Orthogonal Search-Based Rule Extraction (OSRE) for Trained Neural Networks: A Practical and Efficient Approach

Deep Logic Networks: Inserting and Extracting Knowledge From Deep Belief Networks

Developments in deep learning have seen the use of layerwise unsupervised learning combined with supervised learning for fine-tuning. With this layerwise approach, a deep network can be seen as a more modular system that lends itself well to learning representations. In this paper, we investigate whether such modularity can be useful to the insertion of background knowledge into deep networks, whether it can improve learning performance when it is available, and to the extraction of knowledge from trained deep networks, and whether it can offer a better understanding of the representations learned by such networks. To this end, we use a simple symbolic language-a set of logical rules that we call confidence rules-and show that it is suitable for the representation of quantitative reasoning in deep networks. We show by knowledge extraction that confidence rules can offer a low-cost representation for layerwise networks (or restricted Boltzmann machines). We also show that layerwise extraction can produce an improvement in the accuracy of deep belief networks. Furthermore, the proposed symbolic characterization of deep networks provides a novel method for the insertion of prior knowledge and training of deep networks. With the use of this method, a deep neural-symbolic system is proposed and evaluated, with the experimental results indicating that modularity through the use of confidence rules and knowledge insertion can be beneficial to network performance.

Rule Extraction from Training Data Using Neural Network

Data Mining is a powerful technology to help organization to concentrate on most important data by extracting useful information from large database. One of the most commonly used techniques in data mining is Artificial Neural Network due to its high performance in many application domains. Despite many advantages of Artificial Neural Network, one of its main drawbacks is its inherent black box nature which is the main problem of using Artificial Neural Network in data mining. Therefore, this paper proposes a rule extraction algorithm from neural network using classified and misclassified data to convert the black box nature of Artificial Neural Network into a white box. The proposed algorithm is a modification of the existing algorithm, Rule Extraction by Reverse Engineering (RxREN). The proposed algorithm extracts rules from trained neural network for datasets with mixed mode attributes using pedagogical approach. The proposed algorithm uses both classified as well as misclassified data to find out the data ranges of significant attributes in respective classes, which is the innovation of the proposed algorithm. The experimental results clearly show that the performance of the proposed algorithm is superior to existing algorithms.

Risk predictions for individual patients from logistic regression were visualized with bar-line charts

OBJECTIVE

The interface of a computerized decision support system is crucial for its acceptance among end users. We demonstrate how combined bar-line charts can be used to visualize predictions for individual patients from logistic regression models.

STUDY DESIGN AND SETTING

Data from a previous diagnostic study aiming at predicting the immediate risk of acute coronary syndrome (ACS) among 634 patients presenting to an emergency department with chest pain were used. Risk predictions from the logistic regression model were presented for four hypothetical patients in bar-line charts with bars representing empirical Bayes adjusted likelihood ratios (LRs) and the line representing the estimated probability of ACS, sequentially updated from left to right after assessment of each risk factor.

RESULTS

Two patients had similar low risk for ACS but quite different risk profiles according to the bar-line charts. Such differences in risk profiles could not be detected from the estimated ACS risk alone. The bar-line charts also highlighted important but counteracted risk factors in cases where the overall LR was less informative (close to one).

CONCLUSION

The proposed graphical technique conveys additional information from the logistic model that can be important for correct diagnosis and classification of patients and appropriate medical management.

Prediction of persistence of combined evidence-based cardiovascular medications in patients with acute coronary syndrome after hospital discharge using neural networks

In the PREVENIR-5 study, artificial neural networks (NN) were applied to a large sample of patients with recent first acute coronary syndrome (ACS) to identify determinants of persistence of evidence-based cardiovascular medications (EBCM: antithrombotic + beta-blocker + statin + angiotensin converting enzyme inhibitor-ACEI and/or angiotensin-II receptor blocker-ARB). From October 2006 to April 2007, 1,811 general practitioners recruited 4,850 patients with a mean time of ACS occurrence of 24 months. Patient profile for EBCM persistence was determined using automatic rule generation from NN. The prediction accuracy of NN was compared with that of logistic regression (LR) using Area Under Receiver-Operating Characteristics-AUROC. At hospital discharge, EBCM was prescribed to 2,132 patients (44%). EBCM persistence rate, 24 months after ACS, was 86.7%. EBCM persistence profile combined overweight, hypercholesterolemia, no coronary artery bypass grafting and low educational level (Positive Predictive Value = 0.958). AUROC curves showed better predictive accuracy for NN compared to LR models.

Extracting rules from neural networks as decision diagrams

Rule extraction from neural networks (NNs) solves two fundamental problems: it gives insight into the logic behind the network and in many cases, it improves the network's ability to generalize the acquired knowledge. This paper presents a novel eclectic approach to rule extraction from NNs, named LOcal Rule Extraction (LORE), suited for multilayer perceptron networks with discrete (logical or categorical) inputs. The extracted rules mimic network behavior on the training set and relax this condition on the remaining input space. First, a multilayer perceptron network is trained under standard regime. It is then transformed into an equivalent form, returning the same numerical result as the original network, yet being able to produce rules generalizing the network output for cases similar to a given input. The partial rules extracted for every training set sample are then merged to form a decision diagram (DD) from which logic rules can be extracted. A rule format explicitly separating subsets of inputs for which an answer is known from those with an undetermined answer is presented. A special data structure, the decision diagram, allowing efficient partial rule merging is introduced. With regard to rules' complexity and generalization abilities, LORE gives results comparable to those reported previously. An algorithm transforming DDs into interpretable boolean expressions is described. Experimental running times of rule extraction are proportional to the network's training time.

