Using Formal Methods to Analyze Medical Error and Adverse Events Reports

Adverse events and medical errors occur with some frequency in large medical centers. The extent to which these are reported, analyzed and acted upon (i.e., become targets for intervention or additional controlled study) varies considerably. Events are under-reported for a large number of reasons, but one potential reason is the perception that the information is used only for punitive purposes and is rarely translated into operational improvements. Even when adverse events or incidents are reported, much of the data from these reports are underutilized in terms of generating knowledge and guiding safety interventions and system improvements. Full utilization is hampered by the fundamental properties of the data, itself. Primary sources of adverse event data are typically free narratives describing what happened from the singular perspective of the individual(s) involved. While this narrative format preserves the temporal sequence of events and provides important contextual information, it is relatively unstructured, making it very difficult to extract key elements of the reports and analyze them in a systematic or aggregate manner. Given that adverse event reports are the largest and potentially best institutional memory of complex, patient events, we should try to maximize our use of them to guide interventions and safety improvements.

In order to analyze these reports using formal methods (e.g., apply computational techniques such as statistical pattern analysis), it is necessary to transform the unstructured narratives into a more manageable format. As a first step, we are using a Description Logic to develop a conceptual representation for: patient safety adverse medical and surgical events medical and surgical complications medical risk management human error iatrogenic injury unsafe or hazardous processes of care.
This enables us to classify complex clinical phenomena as concepts (semantic types) and concept values and develop a common Semantic Reference Model. We then combine this with a series of matching algorithms to tag the primary source data. Concepts and their semantic types are grouped according to the clauses in the text in which they occur. All possible instances of semantic relations between concepts are generated, and this final group of concepts and relations is then compared to results derived from manual classification that has been performed on this corpus by a group of physicians, nurses and risk managers.

When we have been able to validate the representation scheme, we will move to the second major aspect of this work, applying latent class analysis to identify causal relationships among variables. This will enable us to develop predictive models of error and medical risk, and identify previously unrecognized features of clinical care that put patients at risk.