Multimodal machine learning-based model for differentiating nontuberculous mycobacteria from mycobacterium...

Area of Science:

• Medical Imaging
• Machine Learning
• Infectious Diseases

Background:

• Differentiating non-tuberculous mycobacteria (NTM) from Mycobacterium tuberculosis (MTB) is crucial for effective treatment.
• Accurate diagnosis can be challenging, necessitating advanced diagnostic tools.

Purpose of the Study:

• To develop and evaluate a multimodal machine learning approach for differentiating NTM from MTB.
• To compare the performance of this multimodal model against clinical-only, radiomics-only, and existing diagnostic methods.

Main Methods:

• Retrospective analysis of clinical data and CT images from 175 patients.
• Development of clinical, radiomics, and multimodal (clinical + radiomics) models using five machine learning algorithms.
• Selection of optimal algorithms and feature identification for each model.
• External validation using new patient data and comparison with radiologist performance and NGS detection.

Main Results:

• The optimal multimodal model, incorporating age, IL-6, and two radiomics features, achieved the highest performance metrics (AUC, accuracy, sensitivity, NPV).
• The multimodal model demonstrated superior accuracy (0.745) and sensitivity (0.900) in the external test dataset.
• Performance surpassed that of individual clinical or radiomics models, radiologists, NGS detection, and existing machine learning models.

Conclusions:

• This study presents the first multimodal model for differentiating NTM from MTB.
• The developed model shows significant potential to assist clinical decision-making for experienced radiologists.
• The multimodal approach offers improved diagnostic accuracy compared to current methods.