[ad_1]
Discussion
Using a database of 230 000 ED attendances spanning 61 months at Meir Medical Center, Israel, we have created a regression model that is capable of predicting total arrival numbers for the concurrent 24-hour period. The model uses historical, environmental and calendric factors and historical arrival numbers as input. It is most accurate when provided with arrivals data for the hours prior to the prediction being made, and the MAPE of the model is 6.85%.
Having postulated that historical physician performance can predict admissions-related decisions, we have constructed several measures of team performance which are intended to correlate with these decisions. We find that physicians’ historical admission rate, rate of zero-day admissions and throughput in terms of patients seen per hour, are highly predictive. However, these measures are strongly collinear and in practice only the historical admission rate acts as an independent predictive variable. We find that the rate of ED returnees has no predictive value for admissions-related decisions.
Using the historical admission rate for the individuals working that shift, together with historical admissions and arrivals data, and environmental and calendric factors, we have constructed a regression model capable of predicting same-day admission numbers. The MAPE of the model is 10.6% and the inclusion of physician performance variables provides a clear boost to performance.
Using similar techniques, a mathematical test was developed that is able to provide a same-day prediction for when numbers of admissions rise above that with which the inpatient system is able to safely cope. When applied to those days at the highest risk of such a crisis, the sensitivity of the test is around 0.75 for detecting an impending crisis, for a false-positive rate of 0.25. Such advanced warning would enable same-day interventions to mitigate the effect of such a crisis.
In relation to the existing literature, the performance of our model is at least as good as those previously published.12 13 The use of team performance metrics to augment admissions modelling is new in the medical literature, and results in a MAPE that already exceeds the best previously published.15–19 The application of our model to predict crises in admissions, and the efficacy of attempts to mitigate their consequences, is new. The application of these techniques to medical admissions data is also new in the context of Israeli healthcare.
The main limitation of the present study is that the formulated models are directly applicable only to our own institution, although similar models could be derived easily for any other ED given the requisite data. Although our principal objective is the modelling of medical admissions, there is no reason why a very similar approach cannot be taken regarding surgical or orthopaedic admissions.
Given the large size of our primary dataset, it is unlikely that we can improve the performance of either our arrivals or admissions model by expanding its size. Conversely, the inclusion of hitherto unmodelled variables, such as local traffic patterns or individual ward occupancy, could well reduce residual variance and enhance predictivity. Technological upgrades involving newer machine learning techniques may offer some improved model performance.
Summarily, we have shown that consideration of physician performance is vital for models that predict ED admission numbers. A preliminary analysis shows that dramatic reductions in admissions for minimal outlay are possible by using such models as tools to optimise ED staffing. Our most immediate task is to demonstrate the practical application of our work to ED workforce planning in our institution.
[ad_2]
Source link
