应用人工神经网络预测个体患原发性高血压病危险度

目的 建立个体患原发性高血压病的预测模型,评价并探讨预测个体患病的新方法.方法 选择3054名社区居民流行病学调查资料,按照年龄、性别均衡性,按4:1分为训练集(2438名)与检验集(616名)两部分,分别用于筛选变量、建立预测模型及对模型的检测和评价.应用人工神经网络(ANN)和logistic回归分析方法建立模型,用ROC方法评价所建立的高血压患病预测模型的优劣.结果 对616名检验集预测,ANN模型的特异性较低,但准确性、灵敏度指标均优于logistic回归模型,ANN2的约登指数为0.8399,明显高于其他两个模型;通过ROC曲线下面积比较模型的预测能力:logistic回归方法曲线下面积(Az=0.732±0.026)小于ANN模型(ANN2和ANN1分别为0.918±0.013、0.900±0.014),即ANN模型有更好的预测判别效能.结论 初步证明在预测个体患高血压病方面,ANN方法预测效能更优,从而为解决个体发病危险预测提供了一个新方法.     

Bayesian neural networks for bivariate binary data: an application to prostate cancer study

Prostate cancer is one of the most common cancers in American men. The cancer could either be locally confined, or it could spread outside the organ. When locally confined, there are several options for treating and curing this disease. Otherwise, surgery is the only option, and in extreme cases of outside spread, it could very easily recur within a short time even after surgery and subsequent radiation therapy. Hence, it is important to know, based on pre-surgery biopsy results how likely the cancer is organ-confined or not.The paper considers a hierarchical Bayesian neural network approach for posterior prediction probabilities of certain features indicative of non-organ confined prostate cancer. In particular, we find such probabilities for margin positivity (MP) and seminal vesicle (SV) positivity jointly. The available training set consists of bivariate binary outcomes indicating the presence or absence of the two. In addition, we have certain covariates such as prostate specific antigen (PSA), gleason score and the indicator for the cancer to be unilateral or bilateral (i.e. spread on one or both sides) in one data set and gene expression microarrays in another data set. We take a hierarchical Bayesian neural network approach to find the posterior prediction probabilities for a test and validation set, and compare these with the actual outcomes for the first data set. In case of the microarray data we use leave one out cross-validation to access the accuracy of our method. We also demonstrate the superiority of our method to the other competing methods through a simulation study. The Bayesian procedure is implemented by an application of the Markov chain Monte Carlo numerical integration technique. For the problem at hand, our Bayesian bivariate neural network procedure is shown to be superior to the classical neural network, Radford Neal's Bayesian neural network as well as bivariate logistic models to predict jointly the MP and SV in a patient in both the data sets as well as in the simulation study.     

C5.0决策树与RBF神经网络模型用于急性缺血性脑卒中出血性转化的风险预测性能比较

目的  比较C5.0决策树与径向基函数（radial basis function,RBF）神经网络用于急性缺血性脑卒中（acute jschemic stroke,AIS）出血性转化（hemorrhagic transformation,HT）风险预测性能。 方法  将AIS住院患者作为研究对象,收集相关资料。根据入院2周内是否发生HT分为HT组与非HT组,建立C5.0决策树与RBF神经网络模型,比较两者的预测性能。 结果  共收集460份病历资料,按照训练集与测试集7 ∶3的比例分为训练集样本和测试集样本。C5.0决策树模型的训练集与测试集准确率分别为96.5%和80.1%,灵敏度为98.1%和82.6%,特异度为94.8%和77.9%,Kappa指数是0.93和0.60,AUC是0.97和0.80。RBF神经网络模型的训练集与测试集准确率分别为72.6%和74.7%,灵敏度为87.6%和88.4%,特异度为56.9%和62.3%,Kappa指数为0.45和0.50,AUC为0.72和0.75;在训练集中,C5.0决策树模型的预测性能优于RBF神经网络模型的预测性能。在测试集中,两者预测性能的差异无统计学意义。 结论  C5.0决策树模型的预测性能优于RBF神经网络模型的预测性能。     

Comparison of logistic regression and neural networks to predict rehospitalization in patients with stroke.

