2 rs1386494 (C/T) polymorphism to be moderately associated with TESI (Univariate tests: TESI vs. non-suicidality: p=0.005; adjusted: p=0.09; TESI vs. non-TESI suicidal patients: p=0.0024; adjusted: p=0.086). This polymorphism remained the only significant genetic factor in addition to clinical predictors in logistic regression and CART analyses. CART analyses suggested interactions with several clinical predictors. Haplotype analyses further supported a contribution of this polymorphism in TESI. The TPH2 rs1386494 (C/T) polymorphism might contribute to the genetic background of TESI. This polymorphism has been previously associated with committed suicide and major depressive disorder. The small number of cases warrants replication in larger patient samples. Lack of a placebo control group hampers definite conclusions on an association with antidepressive treatment.

Discriminative Ability of Heel Quantitative Ultrasound in Postmenopausal Women with Prevalent Vertebral Fractures: Application of Optimal Threshold Cutoff Values Using Classification and Regression Tree Models

Quantitative ultrasound (QUS) of the heel has been proposed as a screening tool to evaluate the bone status and risk of osteoporotic fragility fractures. The aim of this study was to define threshold values that would maximize the predictive ability of QUS to discriminate subjects with vertebral fractures using the classification and regression trees (CART) models. A cross-sectional analysis was made of a cohort of 1,132 postmenopausal women with a mean age of 58years. A total of 205 women (18.1%) presented with a history of vertebral fracture. For all patients, a questionnaire of osteoporosis risk factors was given and measurements of the heel QUS and bone mineral density at the lumbar spine and the proximal femur, obtained by dual-energy X-ray absorptiometry (DXA), were made. Spinal radiographs were assessed for vertebral fractures. Sensitivity, specificity, predictive values, likelihood ratios, and receiver operator characteristics (ROC) curve QUS values were calculated using the optimal threshold identified in the CART models. Cutoff values calculated from best CART model (i.e., a QUS index >90.5%) yielded a sensitivity of 80.3% (95% CI 69.2-88.1), a negative predictive value of 94% (95% CI 90.1-96.5), and a specificity of 68.8% (95% CI 63.3-73.8). This cutoff value would obviate the need to perform DXA in 32.8% of the women of our population at risk for vertebral fractures. The area under the ROC curve of the best model was 0.8071. QUS was shown to discriminate between women with and without a history of vertebral fracture and constitutes a useful tool for assessing vertebral fracture risk. The application of decision trees (CART analyses) was helpful to define the optimal threshold QUS values.

Discriminative Ability of Heel Quantitative Ultrasound in Postmenopausal Women With Prevalent Low-Trauma Fractures: Application of Optimal Threshold Cutoff Values Using CART Models

Quantitative ultrasound (QUS) of the heel has been proposed as a screening tool to evaluate the bone status and risk of osteoporotic fragility fractures. The aim of this study was to define threshold values of QUS that would maximize the predictive ability of this technique to discriminate subjects with fragility fractures. A cross-sectional analysis was made of a cohort of 1132 postmenopausal women with a mean age of 58 yr. A total of 361 women (31.9%) presented with a history of osteoporotic fracture. Most fractures (74.1%) were nonvertebral. For all patients, a questionnaire of osteoporosis risk factors and measurements of the heel QUS and bone mineral density at the lumbar spine and the proximal femur obtained by dual-energy X-ray absorptiometry (DXA) were assessed. Spinal radiographs were assessed for fractures and historical nonvertebral fragility fractures. Sensitivity, specificity, predictive values, likelihood ratios, and receiver operator characteristic (ROC) curve QUS values were calculated using the optimal threshold identified in the classification and regression trees (CART) models. Cutoff values calculated from the best CART model (i.e., a quantitative ultrasound index (QUI) greater than 88.5% in women aged 58 yr or older) yielded 88.8% (95% confidence interval [CI]: 81.4–93.5) for sensitivity, a negative predictive value of 93.8 (95% CI: 89.4–96.4), and 70.4% (95% CI: 64.6–75.7) for specificity. This cutoff value would obviate the need to perform DXA in 43.1% of the population. The area under the ROC curve of the best model was 0.8363 (95% CI: 0.8249–0.8477). In conclusion, QUS was shown to discriminate between women with and without a history of fragility fracture and constitutes a useful tool for assessing fracture risk. The application of decision trees (CART analyses) was helpful to define the optimal threshold QUS values.

