Page - 165 - in Applied Interdisciplinary Theory in Health Informatics - Knowledge Base for Practitioners

intervention. A total of 18 different theories across educational, psychological, organisational and diffusion of innovation perspectives were mentioned. Arguably, many of these resonate with elements of CT. For instance, Social Cognitive Theory also proposes that feedback processes drive behaviour; and the Theory of Planned Behaviour postulates that attitudes, social norms, self-efficacy, and controllability provide reference values for behaviour2. 2.2. Other uses of Control Theory in health informatics There are other HI areas that draw upon the negative feedback loop depicted in Figure 1 and thus have roots in CT. As with the A&F literature, explicit references to CT are rare – but we would nevertheless argue that there is a clear relationship. We provide three examples here. First, a large number of decision support systems for the management of long-term conditions such as diabetes and cardiovascular diseases have been developed that deploy a negative feedback loop for controlling important clinical parameters. For instance, Athena [14] is a clinical decision support for the management of hypertension that issues an alert to its clinician users whenever a patient’s latest blood pressure measurement is too high. Similarly, Pandit [15] is a web-based diabetes management system for patients that asks them to measure and enter their blood glucose level. Whenever a glucose value is outside the normoglycaemic range, the system responds by suggesting adjustments to the patient’s insulin dose. Similar mechanisms have been used in expert systems for critical care [16]. Second, many smartphone apps that assist in health behaviour change provide users with feedback on their achievements against pre-set goals, such as the number of steps taken, or time spent on physical activity per day. The feedback aims to incentivise users to increase their level of healthy behaviour when it is below target, and maintain it when it is on par. Third, CT plays a central role in supervised machine learning methods such as Hebb’s learning rule [17]; the Newton-Raphson algorithm [18]; gradient boosting [19]; and deep learning [20]. Essentially, each of these methods utilise the negative feedback loop to derive a model of an input-output function from training data. Initially, a default or random model is chosen that bears no relationship to the data, and that model is subsequently ‘trained’ to better fit the data. The feedback is always derived from discrepancies between observed outputs (in the data; typically called training labels) and predicted outputs (predicted by the model). At each iteration of the feedback loop the classifier will better approximate the input-output function that produced the data, and the discrepancies will disappear after which the process is terminated. 3. Explanation of success or failure of audit and feedback Despite being commonly applied as a healthcare quality improvement strategy, A&F interventions yield variable and often only marginal effects [8]. Moreover, over four decades of research in the field seems to have failed to enable A&F researchers to successfully enhance intervention designs and achieve larger effects consistently. It has 2 See also Chapter 4, “Assessing technology success and failure using Information Value Chain Theory”, where Information Value Chain theory has been applied to A&F interventions. W.T.GudeandN.Peek /ControlTheory toDesignandEvaluateAuditandFeedback Interventions 165