Clinical pathways have been widely adopted in the past years to standardize patient treatments [1], [2]. A number of clinical trials have shown that implementing clinical pathways usually results in higher quality and lower cost of clinical processes [3], [4], [5]. However, these studies also highlighted the discrepancy that often exists between the clinical pathways and the actual clinical practice in that clinical pathways are often designed for ideal conditions [6], [7], regarding, e.g., doctors’ experience or patients’ characteristics, which often do not hold for real-world contexts [8].

The pervasive use of health information systems in modern hospitals has enabled the tracking of the execution of clinical processes, which, in turn, has boosted the development of data-driven techniques to analyze process behaviors [9], [10], [11], [12], [13]. In particular, conformance checking techniques emerged as an effective method to assess the compliance of the execution of a clinical process to the corresponding clinical pathway. These techniques take as input a process model representing the clinical pathway and a so-called event log storing the recorded process executions, and allow to pinpoint discrepancies between each execution and the model [14], [15]. Although they were originally developed for analyzing business processes, a number of studies in the literature have shown the benefits of applying conformance checking techniques on clinical processes [16], [17], [18], [19], [20], [21], [22], [23].

Nevertheless, there are still challenges in applying conformance checking to clinical processes. Among them, a well-known issue is the uncertainty usually affecting the event log that tracks the clinical process, which often results in unreliable logging of activity timestamps [24], [25], [26]. In other words, it is known that in clinical practice the logging timestamp of the process activities can differ from the real execution timestamp. This is due to the fact that the logging of process activities in hospital information systems, such as the execution of a set of lab tests or the preparation of a patient for surgery, is still largely a manual task executed by the hospital personnel. As a result, the activity logging may happen at a different moment in time than the execution of the activity itself. In turn, this implies that the order in which activities are recorded in the system can differ from the real execution order. Applying standard conformance checking techniques in the presence of such logging uncertainty can lead to identifying too many “false positives”, i.e., deviations in the recorded log which actually do not correspond to relevant deviations in the real executions [27], which hampers the quality of the diagnostics and, hence, the trust of the users in the system.

We argue that a promising strategy to deal with this issue consists of explicitly taking into account the uncertainty related to the execution of the clinical process, using experts’ knowledge of the process to assess the magnitude of the deviations. By doing so, it is possible to rank the detected deviations, allowing the analyst to focus on the most interesting ones. The underlying hypothesis is that small deviations are less interesting, since they are likely due to logging errors or workaround practices than signaling real problems occurring within the process execution.

Elaborating upon the previous observation, in this work we introduce a novel conformance checking algorithm to detect violations of activity ordering constraints, using their logging time to assess the magnitude of such deviations. In doing so, we explicitly use experts’ knowledge to determine which time intervals can be used to distinguish between real process violations and violations that are likely due to logging issues. By leveraging principles of the fuzzy set theory, our approach is capable of quantifying the deviation degree of an incorrect timestamp, taking into account the user’s definition of an acceptable range for the timestamp uncertainty. To the best of our knowledge, this approach is the first to quantify the degree of the control-flow violations by utilizing fuzzy sets, and integrating them into the generation of compliance diagnostics. We ran the experiments by comparing the results of our proposed method to a classical conformance checking method for evaluating the effect of the fuzzy time window, and the differences on the conformance diagnostics. The approach brings two main contributions to conformance checking techniques for clinical deviation detection.

It provides a methodology to account for activity timestamp logging uncertainty and, consequently, to filter potential false-positive deviations.

It enables the generation of diagnostics that is personalized based on the analysts’ preferences, by explicitly incorporating their knowledge in the detection of control-flow violations with fuzzy logic.

The remainder of this paper is organized as follows. Section 2 introduces related work. Section 3 discusses background knowledge used throughout the manuscript. Section 4 introduces the proposed approach. Section 5 discusses results obtained by a case study of a cryo-ablation process in a Dutch hospital. Finally, Section 6 draws some conclusions and discusses future work.

Statement of significance