Incidence and prevalence of non-tuberculous mycobacterial lung disease: predictive modelling with German claims data

Objectives Nontuberculous mycobacteria (NTM) are a diverse group of bacteria other than the Mycobacterium tuberculosis complex that can cause progressive inflammatory lung damage, the NTM lung disease (NTM-LD). A previous study explored the annual prevalence of NTM-LD in Germany based on claims data, using ICD-10 coding for patient identification. However, lack of consistent use of ICD-10 code for NTM-LD may result in underestimation of true disease prevalence. Therefore, we aimed to estimate the epidemiological profile of NTM-LD in Germany with a predictive modelling approach that uses historical claims data of coded cases to predict likely cases with NTM-LD that have not received the corresponding ICD-10 code.

Methods Using 70% of a longitudinal claims dataset of approx. 4 mio. Germans within statutory health insurance between 2011 and 2016 we trained a classification algorithm to predict reported NTM-LD cases identified through ICD-10 code (A31.0). The remaining 30% of the dataset was used to evaluate the algorithm with regards to its predictive power. All individuals in the dataset from 2013 – 2016 without ICD-10=A31.0 code were then scored on probability of having NTM-LD.

Results The model had relatively high predictive power (AUC: 84.7%, error: 19.4%). The most important predictive variables were presence of chronic obstructive pulmonary disease, age, x-ray of chest/thorax, computed tomography, diagnosis of unspecified pneumonia or influenza, antibiotic treatment, and treatment with proton pump inhibitors. Applying the predictive model on the 4 mio. dataset and including only individuals with over 99% likelihood of NTM-LD, we estimated the prevalence and incidence rate in 2016 for both coded and uncoded individuals of 19 and 15 per 100,000, respectively, compared to 4 and 2 per 100,000 for the coded NTM-LD cases only. Compared to coded NTM-LD cases individuals without A31.0 code but high likelihood of NTM-LD had more frequently history of COPD and pneumonia diagnosis but less frequently of tuberculosis, higher rates of prescribed proton pump inhibitors and inhaled corticosteroids, and higher mean age.

Conclusions The results of predictive modelling suggest an almost four times higher estimated number of potentially unreported NTM-LD patients than the number of patients with ICD-10 coded diagnosis.