大腿骨頸部骨折患者の歩行自立に必要な要因―決定木分析による検討―

The purpose of this study was to develop a decision-making process to assess the conditions for predicting gait independence in patients with femoral neck fracture. A total of 108 patients were divided into 2 groups on the basis of their walking abilities at discharge for an unrelated illness; an independent (n=55) and dependent group (n=53). Details regarding age, sex, length of hospital therapy, operative procedures, classification of fracture, past history of stroke and fracture were collected from medical records. Body mass index (BMI), knee extension power, maximum walking speed, functional reach test (FRT) and the mini-mental state examination (MMSE) were measured to evaluate motor ability and cognitive status at discharge. Student's t-test and the chi-squared test were used to test for statistical differences between the 2 groups. On multivariate analysis, classification and regression trees (CART) was used to determine the predictive value of those measures that differed significantly between the 2 groups. On bivariable analysis, significant differences were found in nearly all variables, except for BMI and length of hospital therapy. As a result of this analysis, the decision tree, which consists of knee extension power, FRT, MMSE and a past history of stroke, was created. CART analyses showed that when knee extension power was >0.34 kgf/kg, the MMSE score was >13.5; with no past history of stroke, the rate of independent walking at discharge was 93.8%. In contrast, when knee extension power was ≤0.33 kgf/kg, FRT was ≤25.5 cm, the MMSE score was ≤13.5, and the rate of dependent walking at discharge was 100%. Our findings indicate that the decision tree can be helpful in predicting gait independence in patients with femoral neck fracture.

Risk Factors for SARS Transmission from Patients Requiring Intubation: A Multicentre Investigation in Toronto, Canada

BackgroundIn the 2003 Toronto SARS outbreak, SARS-CoV was transmitted in hospitals despite adherence to infection control procedures. Considerable controversy resulted regarding which procedures and behaviours were associated with the greatest risk of SARS-CoV transmission.MethodsA retrospective cohort study was conducted to identify risk factors for transmission of SARS-CoV during intubation from laboratory confirmed SARS patients to HCWs involved in their care. All SARS patients requiring intubation during the Toronto outbreak were identified. All HCWs who provided care to intubated SARS patients during treatment or transportation and who entered a patient room or had direct patient contact from 24 hours before to 4 hours after intubation were eligible for this study. Data was collected on patients by chart review and on HCWs by interviewer-administered questionnaire. Generalized estimating equation (GEE) logistic regression models and classification and regression trees (CART) were used to identify risk factors for SARS transmission.Results45 laboratory-confirmed intubated SARS patients were identified. Of the 697 HCWs involved in their care, 624 (90%) participated in the study. SARS-CoV was transmitted to 26 HCWs from 7 patients; 21 HCWs were infected by 3 patients. In multivariate GEE logistic regression models, presence in the room during fiberoptic intubation (OR = 2.79, p = .004) or ECG (OR = 3.52, p = .002), unprotected eye contact with secretions (OR = 7.34, p = .001), patient APACHE II score ≥20 (OR = 17.05, p = .009) and patient Pa02/Fi02 ratio ≤59 (OR = 8.65, p = .001) were associated with increased risk of transmission of SARS-CoV. In CART analyses, the four covariates which explained the greatest amount of variation in SARS-CoV transmission were covariates representing individual patients.ConclusionClose contact with the airway of severely ill patients and failure of infection control practices to prevent exposure to respiratory secretions were associated with transmission of SARS-CoV. Rates of transmission of SARS-CoV varied widely among patients.

Medical event profiling of COPD patients.

This analysis proposes appropriate descriptive and analytical statistical techniques to profile medical events requiring GP consultation using the example of chronic obstructive pulmonary disease (COPD). Consultation patterns were examined for 1807 cases with a diagnosis compatible with COPD in 1998 and a group of controls matched by sex, age and practice from the General Practice Research Database, a nationally representative UK primary care database. Consulting patterns by Read code chapter and chapter subdivision were examined using cluster analysis, logistic regression and classification and regression trees (CART). CART and multivariate logistic regression analyses suggested that COPD patients were more likely to consult with pulmonary circulatory disease (multivariate OR: 7.46 (95% confidence intervals: 2.05, 20.01) ), non-COPD respiratory diseases (2.77 (2.28, 3.37)), mycoses (2.0 (1.43, 2.71)), 'symptoms or ill-defined conditions' (1.95 (1.68, 2.29)), or other forms of heart disease (1.84 (1.92, 2.64)), and less likely to consult with hypertensive diseases (0.73 (0.57, 0.96)). Regression also showed positive associations with digestive system diseases (OR: 1.31 (1.02, 1.68)) and negative associations with 'other viral or chlamydial disease' (0.16 (0.03, 0.88)). A borderline significance reduced risk for cancers was seen in univariate logistic regression analyses. Cluster analyses were not useful in discriminating between cases and controls. These analyses provide information about the natural history of COPD and could be used to help interpret or detect adverse drug reactions if repeated before and after introduction of a treatment. COPD can be considered a multicomponent disease with more frequent comorbidities than age- and gender-matched individuals without COPD.