Predicting Pain Recovery: A Multimodality Machine-learning Approach using Harmonized Electronic Health Record Data

We will develop robust, ethical, and fair predictive models using harmonized electronic health record data including brain imaging from two institutions (Mayo Clinic and the VA Health Care System) from patients with headache pain diagnoses (migraine, post-traumatic headache, post-stroke headache) of various ages, genders, medical histories, ethnic backgrounds and sociodemographic statuses to identify generalizable factors that contribute to pain persistence and to develop prognostic models that accurately predict pain recovery using multimodal fused data. Model development will be first conducted on each dataset separately due to the ‘uniqueness’ of data available within each dataset. Subsequently, will we develop a novel contrastive learning model to fuse common and unique modalities from both databases to create a harmonized, clinically useful model for predicting pain recovery that can harness data incompleteness and is flexible to real-world settings where not all patients have all data available. Supervised learning and a vision-language model will be used to extract unique features and to connect data of different types across modalities. Testing and validation accuracies of the harmonized model will be compared to the prediction accuracies of the individual models. Code will be released using Huggingface model zoo and GitHub. Throughout our milestones, we will build in opportunities for stakeholder considerations of multimodal AI for chronic pain prediction to ensure that potential uses of the model are aligned with the needs and values of healthcare providers (the end users) and patients (whose data informs the AI) ensuring sensitivity of the impact of multimodal AI on healthcare delivery.