OBJECTIVES

: The aim of this study is to utilize machine learning based algorithms to provide an estimation of both the total and the ambulatory prescription costs per patient related to their health state.

METHODS

An available database was used. This contained information from the IT systems of the Valencia region on the assigned population of a health department. The variables include socio-demographic data (age and sex), Clinical Risk Group’s classification as the health state stratification, and the cost per patient of 2013.

RESULTS

: A double part algorithm is purposed to correctly address the problem. The first stage aims to predict if the patient will cause any cost to the Health Service. A logistic regression was calculated, achieving over 90% of accuracy with over 95% specificity. The second part of the algorithm acts only on the patients marked as expenders in the first half, and focus on the numerical prediction of the cost. In this step an artificial neural network was trained, with significantly different results, higher for the ambulatory prescription costs (64% R²). The combined result of the hole system ends up over 70% R² for the prescription cost, and close to 90% for the total costs.

CONCLUSIONS

: The identification of the expending patients was correctly achieved, been able to find them with a high confidence level. For the numerical predictions, the results may seem insufficient, but considering the system as a whole the global predictive value is higher than the offered in previous studies utilizing a statistical approach. This study proves that machine learning and artificial intelligence tools could improve the predictive models used in health management.