Livia Jakaite

Feature Selection for Bayesian Evaluation of Trauma Death Risk

In the last year more than 70,000 people have been brought to the UK hospitals with serious injuries. Each time a clinician has to urgently take a patient through a screening procedure to make a reliable decision on the trauma treatment. Typically, such procedure comprises around 20 tests; however the condition of a trauma patient remains very difficult to be tested properly. What happens if these tests are ambiguously interpreted, and information about the severity of the injury will come misleading? The mistake in a decision can be fatal — using a mild treatment can put a patient at risk of dying from posttraumatic shock, while using an overtreatment can also cause death. How can we reduce the risk of the death caused by unreliable decisions? It has been shown that probabilistic reasoning, based on the Bayesian methodology of averaging over decision models, allows clinicians to evaluate the uncertainty in decision making. Based on this methodology, in this paper we aim at selecting the most important screening tests, keeping a high performance. We assume that the probabilistic reasoning within the Bayesian methodology allows us to discover new relationships between the screening tests and uncertainty in decisions. In practice, selection of the most informative tests can also reduce the cost of a screening procedure in trauma care centers. In our experiments we use the UK Trauma data to compare the efficiency of the proposed technique in terms of the performance. We also compare the uncertainty in decisions in terms of entropy.

Comparing robustness of pairwise and multiclass neural-network systems for face recognition

Noise, corruptions, and variations in face images can seriously hurt the performance of face-recognition systems. To make these systems robust to noise and corruptions in image data, multiclass neural networks capable of learning from noisy data have been suggested. However on large face datasets such systems cannot provide the robustness at a high level. In this paper, we explore a pairwise neural-network system as an alternative approach to improve the robustness of face recognition. In our experiments, the pairwise recognition system is shown to outperform the multiclass-recognition system in terms of the predictive accuracy on the test face images.

Bayesian Decision Trees for predicting survival of patients: A study on the US National Trauma Data Bank

Trauma and Injury Severity Score (TRISS) models have been developed for predicting the survival probability of injured patients the majority of which obtain up to three injuries in six body regions. Practitioners have noted that the accuracy of TRISS predictions is unacceptable for patients with a larger number of injuries. Moreover, the TRISS method is incapable of providing accurate estimates of predictive density of survival, that are required for calculating confidence intervals. In this paper we propose Bayesian inference for estimating the desired predictive density. The inference is based on decision tree models which split data along explanatory variables, that makes these models interpretable. The proposed method has outperformed the TRISS method in terms of accuracy of prediction on the cases recorded in the US National Trauma Data Bank. The developed method has been made available for evaluation purposes as a stand-alone application.

Prediction of survival probabilities with Bayesian Decision Trees

Abstract Practitioners use Trauma and Injury Severity Score (TRISS) models for predicting the survival probability of an injured patient. The accuracy of TRISS predictions is acceptable for patients with up to three typical injuries, but unacceptable for patients with a larger number of injuries or with atypical injuries. Based on a regression model, the TRISS methodology does not provide the predictive density required for accurate assessment of risk. Moreover, the regression model is difficult to interpret. We therefore consider Bayesian inference for estimating the predictive distribution of survival. The inference is based on decision tree models which recursively split data along explanatory variables, and so practitioners can understand these models. We propose the Bayesian method for estimating the predictive density and show that it outperforms the TRISS method in terms of both goodness-of-fit and classification accuracy. The developed method has been made available for evaluation purposes as a stand-alone application.

Evolving polynomial neural networks for detecting abnormal patterns

Abnormal patterns, existing e.g. in raw data, affect decision making process and have to be accurately detected and removed in order to reduce the risk of making wrong decisions. Existing Machine Learning (ML) approaches known from the literature require the user to set and experimentally adjust parameters of a decision model to achieve the best result. When artificial neural networks (ANNs) are employed, a typical problem is setting of a proper network structure and learning parameters that are required to minimise possible over-fitting. We propose a new evolutionary strategy of learning an ANN structure of a near optimal connectivity from the given data and show that such structures are less prone to over-fitting. The proposed method starts to learn with one input variable and one neuron and then adds a new input and a new neuron to the network while its validation error decreases. The resultant ANN consists of a reasonably small number of neurons that are concisely described by a set of short-term polynomial functions of variables that make a distinct contribution to the output. The proposed technique has been tested on the ML benchmarks and the results showed that the performance is comparable with that obtained by the conventional ML methods that require ad hoc tuning.

Bayesian averaging over decision tree models: An application for estimating uncertainty in trauma severity scoring

INTRODUCTION For making reliable decisions, practitioners need to estimate uncertainties that exist in data and decision models. In this paper we analyse uncertainties of predicting survival probability for patients in trauma care. The existing prediction methodology employs logistic regression modelling of Trauma and Injury Severity Score (TRISS), which is based on theoretical assumptions. These assumptions limit the capability of TRISS methodology to provide accurate and reliable predictions. METHODS We adopt the methodology of Bayesian model averaging and show how this methodology can be applied to decision trees in order to provide practitioners with new insights into the uncertainty. The proposed method has been validated on a large set of 447,176 cases registered in the US National Trauma Data Bank in terms of discrimination ability evaluated with receiver operating characteristic (ROC) and precision-recall (PRC) curves. RESULTS Areas under curves were improved for ROC from 0.951 to 0.956 (p = 3.89 × 10-18) and for PRC from 0.564 to 0.605 (p = 3.89 × 10-18). The new model has significantly better calibration in terms of the Hosmer-Lemeshow Hˆ statistic, showing an improvement from 223.14 (the standard method) to 11.59 (p = 2.31 × 10-18). CONCLUSION The proposed Bayesian method is capable of improving the accuracy and reliability of survival prediction. The new method has been made available for evaluation purposes as a web application.

Bayesian model averaging over decision trees for assessing newborn brain maturity from electroencephalogram

We use the Bayesian Model Averaging (BMA) over Decision Trees (DTs) for assessing newborn brain maturity from clinical EEG. We found that within this methodology an appreciable part of EEG features is rarely used in the DT models, because these features make weak contribution to the assessment. It was identified that the portion of DT models using weak EEG features is large. The negative impact of this is twofold. First, the use of weak features obstructs interpretation of DTs. Second, weak attributes increase dimensionality of a model parameter space needed to be explored in detail. We assumed that discarding the DTs using weak features will reduce the negative impact, and then proposed a new technique. This technique has been tested on some benchmark problems, and the results have shown that the original set of attributes can be reduced without a distinguishable decrease in BMA performance. On the EEG data, we found that the original set of features can be reduced from 36 to 12. Rerunning the BMA on the set of the 12 EEG features has slightly improved the performance.