A framework focused on deployment for pulmonary disease classification multimodal deep learning

Abstract

Globally, pulmonary diseases are a major health burden, especially in settings where resources are constrained and access to specialized radiological expertise is limited. Most chest radiograph-based deep learning models rely only on imaging data, despite showing promising performance when it comes to their diagnostic capabilities. However, clinical decision-making in the real world integrates structured patient information with the aforementioned imaging data. Our study aims to compare current multimodal machine learning approaches that combine clinical data and imaging used in pulmonary disease classification and propose a deployable framework designed for clinical settings in resource-constrained environments. We conducted a review of recent literature around multimodal AI in pulmonary diseases, and we focused on fusion strategies (early-stage, late-stage, and hybrid), techniques for data integration, validation settings, as well as deployment considerations. We performed a comparative synthesis aiming to identify methodological patterns, translational gaps, and performance trends, and on this basis, as a conceptual blueprint to guide our subsequent model development, we propose a modular architecture integrating structured-data encoders with convolutional neural networks for imaging data, along with fusion mechanisms to handle incomplete modalities. We then formalize the fusion operators and present an algorithm for graceful degradation under missing clinical data; empirical validation of the proposed architecture is deferred to a future work. This review indicates that multimodal approaches consistently outperform unimodal imaging models, especially in early-stage or complex cases, with gains in performance reported across many pulmonary conditions. However, the review also reveals several limitations, whether it's the lack of standardized fusion evaluation, insufficiencies in external validation, inadequacies in the handling of missing clinical variables, or limited attention to real-world clinical integration. These obstacles expose the need for system design that is deployment-aware more so than solely performance-driven optimization. Through this work, we aim to contribute a structured synthesis of multimodal pulmonary AI and outline an interpretable, resource-conscious framework intended for integration into healthcare workflows, which we put forward as the design basis for a model to be developed and evaluated in future work. By aligning the model design philosophy with the realities of clinical workflows, this approach should support equitable access to AI-assisted diagnostics and help advance the application of AI for healthcare improvement and social good.