On the prediction of cognitive impairment trajectories from anatomical MRI in ADNI: preliminary results

We predict yearly differences in MMSE and CDR-SOB using data from the ADNI cohort ranging from five years before to five years after the imaging session. We show that models that use embeddings from a deep-learning model trained to predict brain-age or models that use bilateral hippocampi, amygdalae, and accumbens perform approximately the same as models that use baseline cognitive scores as inputs. This has potential ramifications for both clinical machine learning applications and the neurobiology of cognitive decline. In future work, we will finetune the deep-learning model and substantially increase the sample size.