An Adaptive Fuzzy Modeling of Visual Attention in Real-world Interaction with Health Information System

Abstract

Bottom-up and top-down processes are the two mechanisms of visual attention allocation, which allow people to efficiently spot task-relevant stimuli from cluttered and noisy environments, while staying alert to abnormalities within the visual field of view. This paper presents a preliminary study of the physicians’ real-life interaction with Information Communication Technology (ICT) in their own offices, along with extensively analyzing one case of an hour-long interaction of a physician, in which one performs a daily routine of reviewing patient electronic health records (EHRs) and writing diagnostic notes to the system interface. The physician interactions were captured in a time series data by recording display screen, keystrokes and mouse movements, also by simultaneously tracking eye movements. Then, a fuzzy-based model that can distinguish bottom-up and top-down processes were defined by using statistical random variables in terms of eye-movement patterns. The shift between those two attentional processes was detected by tracking the parametric changes of gaze behaviors as input: significant shift of fixation, sustained gazing, and fixation trajectory over time. Based on those gaze metrics, a random variable was assigned to the discrete probability of low (0), medium (0.5), or high (1.0), for a quantified fuzzy output, which was further machine-learned into an Adaptive Neuro-Fuzzy Inference System (ANFIS) model in order to judge how a physician is likely to be dominated by a bottom-up or top-down processes in performing a task at that instance in time. On training the ANFIS model with three different types of fuzzy membership functions (Gaussian, triangular and trapezoidal), the model performed best with the Gaussian function (after 100 iterations, the predicted root mean-square error (RMSE) converged at 0.07%, yielding a smooth linear curve). For a proof-of concept, the model was implemented by using one physician’s gaze behaviors, of which the average, machine-learned fuzzy output probability indicated that the physician was veering toward bottom-up visual attention. This individualized, task-specific pattern of visual attention has implications for the designs of intelligent interface in ICT. Our ANFIS model can scale up to different physicians and tasks to predict the likelihood of bottom-up or top-down information processing based on real-world gaze behaviors.