Making Machine Learning Models Clinically Useful

Recent advances in supervised machine learning have improved diagnostic accuracy and prediction of treatment outcomes, in some cases surpassing the performance of clinicians.¹ In supervised machine learning, a mathematical function is constructed via automated analysis of training data, which consists of input features (such as retinal images) and output labels (such as the grade of macular edema). With large training data sets and minimal human guidance, a computer learns to generalize from the information contained in the training data. The result is a mathematical function, a model, that can be used to map a new record to the corresponding diagnosis, such as an image to grade macular edema. Although machine learning–based models for classification or for predicting a future health state are being developed for diverse clinical applications, evidence is lacking that deployment of these models has improved care and patient outcomes.²