Weight-elimination neural networks applied to coronary surgery mortality prediction

Boundedness and convergence analysis of weight elimination for cyclic training of neural networks

Weight elimination offers a simple and efficient improvement of training algorithm of feedforward neural networks. It is a general regularization technique in terms of the flexible scaling parameters. Actually, the weight elimination technique also contains the weight decay regularization for a large scaling parameter. Many applications of this technique and its improvements have been reported. However, there is little research concentrated on its convergence behavior. In this paper, we theoretically analyze the weight elimination for cyclic learning method and determine the conditions for the uniform boundedness of weight sequence, and weak and strong convergence. Based on the assumed network parameters, the optimal choice for the scaling parameter can also be determined. Moreover, two illustrative simulations have been done to support the theoretical explorations as well.

Two new mathematical models for prediction of early mortality risk in coronary artery bypass graft surgery

OBJECTIVES

The aim of this study was to develop new models for prediction of short-term mortality risk in on-pump coronary artery bypass grafting (CABG) surgery using decision tree (DT) methods.

METHODS

Between September 2005 and April 2006, 948 consecutive patients underwent CABG surgery at Rajaie Heart Center. Potential risk factors were reviewed and univariate and multivariate analysis for short-term mortality were performed. The whole dataset was divided into mutually exclusive subsets. An entropy error fuzzy decision tree (EEFDT) and an entropy error crisp decision tree (EECDT) were implemented using 650 (68.6%) patient data and tested with 298 (31.4%) patient data. Ten times hold-out cross validation was done and the area under the receiver operative characteristic curve (AUC) was reported as model performance. The results were compared with the logistic regression (LR) model and EuroSCORE.

RESULTS

The overall short-term mortality rate was 3.8%, and was statistically higher in women than men (P<.001). The final EEFDT selected 19 variables and resulted in a tree with 39 nodes, 20 conditional rules, and AUC of 0.90±0.008. The final EECDT selected 15 variables and resulted in a tree with 35 nodes, 18 conditional rules, and AUC of 0.86±0.008. The LR model selected 10 variables and resulted in an AUC of 0.78±0.008; the AUC for EuroSCORE was 0.77±0.003. There were no differences in the discriminatory power of EEFDT and EECDT (P=.066) and their performance was superior to LR and EuroSCORE.

CONCLUSIONS

EEFDT, EECDT, LR, and EuroSCORE had clinical acceptance but the performance and accuracy of the DTs were superior to the other models.

Integration of new technology in a legacy system for collecting medical data - challenges and lessons learned

Integrating new technology into a legacy medical system can be very challenging. Completely new systems cannot always be built due to the high cost of medical equipment, thus integrating some new technology into an existing system may be required. This paper looks at the issues and challenges surrounding the integration of new components into a legacy system for collecting medical data. We discuss how the issues were solved, the lessons learned, and how future upgrades can be made more easily.

Predicting adverse outcomes of cardiac surgery with the application of artificial neural networks

Risk-stratification models based on pre-operative patient and disease characteristics are useful for providing individual patients with an insight into the potential risk of complications and mortality, for aiding the clinical decision for surgery vs non-surgical therapy, and for comparing the quality of care between different surgeons or hospitals. Our study aimed to apply artificial neural networks (ANN) models to predict mortality and morbidity after cardiac surgery, and also to compare the efficacy of this model to that of the logistic regression model and Parsonnet score. The accuracy of the ANN, logistic regression and Parsonnet score in predicting mortality was 83.8%, 87.9% and 78.4%. The accuracy of the ANN, logistic regression and Parsonnet score in predicting major morbidity was 79.0%, 74.3% and 68.6%. The area under the receiver operating characteristic curves (AUC) of the ANN, logistic regression and Parsonnet score in predicting in-hospital mortality were 0.873, 0.852 and 0.829. The AUCs of the ANN, logistic regression and Parsonnet score in predicting major morbidity were 0.852, 0.789 and 0.727. The results showed the ANN models have the best discriminating power in predicting in-hospital mortality and morbidity among these models.

Predicting high-risk preterm birth using artificial neural networks

A reengineered approach to the early prediction of preterm birth is presented as a complimentary technique to the current procedure of using costly and invasive clinical testing on high-risk maternal populations. Artificial neural networks (ANNs) are employed as a screening tool for preterm birth on a heterogeneous maternal population; risk estimations use obstetrical variables available to physicians before 23 weeks gestation. The objective was to assess if ANNs have a potential use in obstetrical outcome estimations in low-risk maternal populations. The back-propagation feedforward ANN was trained and tested on cases with eight input variables describing the patient's obstetrical history; the output variables were: 1) preterm birth; 2) high-risk preterm birth; and 3) a refined high-risk preterm birth outcome excluding all cases where resuscitation was delivered in the form of free flow oxygen. Artificial training sets were created to increase the distribution of the underrepresented class to 20%. Training on the refined high-risk preterm birth model increased the network's sensitivity to 54.8%, compared to just over 20% for the nonartificially distributed preterm birth model.

Risk factor identification and mortality prediction in cardiac surgery using artificial neural networks

OBJECTIVE

The artificial neural network model is a nonlinear technology useful for complex pattern recognition problems. This study aimed to develop a method to select risk variables and predict mortality after cardiac surgery by using artificial neural networks.