CONTEXT: Rehospitalization following inpatient medical rehabilitation has important health and economic implications for patients who have experienced a stroke. OBJECTIVE: Compare logistic regression and neural networks in predicting rehospitalization at 3-6-month follow-up for patients with stroke discharged from medical rehabilitation. DESIGN: The study was retrospective using information from a national database representative of medical rehabilitation patients across the US. SETTING: Information submitted to the Uniform Data System for Medical Rehabilitation from 1997 and 1998 by 167 hospital and rehabilitation facilities from 40 states was examined. PARTICIPANTS: 9584 patient records were included in the sample. The mean age was 70.74 years (SD = 12.87). The sample included 51.6% females and was 77.6% non-Hispanic White with an average length of stay of 21.47 days (SD = 15.47). MAIN OUTCOME MEASURES: Hospital readmission from 80 to 180 days following discharge. Results: Statistically significant variables (P <.05) in the logistic model included sphincter control, self-care ability, age, marital status, ethnicity and length of stay. Area under the ROC curves were 0.68 and 0.74 for logistic regression and neural network analysis, respectively. The Hosmer-Lemeshow goodness-of-fit chi-square was 11.32 (df = 8, P = 0.22) for neural network analysis and 16.33 (df = 8, P = 0.11) for logistic regression. Calibration curves indicated a slightly better fit for the neural network model. CONCLUSION: There was no statistically significant or practical advantage in predicting hospital readmission using neural network analysis in comparison to logistic regression for persons who experienced a stroke and received medical rehabilitation during the period of the study.     

Modelling time-dependent hazard ratios in relative survival: application to colon cancer

The Cox model is widely used in the evaluation of prognostic factors in clinical research. In population-based studies, however, which assess long-term survival of unselected populations, relative survival models are often considered more appropriate. In both approaches, the validity of proportional hazard hypothesis should be evaluated. To explore the validity of the proportional hazard assumption in a population-based study of colon cancer, to propose non-proportional hazard relative survival models and to evaluate their utility. The use of a piecewise proportional hazard relative survival model in colon cancer has shown that the effects of most clinical prognostic factors such as age, period of diagnosis and stage are non-proportional. The non-proportional hazard relative survival models developed in this article have been found to be efficient tools for better understanding the time-dependent aspect of prognostic factors.     

Cancer-specific survival of patients with multiple cancers: an application to patients with multiple breast cancers

In the analysis of cause-specific survival, the causes of death must be known. For single-cancer patients with a known cause of death, the estimation of the cause-specific survival rate is straightforward. For multiple-cancer patients with two primary cancers, however, the analysis of cause-specific survival rates is more complex, particularly if the cancers are of the same primary site. In these situations, a concept of cancer-specific survival may also be distinguished from cause-specific survival. Cancer-specific survival rates are studied here by introducing two models, the primary one where the death from cancer is attributed to one of the cancers, and an alternative where such an attribution is not necessary. The models are illustrated using data on patients with multiple breast cancers. The model-based survival rates are compared with each other and with the corresponding relative survival rates based on analogous modelling of relative survival. The results show that for the subsequent breast cancer, the cancer-specific survival rates based on the alternative, where the distinction between the cancers as a cause of death was not necessary, tended to be higher than those based on that distinction. It is thus possible that the subsequent cancer was too often coded as a cause of death, particularly when being localized at diagnosis.     

Understanding neural networks using regression trees: an application to multiple myeloma survival data

Neural networks are becoming very popular tools for analysing data. It is however quite difficult to understand the neural network output in terms of the original covariates or input variables. In this paper we provide, using readily available software, an easy way of understanding the output of the neural network using regression trees. We focus on the problem in the context of censored survival data for patients with multiple myeloma, where identifying groups of patients with different prognosis is an important aspect of clinical studies. The use of regression trees to help understand neural networks can be easily applied to uncensored situations.     