Guiding hidden layer representations for improved rule extraction from neural networks

The production of relatively large and opaque weight matrices by error backpropagation learning has inspired substantial research on how to extract symbolic human-readable rules from trained networks. While considerable progress has been made, the results at present are still relatively limited, in part due to the large numbers of symbolic rules that can be generated. Most past work to address this issue has focused on progressively more powerful methods for rule extraction (RE) that try to minimize the number of weights and/or improve rule expressiveness. In contrast, here we take a different approach in which we modify the error backpropagation training process so that it learns a different hidden layer representation of input patterns than would normally occur. Using five publicly available datasets, we show via computational experiments that the modified learning method helps to extract fewer rules without increasing individual rule complexity and without decreasing classification accuracy. We conclude that modifying error backpropagation so that it more effectively separates learned pattern encodings in the hidden layer is an effective way to improve contemporary RE methods.

Determination of mode of ventilation using OSRE

This study classifies the mode of ventilation using respiratory rate, inhaled and exhaled carbon dioxide concentrations in anaesthetised patients. Thirty seven patients were breathing spontaneously (SPONT) and 50 were on a ventilator (intermittent positive pressure ventilation, IPPV). A data-based methodology for rule inference from trained neural networks, orthogonal search-based rule extraction, identified two sets of low-order Boolean rules for differential identification of the mode of ventilation. Combining both models produced three possible outcomes; IPPV, SPONT and 'Uncertain'. The true positive rates were approximately maintained at 96% for IPPV and 93% for SPONT, with false positive rates of 0.4% for each category and 4.3% 'Uncertain' inferences.

How to find simple and accurate rules for viral protease cleavage specificities

Background

Proteases of human pathogens are becoming increasingly important drug targets, hence it is necessary to understand their substrate specificity and to interpret this knowledge in practically useful ways. New methods are being developed that produce large amounts of cleavage information for individual proteases and some have been applied to extract cleavage rules from data. However, the hitherto proposed methods for extracting rules have been neither easy to understand nor very accurate. To be practically useful, cleavage rules should be accurate, compact, and expressed in an easily understandable way.

Results

A new method is presented for producing cleavage rules for viral proteases with seemingly complex cleavage profiles. The method is based on orthogonal search-based rule extraction (OSRE) combined with spectral clustering. It is demonstrated on substrate data sets for human immunodeficiency virus type 1 (HIV-1) protease and hepatitis C (HCV) NS3/4A protease, showing excellent prediction performance for both HIV-1 cleavage and HCV NS3/4A cleavage, agreeing with observed HCV genotype differences. New cleavage rules (consensus sequences) are suggested for HIV-1 and HCV NS3/4A cleavages. The practical usability of the method is also demonstrated by using it to predict the location of an internal cleavage site in the HCV NS3 protease and to correct the location of a previously reported internal cleavage site in the HCV NS3 protease. The method is fast to converge and yields accurate rules, on par with previous results for HIV-1 protease and better than previous state-of-the-art for HCV NS3/4A protease. Moreover, the rules are fewer and simpler than previously obtained with rule extraction methods.

Conclusion

A rule extraction methodology by searching for multivariate low-order predicates yields results that significantly outperform existing rule bases on out-of-sample data, but are more transparent to expert users. The approach yields rules that are easy to use and useful for interpreting experimental data.

Exploring new possibilities for case-based explanation of artificial neural network ensembles

Artificial neural network (ANN) ensembles have long suffered from a lack of interpretability. This has severely limited the practical usability of ANNs in settings where an erroneous decision can be disastrous. Several attempts have been made to alleviate this problem. Many of them are based on decomposing the decision boundary of the ANN into a set of rules. We explore and compare a set of new methods for this explanation process on two artificial data sets (Monks 1 and 3), and one acute coronary syndrome data set consisting of 861 electrocardiograms (ECG) collected retrospectively at the emergency department at Lund University Hospital. The algorithms managed to extract good explanations in more than 84% of the cases. More to the point, the best method provided 99% and 91% good explanations in Monks data 1 and 3 respectively. Also there was a significant overlap between the algorithms. Furthermore, when explaining a given ECG, the overlap between this method and one of the physicians was the same as the one between the two physicians in this study. Still the physicians were significantly, p-value<0.001, more similar to each other than to any of the methods. The algorithms have the potential to be used as an explanatory aid when using ANN ensembles in clinical decision support systems.

