Quantifying Trust in Wearable Medical Devices

Abstract

Advances in sensor and digital communication technologies have revolutionized the capabilities of wearable medical device (WMD) to monitor patients’ health remotely, raising growing concerns about trust in these devices. There is a need to quantify trust in WMD for their continued acceptance and adoption by different users. Quantifying trust in WMD poses two significant challenges due to their subjective and stochastic nature. The first challenge is identifying the factors that influence trust in WMD, and the second is developing a formal framework for precise quantification of trust while taking into account the uncertainty and variability of trust factors. This thesis proposes a methodology to quantify trust in WMD, addressing these challenges.

In this thesis, first, we devise a method to empirically validate dominant factors that influence the trustworthiness of WMD from the perspective of device users. We identified the users’ awareness of trust factors reported in the literature and additional user concerns influencing their trust. These factors are stepping stones for defining the specifications and quantification of trust in WMD.

Second, we develop a probabilistic graph using Bayesian network to quantify trust in WMD. Using the Bayesian network, the stochastic nature of trust is viewed in terms of probabilities as subjective degrees of belief by a set of random variables in the domain. We define each random variable in the network by the trust factors that are identified from the literature and validated by our empirical study. We construct the trust structure as an acyclic-directed graph to represent the relationship between the variables compactly and transparently. We set the inter-node relationships, using the goal refinement technique, by refining a high-level goal of trustworthiness to lower-level goals that can be objectively implemented as measurable factors.

Third, to learn and estimate the parameters of the Bayesian network, we need access to the probabilities of all nodes, as assuming a uniform Gaussian distribution or using values based on expert opinions may not fully represent the complexities of the factors influencing trust. We propose a data-driven approach to generate priors and estimate Bayesian parameters, in which we use data collected from WMD for all the measurable factors (nodes) to generate priors. We use non-functional requirement engineering techniques to quantify the impacts between the node relationships in the Bayesian network. We design propagation rules to aggregate the quantified relationships within the nodes of the network. This approach facilitates the computation of conditional probability distributions and enables query-based inference on any node, including the high-level trust node, based on the given evidence.

The results of this thesis are evaluated through several experimental validations. The factors influencing trust in WMD are empirically validated by an extensive survey of 187 potential users. The learnability, and generalizability of the proposed trust network are validated with a real dataset collected from three users of WMD in two conditions, performing predefined activities and performing regular daily activities. To extend the variability of conditions, we generated an extensive and representative synthetic dataset and validated the trust network accordingly. Finally, to test the practicality of our approach, we implemented a user-configurable, parameterized prototype that allows users of WMD to construct a customizable trust network and effectively compare the trustworthiness of different devices. The prototype enables the healthcare industry to adapt and adopt this method to evaluate the trustworthiness of WMD for their own specific use cases.