Modelling smoking history using a comprehensive smoking index: application to lung cancer

The mathematical representation of smoking history is an important tool in analysis of epidemiological and clinical data. Hoffmann and colleagues recently proposed a single aggregate measure of smoking exposure that incorporates intensity, duration, and time since cessation. This comprehensive smoking index (CSI), which may be incorporated in any regression model, depends on a half-life (tau) and a lag (delta) parameters that have to be fixed a priori, or estimated by maximizing the fit. The CSI has not previously been used for analysis of cancer data. Following some preliminary results on smoking and lung cancer, the authors proposed a new version of the CSI for lung cancer. The aim of this study was to investigate the performance of the original and the new versions of the CSI in the analysis of three data sets from two case-control studies of lung cancer undertaken in Montreal, in 1979-1985 in males, and in 1996-2000 in both males and females. The estimates of tau and delta for both versions of the CSI were similar across data sets. The new version of the CSI fitted the three data sets systematically although moderately better than the original version, and at least as well as other representations of lifetime smoking history that used separate variables for time since cessation and cumulative amount of cigarettes smoked. The results suggest that the CSI may be an attractive and parsimonious alternative to conventional modelling of different aspects of smoking history for lung cancer.     

The effect of hormone therapy on cognitive function in patients with breast cancer

Preclinical and clinical studies suggest that oestrogens have an important role in brain functioning and cognitive ability. Given that hormone therapies for breast cancer reduce oestrogen levels or block oestrogen receptors, it is conceivable that these agents also influence cognitive function. Several small studies have been conducted to address this issue, but many of them are methodologically insufficient. The negative effects of oophorectomy and luteinising hormone-releasing hormone (LHRH) analogues on verbal memory and working memory have been demonstrated the most consistently, albeit only in small studies. Anastrozole and tamoxifen also appear to exert some negative effect on cognition, but well-designed studies are lacking. No data are available on the influence of the aromatase inhibitors exemestane and letrozole on cognitive function. Raloxifene, a drug that has no obvious advantages over tamoxifen and will likely not be developed further for breast cancer treatment, has no negative influence on cognitive functioning. It remains unclear whether the observed effects are transient or permanent, and to what extent age, menopausal status and duration of therapy influence the severity of cognitive effects.     

乳腺癌围手术期抗菌药物应用的临床研究

目的探讨乳腺癌围手术期不使用抗菌药物对患者术后手术部位感染(SSI)的影响。方法回顾性分析2011年1月-2012年12月甲状腺乳腺外科收治的乳腺癌手术患者411例,将其分为对照组202例、研究组209例,对照组围手术期预防性使用抗菌药物,研究组围手术期不使用抗菌药物,观察两组患者SSI发生率,并对SSI发生及相关因素进行分析,数据采用SPSS13.0统计软件进行处理。结果 411例乳腺癌手术患者术后发生SSI2例,发生率为0.48%,对照组患者SSI发生率0.49%,研究组为0.48%,两组比较差异无统计学意义,证实乳腺癌术后SSI与是否应用抗菌药物无关;两组SSI患者在糖尿病、新辅助化疗、高龄等相关因素方面比较,差异无统计学意义;进一步分析发现,2例SSI均与术后皮下血肿及皮瓣坏死有关。结论乳腺癌术后SSI发生率与年龄、肥胖、糖尿病、术后创面血肿及血清肿等高危因素有关,与预防性使用抗菌药物无关,乳腺癌围手术期不使用抗菌药物并不增加SSI发生率,同时可以减少医疗费用,可以避免使用抗菌药物带来二次感染、耐药等问题,值得临床推广应用。     

Regressive logistic and proportional hazards disease models for within-family analyses of measured genotypes,with application to a CYP17 polymorphism and breast cancer

