Track: Machine Learning

Abstract

At an emergency room (ER) supplies, personnel and infrastructure must be ready and available to provide immediate care whenever is needed. Since these are limited resources, a good estimation of the near future patient volume might help when planning in advance. This research approaches the forecast of the patient volume at an ER by means of applying machine learning algorithms based on Support Vector Machines (SVM).

The research was carried out following a 4-phase methodology: analysis, design, development, and validation. During the analysis, over 50,000 ER records preprocessed and sorted as daily time series. During the design, datasets, forecast horizons, SVM-based forecasting algorithms, and performance metrics were defined. The development was carried out entirely using the data processing software WEKA. Lastly, the validation was completed using held out data so that forecasts were compared to actual values. Forecasts based on linear regressions (LR) were used as baseline.

The first approach consisted in using the cumulative data to forecast one month ahead, while the second approach considered only the last available month to forecast one month ahead. Besides MAE, MSE, RMSE, and MAPE, the number of months where the forecast underestimated the actual value was also used to compare approaches.

The results revealed that, when using cumulative time series, LR-based forecast underestimated the patient volume 8 out of 10 months. SVM-based forecast underestimated only 2 out 10 months. When forecasting with the last available month, the underestimation was 6 out of 10 and 3 out of 10 respectively. 

In conclusion, in some cases SVM-based forecasts can outperform LR-based forecasts. The availability of accurate patient volume estimations might serve well when planning and allocating resources.