Significant Features for Human Activity Recognition Using Tri-Axial Accelerometers.

Activity recognition using wearable sensors has become essential for a variety of applications. Tri-axial accelerometers are the most widely used sensor for activity recognition. Although various features have been used to capture patterns and classify the accelerometer signals to recognise activities, there is no consensus on the best features to choose. Reducing the number of features can reduce the computational cost and complexity and enhance the performance of the classifiers. This paper identifies the signal features that have significant discriminative power between different human activities. It also investigates the effect of sensor placement location, the sampling frequency, and activity complexity on the selected features. A comprehensive list of 193 signal features has been extracted from accelerometer signals of four publicly available datasets, including features that have never been used before for activity recognition. Feature significance was measured using the Joint Mutual Information Maximisation (JMIM) method. Common significant features among all the datasets were identified. The results show that the sensor placement location does not significantly affect recognition performance, nor does it affect the significant sub-set of features. The results also showed that with high sampling frequency, features related to signal repeatability and regularity show high discriminative power.

Digital Detectives: Websleuthing Reduces Eyewitness Identification Accuracy in Police Lineups.

Eyewitnesses to crimes sometimes search for a culprit on social media before viewing a police lineup, but it is not known whether this affects subsequent lineup identification accuracy. The present online study was conducted to address this. Two hundred and eighty-five participants viewed a mock crime video, and after a 15-20 min delay either (i) viewed a mock social media site including the culprit, (ii) viewed a mock social media site including a lookalike, or (iii) completed a filler task. A week later, participants made an identification from a photo lineup. It was predicted that searching for a culprit on social media containing the lookalike (rather than the culprit) would reduce lineup identification accuracy. There was a significant association between social media exposure and lineup accuracy for the Target Present lineup (30% more of the participants who saw the lookalike on social media failed to positively identify the culprit than participants in the other conditions), but for the Target Absent lineup (which also included the lookalike) there was no significant association with lineup identification accuracy. The results suggest that if an eyewitness sees a lookalike (where they are expecting to see the culprit) when conducting a self-directed search on social media, they are less likely to subsequently identify the culprit in the formal ID procedure.

Minimal Patient Clinical Variables to Accurately Predict Stress Echocardiography Outcome: Validation Study Using Machine Learning Techniques.

BACKGROUND: Stress echocardiography is a well-established diagnostic tool for suspected coronary artery disease (CAD). Cardiovascular risk factors are used in the assessment of the probability of CAD. The link between the outcome of stress echocardiography and patients' variables including risk factors, current medication, and anthropometric variables has not been widely investigated. OBJECTIVE: This study aimed to use machine learning to predict significant CAD defined by positive stress echocardiography results in patients with chest pain based on anthropometrics, cardiovascular risk factors, and medication as variables. This could allow clinical prioritization of patients with likely prediction of CAD, thus saving clinician time and improving outcomes. METHODS: A machine learning framework was proposed to automate the prediction of stress echocardiography results. The framework consisted of four stages: feature extraction, preprocessing, feature selection, and classification stage. A mutual information-based feature selection method was used to investigate the amount of information that each feature carried to define the positive outcome of stress echocardiography. Two classification algorithms, support vector machine (SVM) and random forest classifiers, have been deployed. Data from 529 patients were used to train and validate the framework. Patient mean age was 61 (SD 12) years. The data consists of anthropological data and cardiovascular risk factors such as gender, age, weight, family history, diabetes, smoking history, hypertension, hypercholesterolemia, prior diagnosis of CAD, and prescribed medications at the time of the test. There were 82 positive (abnormal) and 447 negative (normal) stress echocardiography results. The framework was evaluated using the whole dataset including cases with prior diagnosis of CAD. Five-fold cross-validation was used to validate the performance of the framework. We also investigated the model in the subset of patients with no prior CAD. RESULTS: The feature selection methods showed that prior diagnosis of CAD, sex, and prescribed medications such as angiotensin-converting enzyme inhibitor/angiotensin receptor blocker were the features that shared the most information about the outcome of stress echocardiography. SVM classifiers showed the best trade-off between sensitivity and specificity and was achieved with three features. Using only these three features, we achieved an accuracy of 67.63% with sensitivity and specificity 72.87% and 66.67% respectively. However, for patients with no prior diagnosis of CAD, only two features (sex and angiotensin-converting enzyme inhibitor/angiotensin receptor blocker use) were needed to achieve accuracy of 70.32% with sensitivity and specificity at 70.24%. CONCLUSIONS: This study shows that machine learning can predict the outcome of stress echocardiography based on only a few features: patient prior cardiac history, gender, and prescribed medication. Further research recruiting higher number of patients who underwent stress echocardiography could further improve the performance of the proposed algorithm with the potential of facilitating patient selection for early treatment/intervention avoiding unnecessary downstream testing.

Assessing the Privacy of mHealth Apps for Self-Tracking: Heuristic Evaluation Approach.

BACKGROUND: The recent proliferation of self-tracking technologies has allowed individuals to generate significant quantities of data about their lifestyle. These data can be used to support health interventions and monitor outcomes. However, these data are often stored and processed by vendors who have commercial motivations, and thus, they may not be treated with the sensitivity with which other medical data are treated. As sensors and apps that enable self-tracking continue to become more sophisticated, the privacy implications become more severe in turn. However, methods for systematically identifying privacy issues in such apps are currently lacking. OBJECTIVE: The objective of our study was to understand how current mass-market apps perform with respect to privacy. We did this by introducing a set of heuristics for evaluating privacy characteristics of self-tracking services. METHODS: Using our heuristics, we conducted an analysis of 64 popular self-tracking services to determine the extent to which the services satisfy various dimensions of privacy. We then used descriptive statistics and statistical models to explore whether any particular categories of an app perform better than others in terms of privacy. RESULTS: We found that the majority of services examined failed to provide users with full access to their own data, did not acquire sufficient consent for the use of the data, or inadequately extended controls over disclosures to third parties. Furthermore, the type of app, in terms of the category of data collected, was not a useful predictor of its privacy. However, we found that apps that collected health-related data (eg, exercise and weight) performed worse for privacy than those designed for other types of self-tracking. CONCLUSIONS: Our study draws attention to the poor performance of current self-tracking technologies in terms of privacy, motivating the need for standards that can ensure that future self-tracking apps are stronger with respect to upholding users' privacy. Our heuristic evaluation method supports the retrospective evaluation of privacy in self-tracking apps and can be used as a prescriptive framework to achieve privacy-by-design in future apps.