Ivan Štajduhar

Impact of censoring on learning Bayesian networks in survival modelling

OBJECTIVE Bayesian networks are commonly used for presenting uncertainty and covariate interactions in an easily interpretable way. Because of their efficient inference and ability to represent causal relationships, they are an excellent choice for medical decision support systems in diagnosis, treatment, and prognosis. Although good procedures for learning Bayesian networks from data have been defined, their performance in learning from censored survival data has not been widely studied. In this paper, we explore how to use these procedures to learn about possible interactions between prognostic factors and their influence on the variate of interest. We study how censoring affects the probability of learning correct Bayesian network structures. Additionally, we analyse the potential usefulness of the learnt models for predicting the time-independent probability of an event of interest. METHODS AND MATERIALS We analysed the influence of censoring with a simulation on synthetic data sampled from randomly generated Bayesian networks. We used two well-known methods for learning Bayesian networks from data: a constraint-based method and a score-based method. We compared the performance of each method under different levels of censoring to those of the naive Bayes classifier and the proportional hazards model. We did additional experiments on several datasets from real-world medical domains. The machine-learning methods treated censored cases in the data as event-free. RESULTS We report and compare results for several commonly used model evaluation metrics. On average, the proportional hazards method outperformed other methods in most censoring setups. As part of the simulation study, we also analysed structural similarities of the learnt networks. Heavy censoring, as opposed to no censoring, produces up to a 5% surplus and up to 10% missing total arcs. It also produces up to 50% missing arcs that should originally be connected to the variate of interest. CONCLUSION Presented methods for learning Bayesian networks from data can be used to learn from censored survival data in the presence of light censoring (up to 20%) by treating censored cases as event-free. Given intermediate or heavy censoring, the learnt models become tuned to the majority class and would thus require a different approach.

Learning Bayesian networks from survival data using weighting censored instances

Different survival data pre-processing procedures and adaptations of existing machine-learning techniques have been successfully applied to numerous fields in clinical medicine. Zupan et al. (2000) proposed handling censored survival data by assigning distributions of outcomes to shortly observed censored instances. In this paper, we applied their learning technique to two well-known procedures for learning Bayesian networks: a search-and-score hill-climbing algorithm and a constraint-based conditional independence algorithm. The method was thoroughly tested in a simulation study and on the publicly available clinical dataset GBSG2. We compared it to learning Bayesian networks by treating censored instances as event-free and to Cox regression. The results on model performance suggest that the weighting approach performs best when dealing with intermediate censoring. There is no significant difference between the model structures learnt using either the weighting approach or by treating censored instances as event-free, regardless of censoring.

Uncensoring censored data for machine learning: A likelihood-based approach

Various machine learning techniques have been applied to different problems in survival analysis in the last decade. They were usually adapted to learning from censored survival data by using the information on observation time. This includes learning from parts of the data or interventions to the learning algorithms. Efficient models were established in various fields of clinical medicine and bioinformatics. In this paper, we propose a pre-processing method for adapting the censored survival data to be used with ordinary machine learning algorithms. This is done by pre-assigning censored instances a positive or negative outcome according to their features and observation time. The proposed procedure calculates the goodness of fit of each censored instance to both the distribution of positives and the spoiled distribution of negatives in the entire dataset and relabels that instance accordingly. We performed a thorough empirical testing of our method in a simulation study and on two real-world medical datasets, using the naive Bayes classifier and decision trees. When compared to one of the popular ML methods dealing with survival, our method provided good results, especially when applied to heavily censored data.

A profile- and community-driven book recommender system

Book lovers often struggle to find something new to read. The choice of possible books to read can be overwhelming and confusing, making this subjective decision-making process quite difficult. Searching the Internet we often stumble upon opinions and book ratings of people who are strangers to us. We do not know what their favorite categories, authors, publishers and taste in books are, nor whether we should trust them. As a possible solution to the above mentioned problems, we present Pickbooks - a Web application that is partly a social network and partly a book database. It integrates the social aspects of todays popular social networks into a book recommender system, and not the other way around. It features book recommendation based on previously read books or personal preferences. It allows users to message and follow people they like, see their updates, reviews and get recommendations from them, and vice versa. The users are encouraged to rate a book they have read before, as those ratings affect the global lists of top rated and currently popular books. There are other features which are in the works and will allow us to further improve the presented application and attract more audience.

