Advances in neuroimaging, genomic, motion tracking, eye-tracking and many other technology-based data collection methods have led to a torrent of high dimensional datasets, which commonly have a small number of samples because of the intrinsic high cost of data collection involving human participants. High dimensional data with a small number of samples is of critical importance for identifying biomarkers and conducting feasibility and pilot work, however it can lead to biased machine learning (ML) performance estimates. Our review of studies which have applied ML to predict autistic from non-autistic individuals showed that small sample size is associated with higher reported classification accuracy. Thus, we have investigated whether this bias could be caused by the use of validation methods which do not sufficiently control overfitting. Our simulations show that K-fold Cross-Validation (CV) produces strongly biased performance estimates with small sample sizes, and the bias is still evident with sample size of 1000. Nested CV and train/test split approaches produce robust and unbiased performance estimates regardless of sample size. We also show that feature selection if performed on pooled training and testing data is contributing to bias considerably more than parameter tuning. In addition, the contribution to bias by data dimensionality, hyper-parameter space and number of CV folds was explored, and validation methods were compared with discriminable data. The results suggest how to design robust testing methodologies when working with small datasets and how to interpret the results of other studies based on what validation method was used.

The electroencephalogram (EEG) is a classic noninvasive method for measuring a person's brainwaves and is used in a large number of fields: from epilepsy and sleep disorder diagnosis to brain-computer interfaces (BCIs). Electrodes are placed on the scalp to detect the microvolt-sized signals that result from synchronized neuronal activity within the brain. Current long-term EEG monitoring is generally either carried out as an inpatient in combination with video recording and long cables to an amplifier and recording unit or is ambulatory. In the latter, the EEG recorder is portable but bulky, and in principle, the subject can go about their normal daily life during the recording. In practice, however, this is rarely the case. It is quite common for people undergoing ambulatory EEG monitoring to take time off work and stay at home rather than be seen in public with such a device. Wearable EEG is envisioned as the evolution of ambulatory EEG units from the bulky, limited lifetime devices available today to small devices present only on the head that can record EEG for days, weeks, or months at a time [see Figure 1(a) and (b)]. Such miniaturized units could enable prolonged monitoring of chronic conditions such as epilepsy and greatly improve the end-user acceptance of BCI systems. In this article, we aim to provide a review and overview of wearable EEG technology, answering the questions: What is it, why is it needed, and what does it entail? We first investigate the requirements of portable EEG systems and then link these to the core applications of wearable EEG technology: epilepsy diagnosis, sleep disorder diagnosis, and BCIs. As a part of our review, we asked 21 neurologists (as a key user group) for their views on wearable EEG. This group highlighted that wearable EEG will be an essential future tool. Our descriptions here will focus mainly on epilepsyand the medical applications of wearable EEG, as this is the historical background of the EEG, our area of expertise, and a core motivating area in itself, but we will also discuss the other application areas. We continue by considering the forthcoming research challenges, principally new electrode technology and lower power electronics, and we outline our approach for dealing with the electronic power issues. We believe that the optimal approach to realizing wearable EEG technology is not to optimize any one part but to find the best set of tradeoffs at both the system and implementation level. In this article, we discuss two of these tradeoffs in detail: investigating the online compression of EEG data to reduce the system power consumption and the optimal method for providing this data compression.

Wearable heart rate sensors such as those found in smartwatches are commonly based upon Photoplethysmography (PPG) which shines a light into the wrist and measures the amount of light reflected back. This method works well for stationary subjects, but in exercise situations, PPG signals are heavily corrupted by motion artifacts. The presence of these artifacts necessitates the creation of signal processing algorithms for removing the motion interference and allowing the true heart related information to be extracted from the PPG trace during exercise. Here, we describe a new publicly available database of PPG signals collected during exercise for the creation and validation of signal processing algorithms extracting heart rate and heart rate variability from PPG signals. PPG signals from the wrist are recorded together with chest electrocardiography (ECG) to allow a reference/comparison heart rate to be found, and the temporal alignment between the two signal sets is estimated from the signal timestamps. The new database differs from previously available public databases because it includes wrist PPG recorded during walking, running, easy bike riding and hard bike riding. It also provides estimates of the wrist movement recorded using a 3-axis low-noise accelerometer, a 3-axis wide-range accelerometer, and a 3-axis gyroscope. The inclusion of gyroscopic information allows, for the first time, separation of acceleration due to gravity and acceleration due to true motion of the sensor. The hypothesis is that the improved motion information provided could assist in the development of algorithms with better PPG motion artifact removal performance.

This study identifies characteristic features in scalp EEG that simultaneously give the best discrimination between epileptic seizures and background EEG in minimally pre-processed scalp data; and have minimal computational complexity to be suitable for online, real-time analysis. The discriminative performance of 65 previously reported features has been evaluated in terms of sensitivity, specificity, area under the sensitivity–specificity curve (AUC), and relative computational complexity, on 47 seizures (split in 2,698 2 s sections) in over 172 h of scalp EEG from 24 adults. The best performing features are line length and relative power in the 12.5–25 Hz band. Relative power has a better seizure detection performance (AUC = 0.83; line length AUC = 0.77), but is calculated after the discrete wavelet transform and is thus more computationally complex. Hence, relative power achieves the best performance for offline detection, whilst line length would be preferable for online low complexity detection. These results, from the largest systematic study of seizure detection features, aid future researchers in selecting an optimal set of features when designing algorithms for both standard offline detection and new online low computational complexity detectors.

Edge computing aims to integrate computing into everyday settings, enabling the system to be context-aware and private to the user. With the increasing success and popularity of deep learning methods, there is an increased demand to leverage these techniques in mobile and wearable computing scenarios. In this paper, we present an assessment of a deep human activity recognition system’s memory and execution time requirements, when implemented on a mid-range smartphone class hardware and the memory implications for embedded hardware. This paper presents the design of a convolutional neural network (CNN) in the context of human activity recognition scenario. Here, layers of CNN automate the feature learning and the influence of various hyper-parameters such as the number of filters and filter size on the performance of CNN. The proposed CNN showed increased robustness with better capability of detecting activities with temporal dependence compared to models using statistical machine learning techniques. The model obtained an accuracy of 96.4% in a five-class static and dynamic activity recognition scenario. We calculated the proposed model memory consumption and execution time requirements needed for using it on a mid-range smartphone. Per-channel quantization of weights and per-layer quantization of activation to 8-bits of precision post-training produces classification accuracy within 2% of floating-point networks for dense, convolutional neural network architecture. Almost all the size and execution time reduction in the optimized model was achieved due to weight quantization. We achieved more than four times reduction in model size when optimized to 8-bit, which ensured a feasible model capable of fast on-device inference.