A divided and prioritized experience replay approach for streaming regression.

In the streaming learning setting, an agent is presented with a data stream on which to learn from in an online fashion. A common problem is catastrophic forgetting of old knowledge due to updates to the model. Mitigating catastrophic forgetting has received a lot of attention, and a variety of methods exist to solve this problem. In this paper, we present a divided and prioritized experience replay approach for streaming regression, in which relevant observations are retained in the replay, and extra focus is added to poorly estimated observations through prioritization. Using a real-world dataset, the method is compared to the standard sliding window approach. A statistical power analysis is performed, showing how our approach improves performance on rare, important events at a trade-off in performance for more common observations. Close inspections of the dataset are provided, with emphasis on areas where the standard approach fails. A rephrasing of the problem to a binary classification problem is performed to separate common and rare, important events. These results provide an added perspective regarding the improvement made on rare events.•We divide the prediction space in a streaming regression setting•Observations in the experience replay are prioritized for further training by the model's current error.

Frequency Analysis and Feature Reduction Method for Prediction of Cerebral Palsy in Young Infants.

The aim of this paper is to achieve a model for prediction of cerebral palsy based on motion data of young infants. The prediction is formulated as a classification problem to assign each of the infants to one of the healthy or with cerebral palsy groups. Unlike formerly proposed features that are mostly defined in the time domain, this study proposes a set of features derived from frequency analysis of infants' motions. Since cerebral palsy affects the variability of the motions, and frequency analysis is an intuitive way of studying variability, suggested features are suitable and consistent with the nature of the condition. In the current application, a well-known problem, few subjects and many features, was initially encountered. In such a case, most classifiers get trapped in a suboptimal model and, consequently, fail to provide sufficient prediction accuracy. To solve this problem, a feature selection method that determines features with significant predictive ability is proposed. The feature selection method decreases the risk of false discovery and, therefore, the prediction model is more likely to be valid and generalizable for future use. A detailed study is performed on the proposed features and the feature selection method: the classification results confirm their applicability. Achieved sensitivity of 86%, specificity of 92% and accuracy of 91% are comparable with state-of-the-art clinical and expert-based methods for predicting cerebral palsy.

Frequency-based features for early cerebral palsy prediction.

In this paper we aim at predicting cerebral palsy, the most serious and lifelong motor function disorder in children, at an early age by analysing infants' motion data. An essential step for doing so is to extract informative features with high class separability. We propose a set of features derived from frequency analysis of the motion data. Then, we evaluate the practicality of our features on one of the richest data sets collected to study this disease. In this data set, the motion data are extracted from both electromagnetic sensors as well as video camera. The proposed features are used for classifying both data sets. Using these features, we manage to achieve promising classification performance. Classification accuracy of 91% for the sensor data and 88% for the video-derived data show not only the advantage of employing these features for predicting cerebral palsy, but also that replacing electromagnetic sensors with a video camera is feasible.

Video-based early cerebral palsy prediction using motion segmentation.

Analysing distinct motion patterns that occur during infancy can be a way through early prediction of cerebral palsy. This analysis can only be performed by well-trained expert clinicians, and hence can not be widespread, specially in poor countries. In order to decrease the need for experts, computer-based methods can be applied. If individual motions of different body parts are available, these methods could achieve more accurate results with better clinical insight. Thus far, motion capture systems or the like were needed in order to provide such data. However, these systems not only need laboratory and experts to set up the experiment, but they could be intrusive for the infant's motions. In this paper we build up our prediction method on a solution based on a single video camera, that is far less intrusive and a lot cheaper. First, the motions of different body parts are separated, then, motion features are extracted and used to classify infants to healthy or affected. Our experimental results show that visually obtained motion data allows cerebral palsy detection as accurate as state-of-the-art electromagnetic sensor data.

An optical flow-based method to predict infantile cerebral palsy.

Cerebral palsy (CP) is a perinatally acquired nonprogressive brain damage resulting in motor impairment affecting mobility and posture. Early identification of infants with CP is desired to target early interventions and follow-up. During early infancy, distinct motion patterns occur which are highly predictive for later disability. These motor patterns can be observed and recorded. In this paper, a method to predict later CP based on early video recordings of the infants' spontaneous movements, applying optical flow and statistical pattern recognition, is presented. We extract motion information from video recordings of young infants using a total variation related optical flow method. By using wavelet analysis features from motion trajectories of points initiated on a regular grid were extracted and classified using a support vector machine.