A Systematic Review on Application of Multimodal Learning and Explainable AI in Tuberculosis Detection

Physicians rely on various data sources when diagnosing Tuberculosis (TB). This includes the patient’s historical data, demographic data, clinical laboratory results, and imaging data. Traditionally, the application of machine learning and deep learning in detecting TB has focused more on using single modes of data. This constrains the capabilities of the artificial intelligence (AI) techniques to replicate the clinical practice of incorporating multiple sources of information in decision-making. Recent advancements in deep learning and machine learning have enabled the integration of multimodal data which has led to the development of applications that more accurately reflect the clinician’s approach. However, the operations of deep learning techniques are still blackbox in nature, which makes it hard to understand their internal work mechanisms. As a result, it is necessary to incorporate explainable AI techniques to assist AI model users understand how the models make decisions. In this paper, we carried out a systematic review of two areas: First, we reviewed recent studies on the application of multimodal learning in TB detection. Here we have provided a summary of the public datasets used in the studies, data modalities used, the fusion techniques, and finally identified AI techniques that can be used with multimodal data. Then we looked at papers that used explainable AI techniques in TB diagnosis and prognosis. This study followed PRISMA guidelines to ensure replicability and accurate reporting of the main findings of the reviewed studies. To stay up-to-date with the state of the art, we specifically examined papers published between 2019 and June 2024. We reviewed thirty-one journal and conference papers we found using Web of Science, Scopus and Pubmed databases. The review indicated that models trained on multiple data modalities outperformed those trained on single data modalities. This is due to the additional information extracted from each data modality. Therefore, multimodal learning can improve clinical decision-making and TB diagnostic precision, but faces challenges like insufficient datasets and interpretability issues in model prediction processes.