PREDICTING HEALTHCARE COSTS OF DIABETES USING MACHINE LEARNING MODELS

Abstract

Objectives To design a predictive model of the in-hospital pharmaceutical expenditure of type II diabetes mellitus, derived from the risk index determined by the associated comorbidities, in a high-complexity hospital in Colombia.

Methods Data are collected from patients hospitalized in the years 2017 and 2018 at one University Hospital in Bogotá. A cross-sectional descriptive study of costs by typology is performed, then the total risk index derived from the number and severity of comorbidities is calculated, and its results from laboratory indicators. Factorial analysis establishes the prevalence of co-occurrence of chronic disease and associated morbidities. Based on these data, we use different Machine Learning models (such as logistic regression and supervised learning/random forest) that allows to identify the main variables related with the prediction of the pharmaceutical cost or expenditure.

Results Diabetic patients account for approximately 25% and have an average stay of 5 days. Differences in the number of comorbidities and variability of clinical severity were observed. Laboratory results were available for indicators such as: glycosylated haemoglobin, creatinine and uric acid, as well as lipid profile and 24-hour mean glycemia. The total estimated expense for diabetic patients, with and without associated comorbidities, was US$ 2,204,860 (30.1%) by 2017 and US$ 2,192,365 (27.2%) by 2018. The main co-morbidities associated with diabetes are: Urinary Tract Infection, Hypertension, Heart Failure and Coronary Ischemic Disease. There is an observable relationship in the increasing of costs with the number of associated comorbidities.

Conclusions Based on the aforementioned data, simulations are being carried out for the development of the model, which is expected to be able to predict the variations that emerge from the existing relationship between the risk index and pharmaceutical expenditure.