Back

Fusion of magnet resonance imaging and electronic health records promotes the multimodal prediction of postoperative delirium

Giesa, N.; Dell'Orco, A.; Scheel, M.; Finke, C.; Balzer, F.; Spies, C. D.; Sekutowicz, M.

2025-02-21 health informatics
10.1101/2025.02.20.25322564 medRxiv
Show abstract

The role of individual brain morphometry, derived from clinical imaging, for the analysis as well as the early prediction of postoperative delirium (POD), as a severe complication after a surgery, is currently under-explored. We extensively analyzed patient trajectories of magnet resonance imaging (MRI) and electronic health records (EHRs) for two POD definitions covering 557 and 201 patients. Age-adjusted correlations with linear mixed-effect models identified middle temporal and superior temporal cortical thickness as well as thalamus and brain stem volumes. We trained highly non-linear multi-layer perceptrons (MLPs) on EHRs, MRI measures, and the combination of both as multimodal fusions. MLP models achieved high performance metrics, as are under receiver operating characteristics (AUROC) values up to 86%, outperforming baselines. Multimodal fusion was especially beneficial for 645 less critically ill patients. MLP model weights for this cohort focused on cerebral atrophy measures of higher order cortical areas, such as the temporalpole -, superiofrontal gyrus-, and the insula. Our results pave the way for the far unrecognized potential of clinical MRI features for early multimodal predictions of POD.

Matching journals

The top 5 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.