METHODS

Prospectively collected data from 18,362 patients undergoing cardiac surgery at 128 European institutions in 1995 (the European System for Cardiac Operative Risk Evaluation database) were used. Models to predict the operative mortality were constructed using artificial neural networks. For calibration a sixfold cross-validation technique was used, and for testing a fourfold cross-testing was performed. Risk variables were ranked and minimized in number by calibrated artificial neural networks. Mortality prediction with 95% confidence limits for each patient was obtained by the bootstrap technique. The area under the receiver operating characteristics curve was used as a quantitative measure of the ability to distinguish between survivors and nonsurvivors. Subgroup analysis of surgical operation categories was performed. The results were compared with those from logistic European System for Cardiac Operative Risk Evaluation analysis.

RESULTS

The operative mortality was 4.9%. Artificial neural networks selected 34 of the total 72 risk variables as relevant for mortality prediction. The receiver operating characteristics area for artificial neural networks (0.81) was larger than the logistic European System for Cardiac Operative Risk Evaluation model (0.79; P = .0001). For different surgical operation categories, there were no differences in the discriminatory power for the artificial neural networks (P = .15) but significant differences were found for the logistic European System for Cardiac Operative Risk Evaluation (P = .0072).

CONCLUSIONS

Risk factors in a ranked order contributing to the mortality prediction were identified. A minimal set of risk variables achieving a superior mortality prediction was defined. The artificial neural network model is applicable independent of the cardiac surgical procedure.

Identifying risk factors for two complication types for neonatal intensive care patients (NICU)

This paper discusses the results of applying artificial neural networks to predicting complication for neonatal intensive care patients. Risk factors that lead to necrotizing entero-colitis or broncho-pulmonary dysplasia were identified. Future work will expand this work to other outcomes and add probability information to the estimations.

Mortality assessment in intensive care units via adverse events using artificial neural networks

OBJECTIVE

This work presents a novel approach for the prediction of mortality in intensive care units (ICUs) based on the use of adverse events, which are defined from four bedside alarms, and artificial neural networks (ANNs). This approach is compared with two logistic regression (LR) models: the prognostic model used in most of the European ICUs, based on the simplified acute physiology score (SAPS II), and a LR that uses the same input variables of the ANN model.

MATERIALS AND METHODS

A large dataset was considered, encompassing forty two ICUs of nine European countries. The recorded features of each patient include the final outcome, the case mix (e.g. age) and the intermediate outcomes, defined as the daily averages of the out of range values of four biometrics (e.g. heart rate). The SAPS II score requires 17 static variables (e.g. serum sodium), which are collected within the first day of the patient's admission. A nonlinear least squares method was used to calibrate the LR models while the ANNs are made up of multilayer perceptrons trained by the RPROP algorithm. A total of 13,164 adult patients were randomly divided into training (66%) and test (33%) sets. The two methods were evaluated in terms of receiver operator characteristic (ROC) curves.

RESULTS

The event based models predicted the outcome more accurately than the currently used SAPS II model (P<0.05), with ROC areas within the ranges 83.9-87.1% (ANN) and 82.6-85.2% (LR) versus 80% (LR SAPS II). When using the same inputs, the ANNs outperform the LR (improvement of 1.3-2%).

CONCLUSION

Better prognostic models can be achieved by adopting low cost and real-time intermediate outcomes rather than static data.

A novel dengue fever (DF) and dengue haemorrhagic fever (DHF) analysis using artificial neural network (ANN)

Dengue fever (DF) is an acute febrile viral disease frequently presented with headache, bone or joint and muscular pains, and rash. A significant percentage of DF patients develop a more severe form of disease, known as dengue haemorrhagic fever (DHF). DHF is the complication of DF. The main pathophysiology of DHF is the development of plasma leakage from the capillary, resulting in haemoconcentration, ascites, and pleural effusion that may lead to shock following defervescence of fever. Therefore, accurate prediction of the day of defervescence of fever is critical for clinician to decide on patient management strategy. To date, no known literature describes of any attempt to predict the day of defervescence of fever in DF patients. This paper describes a non-invasive prediction system for predicting the day of defervescence of fever in dengue patients using artificial neural network. The developed system bases its prediction solely on the clinical symptoms and signs and uses the multilayer feed-forward neural networks (MFNN). The results show that the proposed system is able to predict the day of defervescence in dengue patients with 90% prediction accuracy.

Improvement and automation of artificial neural networks to estimate medical outcomes

The lengthy process of manually optimizing a feedforward backpropagation artificial neural network (ANN) provided the incentive to develop an automated system that could fine-tune the network parameters without user supervision. A new stopping criterion was introduced--the logarithmic-sensitivity index--that manages a good balance between sensitivity and specificity of the output classification. The automated network automatically monitored the classification performance to determine when was the best time to stop training-after no improvement in the performance measure (either highest correct classification rate, lowest mean squared error or highest log-sensitivity index value) occurred in the subsequent 500 epochs. Experiments were performed on three medical databases: an adult intensive care unit, a neonatal intensive care unit and a coronary surgery patient database. The optimal network parameter settings found by the automated system were similar to those found manually. The results showed that the automated networks performed equally well or better than the manually optimized ANNs, and the best classification performance was achieved using the log-sensitivity index as a stopping criterion.