Resource frequency prediction in healthcare: Machine learning approach

Determining the minimal amount of resources needed to ensure minimal number of bottlenecks in the patient flow not only promotes patient satisfaction but also provides financial benefits to hospitals. The increase of data gathering by healthcare facilities in the last years have brought new opportunities to apply machine learning techniques to tackle this problem. This work makes use of data gathered from the Oulu University Hospital in Finland between 2011 and 2014 to study the effectiveness of machine learning techniques to predict resources usage. This work investigates the problem of resource frequency prediction and compares the performance of Nearest Neighbours and Random Forest. The application of data clustering as a preprocessing step is also explored as a way to improve the prediction accuracy of resources whose behavior change over time. The results indicate that 1) highly frequented resources can be predicted with higher accuracy than the lowly frequented resources, 2) the Random Forest have similar performance to the Nearest Neighbours although Random Forest performs better, 3) clustering improves the performance of the Nearest Neighbours but not of Random Forest, and 4) if averages are used to determine the resource frequency then cluster averages yields higher accuracy than all data averages.