Analogies are one of the core components of human reasoning: when solving a problem, we try to find solutions by looking for similar situations from our experience or elsewhere. Medical diagnosis is not different, physicians often reach a diagnosis or treatment decision by finding patient cases they had experienced before. Unfortunately, computationally determining patient similarity is a highly complex task due to the multi-modal and multi-sourced nature of clinical information. This is especially visible in the field of rare diseases, where more than 7000 rare diseases afflict individually less than 2000 people. Thus, in the case of rare diseases, it is highly unlikely that a treating physician may experience several times a similar patient case.
In this three-year project, we aim to generate innovation and impact in practice through a three-layer development pipeline for diagnosing patients with rare disease. We propose to develop novel proof-of-concept algorithms/approaches for (i) classifying patients with ORPHA codes and (ii) computing/inferring patient similarity functions to support physicians in their analogical tasks of patient diagnosing as well as, possibly, deciding on treatment, and monitoring of disease course. We will address the first goal by developing an ORPHA-coder helper, i.e., ontology-based classification method that, based on patient information, will help clinicians determine the correct ORPHA code and standardize coding for patients with rare diseases. We will address the second goal by using existing patient information to learn embedding functions (i.e., to “project” patient information into a metric space, where patients with similar properties lie close to each other) to find similar patients. We will develop a rich catalog of data-driven methods for computing patient similarity. We will also exploit active learning techniques to improve the computational methods for patient similarity. Finally, we will develop an exploration interface for electronic health records that allows physicians to inspect the available data on a patient, assess similarities, and form cohorts.
In terms of public health impact, our project will help hospital clinics to attain their classification as a reference center recognized by national and international organizations by using our ORPHA-classifications. We also plan to release the prototype of the EHR exploratory interface to enhance medical practice by enabling planning and selecting cohorts of patients and strengthening the development of medical discovery tools. Furthermore, by providing AI-driven support in diagnostic procedures, our project should lower the psychological burden of undiagnosed patients and shorten their diagnostic odyssey.