Ethical Considerations of Using Machine Learning for Decision Support in Occupational Health: An Example Involving Periodic Workers' Health Assessments.

Purpose Computer algorithms and Machine Learning (ML) will be integrated into clinical decision support within occupational health care. This will change the interaction between health care professionals and their clients, with unknown consequences. The aim of this study was to explore ethical considerations and potential consequences of using ML based decision support tools (DSTs) in the context of occupational health. Methods We conducted an ethical deliberation. This was supported by a narrative literature review of publications about ML and DSTs in occupational health and by an assessment of the potential impact of ML-DSTs according to frameworks from medical ethics and philosophy of technology. We introduce a hypothetical clinical scenario from a workers' health assessment to reflect on biomedical ethical principles: respect for autonomy, beneficence, non-maleficence and justice. Results Respect for autonomy is affected by uncertainty about what future consequences the worker is consenting to as a result of the fluctuating nature of ML-DSTs and validity evidence used to inform the worker. A beneficent advisory process is influenced because the three elements of evidence based practice are affected through use of a ML-DST. The principle of non-maleficence is challenged by the balance between group-level benefits and individual harm, the vulnerability of the worker in the occupational context, and the possibility of function creep. Justice might be empowered when the ML-DST is valid, but profiling and discrimination are potential risks. Conclusions Implications of ethical considerations have been described for the socially responsible design of ML-DSTs. Three recommendations were provided to minimize undesirable adverse effects of the development and implementation of ML-DSTs.

Inferring Human Activity Recognition with Ambient Sound on Wireless Sensor Nodes.

A wireless sensor network that consists of nodes with a sound sensor can be used to obtain context awareness in home environments. However, the limited processing power of wireless nodes offers a challenge when extracting features from the signal, and subsequently, classifying the source. Although multiple papers can be found on different methods of sound classification, none of these are aimed at limited hardware or take the efficiency of the algorithms into account. In this paper, we compare and evaluate several classification methods on a real sensor platform using different feature types and classifiers, in order to find an approach that results in a good classifier that can run on limited hardware. To be as realistic as possible, we trained our classifiers using sound waves from many different sources. We conclude that despite the fact that the classifiers are often of low quality due to the highly restricted hardware resources, sufficient performance can be achieved when (1) the window length for our classifiers is increased, and (2) if we apply a two-step approach that uses a refined classification after a global classification has been performed.

A survey on the feasibility of sound classification on wireless sensor nodes.

Wireless sensor networks are suitable to gain context awareness for indoor environments. As sound waves form a rich source of context information, equipping the nodes with microphones can be of great benefit. The algorithms to extract features from sound waves are often highly computationally intensive. This can be problematic as wireless nodes are usually restricted in resources. In order to be able to make a proper decision about which features to use, we survey how sound is used in the literature for global sound classification, age and gender classification, emotion recognition, person verification and identification and indoor and outdoor environmental sound classification. The results of the surveyed algorithms are compared with respect to accuracy and computational load. The accuracies are taken from the surveyed papers; the computational loads are determined by benchmarking the algorithms on an actual sensor node. We conclude that for indoor context awareness, the low-cost algorithms for feature extraction perform equally well as the more computationally-intensive variants. As the feature extraction still requires a large amount of processing time, we present four possible strategies to deal with this problem.