Answers are grounded in this paper and cite their sources.
Background The aim of this work was to develop and validate, on a held-out test cohort from the same center, a multimodal end-of-surgery model for postoperative delirium and test whether it directs a nursing prevention bundle to reduce incidence. Methods This two-phase retrospective study used stratified five-fold cross-validation and a temporally held-out test cohort from the same center (the most recent 310 of the 1,160 Phase 1 patients). Phase 1 developed DeliGuard-Net, combining a static encoder, a bidirectional LSTM over 18 nursing channels, and an imaging encoder pairing ConvNeXt-Tiny over five axial slices with the Evans Index and two atrophy grades. Of 2,418 records screened, 850 patients formed the development cohort and 310 the held-out test cohort; prediction was anchored at the end of surgery. Comparators included XGBoost (with and without the atrophy descriptors), a five-variable clinical model, and nursing judgment, excluding dementia patients. Phase 2 was an exploratory controlled before-after evaluation across three nursing units. On one unit, the frozen model's risk tier guided allocation of a nursing prevention bundle in 342 patients; the two other units continued usual care (486 concurrent controls). The Phase 1 patients from the same units served as the preintervention period (480 on the implementing unit and 680 on the control units), and the groups were compared by difference-in-differences. Results Delirium occurred in 146/850 development and 56/310 hold-out patients, at a median of 38 h. Hold-out discrimination was 0.82 (95% CI: 0.76–0.87) vs. 0.78 for XGBoost (0.80 when it also received the atrophy descriptors) and 0.66 for nursing judgment; the paired advantage was 0.16 (0.09–0.23) over nursing judgment but 0.04 (0.00–0.08) over XGBoost (0.02, −0.02 to 0.06, with descriptors). Adding modalities improved discrimination, but no architectural component separated from zero once all three modalities were retained. Delirium fell from 16.7% to 11.4% vs. unchanged controls (adjusted difference −4.9 points, 95% CI −14.1 to +4.3); no secondary outcome survived a multiplicity screen. Conclusion The multimodal model discriminated at 0.82 in the hold-out cohort but was not shown to be better than gradient boosting (difference 0.04, 95% CI: 0.00–0.08) despite far greater complexity; the exploratory implementation evaluation was underpowered and inconclusive.
No takes yet. Share an insight, caveat, or question.
Hu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: