Adam Pantanowitz

Evaluating the Impact of Missing Data Imputation

This paper presents an impact assessment for the imputation of missing data. The assessment is performed by measuring the impacts of missing data on the statistical nature of the data, on a classifier, and on a logistic regression system. The data set used is HIV seroprevalence data from an antenatal clinic study survey performed in 2001. Data imputation is performed through the use of Random Forests, selected based on best imputation performance above five other techniques. Test sets are developed which consist of the original data and of imputed data with varying numbers of specifically selected missing variables imputed. Results indicate that, for this data set, the evaluated properties and tested paradigms are fairly immune to missing data imputation. The impact is not highly significant, with, for example, linear correlations of 96 % between HIV status probability prediction with a full set and with a set of two imputed variables using the logistic regression analysis.

Missing Data Imputation Through the Use of the Random Forest Algorithm

This paper presents a comparison of different paradigms used for missing data imputation. The data set used is HIV seroprevalence data from an antenatal clinic study survey performed in 2001. Data imputation is performed through five methods: Random Forests; auto-associative neural networks with genetic algorithms; auto-associative neuro-fuzzy configurations; and two random forest and neural network based hybrids. Results indicate that Random Forests are superior in imputing missing data for the given data set in terms of accuracy and in terms of computation time, with accuracy increases of up to 32 % on average for certain variables when compared with auto-associative networks. While the concept of hybrid systems has promise, the presented systems appear to be hindered by their auto-associative neural network components.

Improvements in feature vector selection and parameter optimisation for continuous gesture recognition

Gesture recognition has attracted significant interest due to diverse potential applications, including: hand writing recognition, robot control and human-computer interfaces. This paper identifies and addresses three shortcomings in current approaches to feature vector selection and parameter optimisation for continuous gesture recognition. First, in selecting the final feature vector, researchers typically analyse only a small subset of possible feature combinations; however, the limited subset is likely to omit the optimum feature vector. Second, selection of the final feature vector is based on performance in isolated recognition; however, the final feature vector may not perform adequately in continuous recognition. No protocol currently exists to evaluate and select the final feature vector in continuous recognition mode, thus a novel scoring system is developed. Finally, optimisation of the number of states in the Hidden Markov Models (HMMs) and the number of clusters (k-means clustering) is performed independently, ignoring any possible interdependency. To investigate and address these shortcomings, a gesture recognition system geared towards sign language interpretation is designed. The system is tested on a 9-word gesture vocabulary, and subsequent analysis confirms the above conjectures: first, the optimum feature vector cannot be intuitively predicted and must be determined through rigorous analysis; second, selecting the final feature vector in continuous mode improved the accuracy score by 5.85 % and the perfect sentence recognition by 47.2 %; finally, optimising the number of states and number of clusters simultaneously improved the accuracy score by 3.0 % and the perfect sentence recognition by 11.1%.

Towards Automated Three-Dimensional Tracking of Nephrons through Stacked Histological Image Sets.

An automated approach for tracking individual nephrons through three-dimensional histological image sets of mouse and rat kidneys is presented. In a previous study, the available images were tracked manually through the image sets in order to explore renal microarchitecture. The purpose of the current research is to reduce the time and effort required to manually trace nephrons by creating an automated, intelligent system as a standard tool for such datasets. The algorithm is robust enough to isolate closely packed nephrons and track their convoluted paths despite a number of nonideal, interfering conditions such as local image distortions, artefacts, and interstitial tissue interference. The system comprises image preprocessing, feature extraction, and a custom graph-based tracking algorithm, which is validated by a rule base and a machine learning algorithm. A study of a selection of automatically tracked nephrons, when compared with manual tracking, yields a 95% tracking accuracy for structures in the cortex, while those in the medulla have lower accuracy due to narrower diameter and higher density. Limited manual intervention is introduced to improve tracking, enabling full nephron paths to be obtained with an average of 17 manual corrections per mouse nephron and 58 manual corrections per rat nephron.

Paediatric Frontal Chest Radiograph Screening with Fine-Tuned Convolutional Neural Networks

Within developing countries, there is a realistic need for technologies that can assist medical practitioners in meeting the increasing demand for patient screening and monitoring. To this end, computer aided diagnosis (CAD) based approaches to chest radiograph screening can be utilised in areas where there is a high burden of diseases such as tuberculosis and pneumonia. In this work, we investigate the efficacy of a purely data-driven approach to chest radiograph classification through the use of fine-tuned convolutional neural networks (CNN). We use two popular CNN models that are pre-trained on a large natural image dataset and two distinct datasets containing paediatric and adult radiographs respectively. Evaluation is performed using a 5-fold cross-validation analysis at an image level. The promising results, with top AUC metrics of 0.87 and 0.84 for the respective datasets, along with several characteristics of our data-driven approach motivate for the use of fine-tuned CNN models within this application of CAD.

Analysis of MEG Signals for Selective Arithmetic Tasks

In this study, magnetoencephalography (MEG) data were recorded as subjects carried out a basic numerical task: deciding whether a number is even or odd. Signal processing techniques were applied to the MEG data so as to characterise the spatial and temporal dynamics of the brain during the decision-making process. Event-related fields (ERFs) were found by averaging all the trials in the time domain. Induced potentials or oscillatory rhythms were found by averaging the time-frequency representations (TFRs) for all the trials. The TFRs were found using the Wavelet transform. The results show that typical ERF components are present just after the onset of the stimulus. N100, P200 and N200 waveforms indicate that the brain carries out higher-order perceptual processing modulated by attention, and that memory may play a critical role in parity selection. TFRs show beta-band synchronisation as the individual concentrates on the mental task, followed by desynchronisation as the motor response is carried out. Activity is pronounced in the left general interpretive area (Wernicke’s area) with a latency of around 610 ms.

Threshold-based Lip Segmentation using Feedback of Shape Information

The shape and movement of the human lips convey valuable visual information which is used in various applications including: automatic lip-reading (ALR), emotion recognition, biometric speaker identification, and virtual face animation. The image processing to extract visual information from the lips typically involves three stages: face detection, location of the region of interest (ROI), and lip segmentation. This research focuses on lip segmentation as the accuracy of this component is crucial to the performance of the overall system. The challenge of lip segmentation arises from variability in the speaker profile (colour, shape, facial hair, make-up); and the dynamic ROI which changes as the teeth and tongue appear during speech movements.

Hybrid BCI for Controlling a Robotic Arm over an IP Network

A fully online EEG-based hybrid brain-computer interface (BCI) is presented. The BCI is used to control a robotic arm in three degrees of freedom. The system utilises the commercially available Emotiv SDK and EPOC neuroheadset. Using facial gestures, test subjects were able to carry out six different actions with an average detection accuracy of 66.1%. SSVEPs are used as a biofeedback mechanism to implement an attention-based “brain switch”. When the user gazes at the light stimulus, an SSVEP is detected in the brain and the system is either activated or deactivated. Since SSVEPs cannot be detected simultaneously with facial expressions due to noise, the “brain switch” may only be used to provide a user with the ability to mentally turn the system on or off, thereby reducing the number of false positives during rest periods. All 13 test subjects used in the experiment had different responses to SSVEP frequencies in the range of 3-20 Hz. Each subject has an optimal or resonant frequency. The average true-positive detection rate or accuracy at each individual’s optimal frequency is 74.2% (11.8% false positives) when using the Minimum Energy Classification (MEC) algorithm. A defining feature of the system is that it is highly extensible. The inter process communication (IPC) framework enables users to interact with multiple client objects over an IP network.