Various statistical methods have been proposed to evaluate associations between measured genetic variants and disease, including some using family designs. For breast cancer and rare variants, we applied a modified segregation analysis method that uses the population cancer incidence and population-based case families in which a mutation is known to be segregating. Here we extend the method to a common polymorphism, and use a regressive logistic approach to model familial aggregation by conditioning each individual on their mother's breast cancer history. We considered three models: 1) class A regressive logistic model; 2) age-of-onset regressive logistic model; and 3) proportional hazards familial model. Maximum likelihood estimates were calculated using the software MENDEL. We applied these methods to data from the Australian Breast Cancer Family Study on the CYP17 5'UTR T-->C MspA1 polymorphism measured for 1,447 case probands, 787 controls, and 213 relatives of case probands found to have the CC genotype. Breast cancer data for first- and second-degree relatives of case probands were used. The three methods gave consistent estimates. The best-fitting model involved a recessive inheritance, with homozygotes being at an increased risk of 47% (95% CI, 28-68%). The cumulative risk of the disease up to age 70 years was estimated to be 10% or 22% for a CYP17 homozygote whose mother was unaffected or affected, respectively. This analytical approach is well-suited to the data that arise from population-based case-control-family studies, in which cases, controls and relatives are studied, and genotype is measured for some but not all subjects.     

Modelling the progression towards duodenal cancer among patients with familial polyposis on the basis of two different score profiles

The objective was to design a method that considers, on clinical arguments, the likely existence of patient subgroups with different evolution profiles. The method is applied in familial adenomatous polyposis to predict the proportion of patients that would develop duodenal cancer. A subject-specific linear mixed-effects model was elaborated to explicitly model heterogeneity in regression parameters. The estimates of the parameters were obtained by Bayesian inference using Gibbs sampling. The application concerned two potential polyposis subgroups stable-state and progressive. Each patients score was expressed in function of his putative subgroup, the reference subgroup mean score (intercept), the rate of change (slope), and time. The estimated proportion of stable-state patients was 35. In progressive-state patients, the estimated annual score increase was 0.38 (95% CI: 0.27–0.48). The regression model predicted that the proportion of patients with a score 9 is near 43% at age 60 (36–50) and 50 at 70 (43–57). The method indicates the evolution profile of each subject, which facilitates therapeutic decisions. The modelling may be extended to other more complex situations with several subgroups, with different change rates, or with various genetic or therapeutic profiles.     

Approaches in modelling long-term survival: an application to breast cancer

Several modelling techniques have been proposed for non-proportional hazards. In this work we consider different models which can be classified into three wide categories: models with time-varying effects of the covariates; frailty models and cure rate models. We present those different extensions of the proportional hazards model on an application of 2433 breast cancer patients with a long follow-up. We comment on the differences and similarities among the models and evaluate their performance using survival and hazard plots, Brier scores and pseudo-observations.     

Using artificial neural networks to model the urinary excretion of total and purine derivative nitrogen fractions in cows.

A dataset of 177 individual nitrogen balances from dry and lactating cows was split in two independent groups: training dataset (n = 130) and challenge dataset (n = 47). The training dataset was used to develop multiple linear regressions (MLR) and artificial neural networks (ANN) aimed at predicting the urinary excretion of total (NURI) and that of purine derivative nitrogen (PDN). Input variables for the prediction of NURI were crude protein (CP) intake, effective degradability of non-protein dry matter (DM), neutral detergent fiber (NDF) content of the diet, live weight and milk yield. Live weight, total carbohydrate intake, the ratio of non-protein DM degraded to CP degraded and milk yield corrected for DM intake were entered to predict PDN. The regression between predicted and observed values for the training dataset showed a better statistical accuracy of ANN than did MLR models, especially for PDN. The evaluation of the two models on the challenge dataset showed similar determination coefficients, either when predicting total nitrogen excretion (0.623 and 0.614 for ANN and MLR, respectively) or PDN (0.688 and 0.666, for ANN and MLR, respectively). Moreover, both approaches were affected by a tendency to under-predict both targets at high levels of NURI and PDN. However, with the ANN approach, it is possible to study the response of the model to modifications of individual inputs by the so-called response analysis. This unique feature could be used to study the effect of different physiological situations as well as providing hypotheses for additional research.     

