Prediction of delayed graft function after kidney transplantation: comparison between logistic regression and machine learning methods.

BACKGROUND: Predictive models for delayed graft function (DGF) after kidney transplantation are usually developed using logistic regression. We want to evaluate the value of machine learning methods in the prediction of DGF. METHODS: 497 kidney transplantations from deceased donors at the Ghent University Hospital between 2005 and 2011 are included. A feature elimination procedure is applied to determine the optimal number of features, resulting in 20 selected parameters (24 parameters after conversion to indicator parameters) out of 55 retrospectively collected parameters. Subsequently, 9 distinct types of predictive models are fitted using the reduced data set: logistic regression (LR), linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), support vector machines (SVMs; using linear, radial basis function and polynomial kernels), decision tree (DT), random forest (RF), and stochastic gradient boosting (SGB). Performance of the models is assessed by computing sensitivity, positive predictive values and area under the receiver operating characteristic curve (AUROC) after 10-fold stratified cross-validation. AUROCs of the models are pairwise compared using Wilcoxon signed-rank test. RESULTS: The observed incidence of DGF is 12.5 %. DT is not able to discriminate between recipients with and without DGF (AUROC of 52.5 %) and is inferior to the other methods. SGB, RF and polynomial SVM are mainly able to identify recipients without DGF (AUROC of 77.2, 73.9 and 79.8 %, respectively) and only outperform DT. LDA, QDA, radial SVM and LR also have the ability to identify recipients with DGF, resulting in higher discriminative capacity (AUROC of 82.2, 79.6, 83.3 and 81.7 %, respectively), which outperforms DT and RF. Linear SVM has the highest discriminative capacity (AUROC of 84.3 %), outperforming each method, except for radial SVM, polynomial SVM and LDA. However, it is the only method superior to LR. CONCLUSIONS: The discriminative capacities of LDA, linear SVM, radial SVM and LR are the only ones above 80 %. None of the pairwise AUROC comparisons between these models is statistically significant, except linear SVM outperforming LR. Additionally, the sensitivity of linear SVM to identify recipients with DGF is amongst the three highest of all models. Due to both reasons, the authors believe that linear SVM is most appropriate to predict DGF.

[Genetic algorithm based multi-objective least square support vector machine for simultaneous determination of multiple components by near infrared spectroscopy].

The near infrared (NIR) spectrum contains a global signature of composition, and enables to predict different proper ties of the material. In the present paper, a genetic algorithm and an adaptive modeling technique were applied to build a multiobjective least square support vector machine (MLS-SVM), which was intended to simultaneously determine the concentrations of multiple components by NIR spectroscopy. Both the benchmark corn dataset and self-made Forsythia suspense dataset were used to test the proposed approach. Results show that a genetic algorithm combined with adaptive modeling allows to efficiently search the LS-SVM hyperparameter space. For the corn data, the performance of multi-objective LS-SVM was significantly better than models built with PLS1 and PLS2 algorithms. As for the Forsythia suspense data, the performance of multi-objective LS-SVM was equivalent to PLS1 and PLS2 models. In both datasets, the over-fitting phenomena were observed on RBFNN models. The single objective LS-SVM and MLS-SVM didn't show much difference, but the one-time modeling convenience al lows the potential application of MLS-SVM to multicomponent NIR analysis.

Split BiRNN for real-time activity recognition using radar and deep learning.

Radar systems can be used to perform human activity recognition in a privacy preserving manner. This can be achieved by using Deep Neural Networks, which are able to effectively process the complex radar data. Often these networks are large and do not scale well when processing a large amount of radar streams at once, for example when monitoring multiple rooms in a hospital. This work presents a framework that splits the processing of data in two parts. First, a forward Recurrent Neural Network (RNN) calculation is performed on an on-premise device (usually close to the radar sensor) which already gives a prediction of what activity is performed, and can be used for time-sensitive use-cases. Next, a part of the calculation and the prediction is sent to a more capable off-premise machine (most likely in the cloud or a data center) where a backward RNN calculation is performed that improves the previous prediction sent by the on-premise device. This enables fast notifications to staff if troublesome activities occur (such as falling) by the on-premise device, while the off-premise device captures activities missed or misclassified by the on-premise device.

Random Survival Forests for Predicting the Bed Occupancy in the Intensive Care Unit.

Predicting the bed occupancy of an intensive care unit (ICU) is a daunting task. The uncertainty associated with the prognosis of critically ill patients and the random arrival of new patients can lead to capacity problems and the need for reactive measures. In this paper, we work towards a predictive model based on Random Survival Forests which can assist physicians in estimating the bed occupancy. As input data, we make use of the Sequential Organ Failure Assessment (SOFA) score collected and calculated from 4098 patients at two ICU units of Ghent University Hospital over a time period of four years. We compare the performance of our system with a baseline performance and a standard Random Forest regression approach. Our results indicate that Random Survival Forests can effectively be used to assist in the occupancy prediction problem. Furthermore, we show that a group based approach, such as Random Survival Forests, performs better compared to a setting in which the length of stay of a patient is individually assessed.

Predictive modelling of survival and length of stay in critically ill patients using sequential organ failure scores.

INTRODUCTION: The length of stay of critically ill patients in the intensive care unit (ICU) is an indication of patient ICU resource usage and varies considerably. Planning of postoperative ICU admissions is important as ICUs often have no nonoccupied beds available. PROBLEM STATEMENT: Estimation of the ICU bed availability for the next coming days is entirely based on clinical judgement by intensivists and therefore too inaccurate. For this reason, predictive models have much potential for improving planning for ICU patient admission. OBJECTIVE: Our goal is to develop and optimize models for patient survival and ICU length of stay (LOS) based on monitored ICU patient data. Furthermore, these models are compared on their use of sequential organ failure (SOFA) scores as well as underlying raw data as input features. METHODOLOGY: Different machine learning techniques are trained, using a 14,480 patient dataset, both on SOFA scores as well as their underlying raw data values from the first five days after admission, in order to predict (i) the patient LOS, and (ii) the patient mortality. Furthermore, to help physicians in assessing the prediction credibility, a probabilistic model is tailored to the output of our best-performing model, assigning a belief to each patient status prediction. A two-by-two grid is built, using the classification outputs of the mortality and prolonged stay predictors to improve the patient LOS regression models. RESULTS: For predicting patient mortality and a prolonged stay, the best performing model is a support vector machine (SVM) with GA,D=65.9% (area under the curve (AUC) of 0.77) and GS,L=73.2% (AUC of 0.82). In terms of LOS regression, the best performing model is support vector regression, achieving a mean absolute error of 1.79 days and a median absolute error of 1.22 days for those patients surviving a nonprolonged stay. CONCLUSION: Using a classification grid based on the predicted patient mortality and prolonged stay, allows more accurate modeling of the patient LOS. The detailed models allow to support the decisions made by physicians in an ICU setting.