Denoising of X-ray Images Using the Adaptive Algorithm Based on the LPA-RICI Algorithm

Diagnostics and treatments of numerous diseases are highly dependent on the quality of captured medical images. However, noise (during both acquisition and transmission) is one of the main factors that reduce their quality. This paper proposes an adaptive image denoising algorithm applied to enhance X-ray images. The algorithm is based on the modification of the intersection of confidence intervals (ICI) rule, called relative intersection of confidence intervals (RICI) rule. For each image pixel apart, a 2D mask of adaptive size and shape is calculated and used in designing the 2D local polynomial approximation (LPA) filters for noise removal. One of the advantages of the proposed method is the fact that the estimation of the noise free pixel is performed independently for each image pixel and thus, the method is applicable for easy parallelization in order to improve its computational efficiency. The proposed method was compared to the Gaussian smoothing filters, total variation denoising and fixed size median filtering and was shown to outperform them both visually and in terms of the peak signal-to-noise ratio (PSNR) by up to 7.99 dB.

Predicting stock market trends using random forests: A sample of the Zagreb stock exchange

Stock market prediction is considered to be a challenging task for both investors and researchers, due to its profitability and intricate complexity. Highly accurate stock market predictive models are very often the basis for the construction of algorithms used in automated trading. In this paper, 5-days-ahead and 10-days-ahead predictive models are built using the random forests algorithm. The models are built on the historical data of the CROBEX index and on a few companies listed at the Zagreb Stock Exchange from various sectors. Several technical indicators, popular in quantitative analysis of stock markets, are selected as model inputs. The proposed method is empirically evaluated using stratified 10-fold cross-validation, achieving an average classification accuracy of 76.5% for 5-days-ahead models and 80.8% for 10-days-ahead models.

TFD thresholding in estimating the number of EEG components and the dominant if using the short-term rényi entropy.

Time-frequency (TF) based EEG signal analysis using the local or short-term Rényi entropy often requires low-energy cross-terms and noise suppression prior to the estimation of the local number of components and the dominant component instantaneous frequency (IF). This can be easily accomplished by thresholding in the TF domain with the preset TF threshold value, often chosen empirically. The paper investigates the sensitivity of the method based on the local Rényi entropy to the chosen threshold value. The study was performed on real-life left and right hand movements EEG signals. As shown in the paper, the number of the EEG components extracted using the short-term Rényi entropy is highly sensitive to the chosen TF threshold value, unlike the dominant IF which was shown to be highly robust to TF thresholding. Hence, characterization of the EEG signals using the short-term Rényi entropy should include both detecting the number of EEG components and the dominant component IF estimation.

Mirroring quasi-symmetric organ observations for reducing problem complexity

Abstract Following an obvious growth of available collections of medical images in recent years, both in number and in size, machine learning has nowadays become an important tool for solving various image-analysis-related problems, such as organ segmentation or injury/pathology detection. The potential of learning algorithms to produce models having good generalisation properties is highly dependent on model complexity and the amount of available data. Bearing in mind that complex concepts require the use of complex models, it is of paramount importance to mitigate representation complexity, where possible, therefore enabling the utilisation of simpler models for performing the same task. When dealing with image collections of quasi-symmetric organs, or imaging observations of organs taken from different quasi-symmetric perspectives, one way of reducing representation complexity would be aligning all the images in a collection for left-right or front-rear orientation. That way, a learning algorithm would not be dealing with learning redundant symmetric representations. In this paper, we study in detail the influence of such within-class variation on model complexity, and present a possible solution, that can be applied to medical-imaging computer-aided diagnosis systems. The proposed method involves compacting the data, extracting features and then learning to separate the mirror-image representation classes from one another. Two efficient approaches are considered for performing such orientation separation: a fully automated unsupervised approach and a semi-automated supervised approach. Both solutions are directly applicable to imaging data. Method performance is illustrated on two 2D and one 3D real-world publicly-available medical datasets, concerning different parts of human anatomy, and observed using different imaging techniques: colour fundus photography, mammography CT scans and volumetric knee-joint MR scans. Experimental results suggest that efficient organ-mirroring orientation-classifier models, having expected classification accuracy greater than 99%, can be estimated using either the unsupervised or the supervised approach. In the presence of noise, however, an equally good performance can be achieved only by using the supervised approach, learning from a small subset of labelled data.

Semantic approach to accommodation & booking related Web services

There are several types of information that tourists most often search for when planning a trip. Commonly, these are information about the location, travel, accommodation and nearby attractions. Almost all the knowledge that an average tourist needs already exists, but the current knowledge is highly decoupled and its aggregation is time consuming. Another problem with tourism-related Web services is that they do not correlate with each other. The information they share is often represented differently, is not connected and is sometimes even in contradiction. In this paper, we provide an overview of the differences in data representation between most popular Web tourist services, and introduce the A&B ontology that covers their data representation and enables the service discovery. Additionally, we propose a framework that is connected to the aforementioned Web services, translates the data to the A&B ontology, links the data and provides a single endpoint for multiple Web services.