Recursive neural network rule extraction for data with mixed attributes

In this paper, we present a recursive algorithm for extracting classification rules from feedforward neural networks (NNs) that have been trained on data sets having both discrete and continuous attributes. The novelty of this algorithm lies in the conditions of the extracted rules: the rule conditions involving discrete attributes are disjoint from those involving continuous attributes. The algorithm starts by first generating rules with discrete attributes only to explain the classification process of the NN. If the accuracy of a rule with only discrete attributes is not satisfactory, the algorithm refines this rule by recursively generating more rules with discrete attributes not already present in the rule condition, or by generating a hyperplane involving only the continuous attributes. We show that for three real-life credit scoring data sets, the algorithm generates rules that are not only more accurate but also more comprehensible than those generated by other NN rule extraction methods.

Development of a rule based prognostic tool for HER 2 positive breast cancer patients

A three stage development process for the production of a hierarchical rule based prognosis tool is described. The application for this tool is specific to breast cancer patients that have a positive expression of the HER 2 gene. The first stage is the development of a Bayesian classification neural network to classify for cancer specific mortality. Secondly, low-order Boolean rules are extracted form this model using an Orthogonal Search based Rule Extraction (OSRE) algorithm. Further to these rules additional information is gathered from the Kaplan-Meier survival estimates of the population, stratified by the categorizations of the input variables. Finally, expert knowledge is used to further simplify the rules and to rank them hierarchically in the form of a decision tree. The resulting decision tree groups all observations into specific categories by clinical profile and by event rate. The practical clinical value of this decision support tool will in future be tested by external validation with additional data from other clinical centres.

An integrated framework for risk profiling of breast cancer patients following surgery

OBJECTIVE

An integrated decision support framework is proposed for clinical oncologists making prognostic assessments of patients with operable breast cancer. The framework may be delivered over a web interface. It comprises a triangulation of prognostic modelling, visualisation of historical patient data and an explanatory facility to interpret risk group assignments using empirically derived Boolean rules expressed directly in clinical terms.

METHODS AND MATERIALS

The prognostic inferences in the interface are validated in a multicentre longitudinal cohort study by modelling retrospective data from 917 patients recruited at Christie Hospital, Wilmslow between 1983 and 1989 and predicting for 931 patients recruited in the same centre during 1990-1993. There were also 291 patients recruited between 1984 and 1998 at the Clatterbridge Centre for Oncology and the Linda McCartney Centre, Liverpool, UK.

RESULTS AND CONCLUSIONS

There are three novel contributions relating this paper to breast cancer cases. First, the widely used Nottingham prognostic index (NPI) is enhanced with additional clinical features from which prognostic assessments can be made more specific for patients in need of adjuvant treatment. This is shown with a cross matching of the NPI and a new prognostic index which also provides a two-dimensional visualisation of the complete patient database by risk of negative outcome. Second, a principled rule-extraction method, orthogonal search rule extraction, generates readily interpretable explanations of risk group allocations derived from a partial logistic artificial neural network with automatic relevance determination (PLANN-ARD). Third, 95% confidence intervals for individual predictions of survival are obtained by Monte Carlo sampling from the PLANN-ARD model.

Time-to-event analysis with artificial neural networks: an integrated analytical and rule-based study for breast cancer

This paper presents an analysis of censored survival data for breast cancer specific mortality and disease-free survival. There are three stages to the process, namely time-to-event modelling, risk stratification by predicted outcome and model interpretation using rule extraction. Model selection was carried out using the benchmark linear model, Cox regression but risk staging was derived with Cox regression and with Partial Logistic Regression Artificial Neural Networks regularised with Automatic Relevance Determination (PLANN-ARD). This analysis compares the two approaches showing the benefit of using the neural network framework especially for patients at high risk. The neural network model also has results in a smooth model of the hazard without the need for limiting assumptions of proportionality. The model predictions were verified using out-of-sample testing with the mortality model also compared with two other prognostic models called TNG and the NPI rule model. Further verification was carried out by comparing marginal estimates of the predicted and actual cumulative hazards. It was also observed that doctors seem to treat mortality and disease-free models as equivalent, so a further analysis was performed to observe if this was the case. The analysis was extended with automatic rule generation using Orthogonal Search Rule Extraction (OSRE). This methodology translates analytical risk scores into the language of the clinical domain, enabling direct validation of the operation of the Cox or neural network model. This paper extends the existing OSRE methodology to data sets that include continuous-valued variables.

Orthogonal search-based rule extraction for modelling the decision to transfuse

Data from an audit relating to transfusion decisions during intermediate or major surgery were analysed to determine the strengths of certain factors in the decision making process. The analysis, using orthogonal search-based rule extraction (OSRE) from a trained neural network, demonstrated that the risk of tissue hypoxia (ROTH) assessed using a 100-mm visual analogue scale, the haemoglobin value (Hb) and the presence or absence of on-going haemorrhage (OGH) were able to reproduce the transfusion decisions with a joint specificity of 0.96 and sensitivity of 0.93 and a positive predictive value of 0.9. The rules indicating transfusion were: 1. ROTH > 32 mm and Hb < 94 g x l(-1); 2. ROTH > 13 mm and Hb < 87 g x l(-1); 3. ROTH > 38 mm, Hb < 102 g x l(-1) and OGH; 4. Hb < 78 g x l(-1).