22 Sep, 15 | by rradecki
This month, Kreindler et al report a review of patient-level predictors of protracted Emergency Department length of stay. These sorts of reports are critical, not solely due to the relative importance of the topic, but also because of the gaps revealed in the current literature. To wit, despite reviewing 30 investigations of Emergency Department length of stay, a partial conclusion of these authors is the available information is insufficient to facilitate its use in service planning.
The value of such predictive tools would be profound. If, based on information readily available during triage, the LOS and resource utilization of a patient might be predicted, the effect spans the entire hospital. Nursing staff needs can be informed for the downstream wards and locations for general and specialty care, not simply for ongoing ED care. The ongoing presence of technical resources might be predicted, such as radiology or laboratory, and staffing maintained in anticipation of future orders. Finally, a prediction of increased LOS may influence models predictive of ED crowding, and prompt additional resources targeted at traditional bottlenecks in care. Clearly, predicting LOS results in a cascade of value.
However, as these authors note, it is simply too challenging to distill simple predictors from these data. Predictors of long LOS include eventual admission, patients with mid-acuity levels, older adults, and certain socioeconomic factors. However, some of these features are colinear with others, and some – like eventual admission – are frequently evident only at the conclusion of an ED stay. Increased intensity of diagnostic testing also predicts increased LOS, which brings to mind another confounding variable: the physicians themselves. Every ED has a range of clinicians, with variable diagnostic skills and risk tolerance, and identical patients with the same complaints may be assessed as immediately safe for discharge by some clinicians, while other clinicians might perform resource-intensive evaluations. This almost certainly ties into the general observation of presenting patient complaint only inconsistently associated with LOS.
What are the ramifications, then, for pursuing models for ED LOS? Essentially, there’s no simple shortcut. And, this is clearly clinically reasonable – as physicians we intrinsically know the vast constellation of patient factors relating to our decision-making process. It does not make sense to suggest a model of reasonable accuracy can be built from such broad brush strokes. Help, however, may yet be on the way from the realm of Clinical Informatics. As EHR data proliferates, the “undifferentiated” ED patient is gradually becoming a thing of the past. No longer would such a model be restricted to the basic triage information, but, perhaps include coded features of the medical history heretofore unavailable to prior models. Alternative analytic approaches, such non-linear, cluster-based techniques – those which best handle “big data” – may also be enabled by access to vastly more robust substrate.
The need for further study is apparent – but it is clear these previously pursued analyses should not be fruitlessly re-replicated. A new approach is necessary.