Identifying diseases in claims data using a machine learning approach – a case from Switzerland

In many health care systems, health data often include either diagnostic or cost data but not both. This poses a challenge for epidemiological or cost-of-illness studies. In this paper, we aim to identify diseases based on the utilization of outpatient services employing statistical learning models. We combine insurance claims data of the hospitalized population of a large Swiss health insurer from 2017 with diagnostic information in the national hospital inpatient registry from 2016 and 2017 at the patient level. We use random forests and boosting algorithms to predict the presence of 32 diseases based on outpatient health care utilization alone. The features include drug spending by four-digit ATC codes, spending by service provider, and spending by subchapter of national fee-for-service catalogues. We use the models to predict the prevalence of a disease in the non-hospitalized population for which no disease labels are available. Disease prediction worked best for diseases with specific treatment options (e.g., diabetes). Random forests achieved the best performance in 56% of all classification problems. For 25 diseases, drug utilization by ATC chapter was the most important feature in the prediction. Prevalence rates predicted for the full population were close to those reported previously for few diseases only, and showed large deviations for other diseases. Information on health care utilization from claims data may be used to predict the presence of diseases, but predictive performance varies across diseases, warranting further research on population-wide disease prevalence rates with incomplete information on diagnostic data.