Predictive modeling of nontuberculous mycobacterial pulmonary disease epidemiology using German health claims data

Highlights

• Machine learning using historical claims data may predict previously undiagnosed NTM-PD.

• A random forest model with a risk threshold >99% performed best (AUC 0.847; total error 19.4%).

• Prevalence increased 5-fold to 19/100,000 for both coded and non-coded vs. coded cases alone.

• Correspondingly, incidence increased 9-fold to 15/100,000 population in 2016.

• A relevant number of previously unreported NTM-PD cases were identified with high probability.

Abstract

Objectives

Administrative claims data are prone to underestimate the burden of non-tuberculous mycobacterial pulmonary disease (NTM-PD).

Methods

We developed machine learning-based algorithms using historical claims data from cases with NTM-PD to predict patients with a high probability of having previously undiagnosed NTM-PD and to assess actual prevalence and incidence. Adults with incident NTM-PD were classified from a representative 5% sample of the German population covered by statutory health insurance during 2011–2016 by the International Classification of Diseases, 10th revision code A31.0. Pre-diagnosis characteristics (patient demographics, comorbidities, diagnostic and therapeutic procedures, and medications) were extracted and compared to that of a control group without NTM-PD to identify risk factors.

Results

Applying a random forest model (area under the curve 0.847; total error 19.4%) and a risk threshold of >99%, prevalence and incidence rates in 2016 increased 5-fold and 9-fold to 19 and 15 cases/100,000 population, respectively, for both coded and non-coded vs. coded cases alone.

Conclusions

The use of a machine learning-based algorithm applied to German statutory health insurance claims data predicted a considerable number of previously unreported NTM-PD cases with high probabilty.