Identification of Misdiagnosis Factors in Allergic Bronchopulmonary Mycosis Using Explainable Machine Learning

Background: Allergic bronchopulmonary aspergillosis/mycosis (ABPA/ABPM) is frequently misdiagnosed as pulmonary tuberculosis due to overlapping clinical and radiological features. Methods: In a retrospective cohort of 89 multidisciplinary team (MDT)–confirmed ABPA/ABPM patients, we investigated determinants of misdiagnosis and disentangled causal drivers from spurious associations. Machine learning models with SHAP interpretation were utilized alongside a Double Machine Learning framework to estimate the average treatment effects of five key features while adjusting for major confounders. Results: Immunological markers were identified as the dominant contributors to misdiagnosis. Total serum IgE showed the strongest protective causal effect against misdiagnosis, followed by Aspergillus fumigatus –specific IgE. Bronchiectasis demonstrated a modest protective effect. These findings were robust across covariate balance, overlap, and placebo analyses. Conclusion: Our results indicate that ABPA/ABPM misdiagnosis is driven primarily by underrecognition of immunological features rather than imaging findings, underscoring the importance of systematic immunologic assessment to reduce diagnostic delay and unnecessary anti-tuberculosis treatment, particularly in tuberculosis-endemic settings. Infographic on ABPA/M diagnosis: data, methods, clinical impact via machine learning.The infographic is divided into three sections: Data and Processing, Methodology and Results and Clinical Implications. The Data and Processing section describes the patient groups, with 89 patients diagnosed with ABPA/M by MDT, including 55 correctly diagnosed and 34 misdiagnosed. It mentions data standardization using z-standardization and feature selection through LASSO with cross-validation. A map of China is shown, highlighting the lack of diagnostic criteria for ABPA/M and its frequent misdiagnosis as tuberculosis. The Methodology section outlines the use of LASSO regression with cross-validation, model selection involving six machine learning models and SHAP analysis for model interpretation. It includes causal inference using a Double Machine Learning framework and Average Treatment Effect estimation. The Results and Clinical Implications section highlights feature selection results, noting the best performance by the Glmnet model with an AUC of 0.808. It identifies slgE and tIgE as key factors, with low levels leading to misdiagnosis and confirms their protective effect. Clinical implications include identifying patients at high risk of misdiagnosis, suggesting combined immunological and radiological assessments and reducing inappropriate treatment. Keywords: allergic bronchopulmonary aspergillosis, allergic bronchopulmonary mycosis, machine learning, SHAP, diagnosis