Artificial Intelligence-Based Multimodal Detection of Pulmonary Tuberculosis Using Chest X-ray Images and Routine Laboratory Findings: A Machine Learning Study
TL;DR
Fusing radiological and hematological information through a multimodal deep learning pipeline substantially improves automated PTB screening performance relative to single-modality models and may serve as a low-cost, scalable triage tool in high-burden settings, pending prospective multi-center validation.
Abstract
Background: Pulmonary tuberculosis (PTB) remains one of the leading infectious causes of mortality worldwide, and delayed or missed diagnosis continues to hinder tuberculosis (TB) elimination programs, particularly in resource-constrained settings. Chest X-ray (CXR) interpretation and routine laboratory investigations are widely accessible, yet unimodal diagnostic approaches individually exhibit limited sensitivity and specificity. Objective: This study proposes and evaluates an artificial intelligence (AI)-based multimodal framework that fuses deep convolutional features extracted from CXR images with routine hematological and inflammatory laboratory markers to improve automated PTB detection. Methods: A retrospective dataset of 5,200 patients (2,540 TB-positive, 2,660 TB-negative), each comprising a digital CXR image and eight routine laboratory parameters (complete blood count indices, erythrocyte sedimentation rate, and C-reactive protein), was used. A ResNet-50 convolutional neural network (CNN) extracted 2048-dimensional image embeddings, while an Extreme Gradient Boosting (XGBoost) encoder generated 64-dimensional clinical embeddings from laboratory data. These were combined through an attention-weighted fusion layer feeding a fully connected classification head. Performance was benchmarked against four unimodal baselines (support vector machine, random forest, XGBoost, and CNN-only) using accuracy, sensitivity, specificity, precision, F1-score, and the area under the receiver operating characteristic curve (AUC). Results: The proposed multimodal fusion model achieved a test-set accuracy of 96.2%, sensitivity of 95.8%, specificity of 96.6%, F1-score of 96.0%, and AUC of 0.984, outperforming the best unimodal baseline (CNN-only, 89.6% accuracy, AUC = 0.932) by a statistically significant margin (p < 0.001, McNemar's test). Erythrocyte sedimentation rate and C-reactive protein emerged as the most influential laboratory predictors. Conclusion: Fusing radiological and hematological information through a multimodal deep learning pipeline substantially improves automated PTB screening performance relative to single-modality models and may serve as a low-cost, scalable triage tool in high-burden settings, pending prospective multi-center validation.