Multimodal Machine Learning for Evidence Base Medicine

Practicing evidence-based medicine relies on identifying related evidences like medical literature or clinical trial which include credible information related to the clinical case where such evidence typically contains imaging and non-imaging data. However, classical machine learning techniques rely on using monomodal learning that process only a single modality (e.g. text, images or videos). Although this type of learning has provided extraordinary results in clinical science including diagnosis and prognosis, it has limited usage capabilities related to precision and evidence-based medicine. Multimodal learning, however, utilizes techniques to process and find relationships between different types of data modalities to comprehend the case description as some clinical cues exist only in certain modalities. Actually in clinical practice, clinicians use imaging modalities like computed tomography (CT) scans or X-ray images beside non-image modalities like electroencephalogram (EEG) data or tabular data used in the electronic healthcare records to describe clinical cases. In this article we are presenting a multimodal machine framework that can be used for evidence-based practice using fine tuned MedCLIP transformer model to fetch relevant PubMed articles based on three different scenarios of usage where clinicians can provide multimodal queries in their search for relevant PubMed articles.