New AI Frameworks for Real-World Clinical Prediction

Abstract

In recent years, artificial intelligence (AI) has seen a string of successes for predicting clinical outcomes from eye disease to skin cancer to mortality. Clinical AI methods usually assume access to a large patient data matrix where rows are patients and columns are clinical variables, as well as a vector of labels representing each patient’s outcomes. However, access to these data matrices is limited in many important clinical prediction tasks, making standard AI methods difficult to apply. We present methods to advance three broad areas of data-constrained clinical prediction. First, when AI models are deployed in fields like emergency medicine, providers may lack the time to gather all variables in the data matrix for input to an AI model when it is deployed. We addressed this problem with CoAI (Cost-Aware AI), which can automatically select highly predictive variables to make any AI model work within a desired clinical time constraint. Second, because most AI models require extremely large numbers of patient samples, it can be hard to train them when very little labeled data is available. We addressed this problem with sparse attribution priors, which enable a form of sparsity regularization in neural networks similar to sparse linear models and significantly improve performance with limited training data. Finally, when studying rare or emerging diseases, scientists may find that no outcome data is available at all to train predictive models. We addressed this problem with decoupled regression, an approach for synthesizing existing associations reported in the literature into full multivariate predictive models without ever using labeled training data. All three of these methods are general AI frameworks that are applicable to any predictive task within or outside of clinical medicine. We hope they will help to bring the impressive predictive power of AI methods into new clinical fields in which their application would have otherwise been impossible.