Predicting progression of Mycobacterium avium complex pulmonary disease with treatment using explainable artificial intelligence techniques.

OBJECTIVES

The incidence of nontuberculous mycobacterial pulmonary disease (NTM-PD) is rising; however, only certain patients experience disease progression requiring therapy.

METHODS

This retrospective cohort study enrolled 303 adults (age ≥ 20 years) who were not receiving treatment at enrollment, including patients with a prior treatment history of Mycobacterium avium complex pulmonary disease (MAC-PD). Clinical, laboratory, microbiologic, physiological, and radiographic variables were extracted. The outcome was clinical progression, defined as treatment initiation.

RESULTS

A point-wise linear neural network screened 63 features; the 22 most informative features yielded a parsimonious two-variable logistic model based on four routinely available inputs (age, treatment history, cavitation, and modified Reiff score). Overall, 124 patients experienced progression: these patients were younger, and frequently had cavitation, higher modified Reiff scores, and treatment history. The best model achieved an area under the curve (AUC) of 0.81. The Bayesian network indicated that age and treatment history influenced clinical progression through the radiographic disease burden.

CONCLUSION

For the first time, we demonstrated that the modified Reiff score is important for predicting clinical progression in MAC-PD using explainable AI models.