Developing the function of 'Magnitude-of-Effect' (MoE) for artificial neural networks to demonstrate the causal effect of exposure variables on outcome variable

Statistical analysis and logistic regression (LR) in particular are among the most popular tools being used by safety professionals and practitioners to assess the association between exposures and possible occupational disorders or diseases and predict the outcome. Recently, artificial neural network (ANN) models are gradually finding their way into safety field. It has been shown that they are capable of predicting outcomes more accurately than LR, but they are incapable of demonstrating the direct correlation between exposure variables and a possible outcome variable. The objective of this study was to develop a mathematical function that can use the result of ANN models to produce a measure for evaluating the direct association between exposure and possible outcome variables. This function was referred to as the function of Magnitude-of-Effect (MoE). Safety experts and practitioners can use the MoE function to interpret how strongly an exposure variable can affect the outcome variable, similar to an odds ratio, which can be calculated by using estimated parameters in LR models. The significance of such achievement is that it can eliminate one of the ANN model's shortcoming and make them more applicable in the occupational safety and health engineering field.     

From policy to patients and back: surgical treatment decision making for patients with breast cancer

Persistent use of mastectomy for breast cancer has motivated concerns about overtreatment by surgeons and lack of patient involvement in decisions. However, recent studies suggest that patients perceive substantial involvement and that some patients prefer more invasive surgery, while other research suggests that surgical treatment choices might be poorly informed. Decision-making quality can be improved by increasing patients' knowledge about treatments' risks and benefits and by optimizing their involvement. The mastectomy story underscores the limitations of utilization measures as quality indicators. Strategies to improve patient outcomes should focus on tools to improve the quality of decision making and innovations in multispecialty practice.     

Risk of contralateral breast cancer: associations with factors related to initial breast cancer

A case-control study was conducted to assess the risk factors associated with the development of a contralateral primary breast cancer among women who had had a first primary breast cancer. Hospital records were reviewed for 292 women who had an incident contralateral breast cancer, diagnosed in one of eight Connecticut hospitals between July 1, 1975 and December 31, 1983, and for a comparison group of 264 surviving unilateral breast cancer patients previously diagnosed in the same hospitals. All subjects were identified through the records of the Connecticut Tumor Registry. A family history of breast cancer in any first- or second-degree relative was associated with an almost threefold increased risk of developing a contralateral cancer (adjusted odds ratio (OR) = 2.8, 95% confidence interval (CI) = 1.6-4.9). Further, this relation was modified by the time elapsed since the initial cancer diagnosis (ratio of OR = 1.9, 95% CI = 1.2-3.0 for a five-year differential in time since initial diagnosis). A delay of 10 years in first full-term pregnancy was associated with a 36% decrease in risk (adjusted OR = 0.6, 95% CI = 0.3-1.2); this estimate excluded the magnitude of increased risk usually observed in studies of initial breast cancer. A conceptual framework is presented for assessing the study findings in the context of previous studies that have examined the corresponding associations for initial primary breast cancers.     

A new approach to training back-propagation artificial neural networks: empirical evaluation on ten data sets from clinical studies

We present a new approach to training back-propagation artificial neural nets (BP-ANN) based on regularization and cross-validation and on initialization by a logistic regression (LR) model. The new approach is expected to produce a BP-ANN predictor at least as good as the LR-based one. We have applied the approach to ten data sets of biomedical interest and systematically compared BP-ANN and LR. In all data sets, taking deviance as criterion, the BP-ANN predictor outperforms the LR predictor used in the initialization, and in six cases the improvement is statistically significant. The other evaluation criteria used (C-index, MSE and error rate) yield variable results, but, on the whole, confirm that, in practical situations of clinical interest, proper training may significantly improve the predictive performance of a BP-ANN.