The problem with existing prediction models, the researchers argue, is practical rather than scientific. The study, published in Frontiers in Psychiatry, set out to build a model from information that is essentially always available at the bedside within minutes of admission. The three variables chosen were age, the Nutritional Risk Screening 2002 score (NRS-2002), and the Activity of Daily Living (ADL) dependency score.

Age serves as a non-modifiable biological baseline. NRS-2002 is a composite nutritional risk measure incorporating body mass index, recent weight loss, reduced dietary intake, and disease stress, which nursing staff can complete quickly. The ADL score, ranging from fully independent to fully dependent, captures pre-fracture functional status and frailty.

Crucially, all three are routinely recorded during nursing admission and require no laboratory work, cognitive evaluation, or additional testing. The team analyzed 387 patients aged 60 and over who underwent hip fracture surgery between January 2019 and August 2023, excluding those with pathological fractures or delirium already present at admission. Of these, 43 patients — 11.1 percent — developed postoperative delirium, ascertained through the Confusion Assessment Method documented in medical records.

The cohort was predominantly female (67.2 percent) with a mean age of 80.8 years, and patients who developed delirium were significantly older, more malnourished by NRS-2002, and more functionally dependent than those who did not. Statistically, all three variables emerged as independent predictors. Each additional year of age raised the odds of delirium by roughly 9.6 percent (odds ratio 1.096), each point on the NRS-2002 more than doubled the odds (OR 2.075), and each step of ADL dependency increased the odds by about 59 percent (OR 1.593).

The model achieved a cross-validated area under the receiver operating characteristic curve of 0.815, indicating moderate discrimination, with a calibration slope of 1.068 and a Brier score of 0.0848 — figures suggesting the predicted probabilities track observed outcomes reasonably well. Below a predicted probability of 0.115, the negative predictive value reached 97.9 percent: among 234 patients classified as low-risk, only five developed delirium, an observed rate of just 2.1 percent. Sensitivity at the cutoff was 88.4 percent, meaning the model caught nearly nine in ten eventual delirium cases.

The flip side is a modest positive predictive value of 24.8 percent — being flagged as high-risk does not mean delirium is likely, only that vigilance is warranted. NRS-2002 was missing in 33.6 percent of patients and ADL in 44.4 percent, largely because the hospital's comprehensive geriatric assessment program was rolled out progressively during the study period — ADL documentation was absent for all 2020 admissions but complete by 2023. The team used multiple imputation by chained equations with 20 imputations, and, notably, fitted all imputation models strictly within training folds of cross-validation to prevent data leakage, computing every reported performance metric on out-of-fold predictions.

When imputed values were deliberately shifted toward worse nutritional and functional status to simulate missing-not-at-random data, the age and ADL estimates remained essentially unchanged, but the NRS-2002 estimate attenuated from 2.075 to 1.204 — still directionally positive, but weaker. The authors therefore interpret the nutritional signal with explicit caution, noting that the true strength of the NRS-2002 association may lie toward the lower end of that range. A complete-case analysis with only 21 events was reported purely as a directional check, not as evidence of superior performance.

The practical takeaway is that this three-variable model is best understood as a bedside rule-out screening tool rather than a diagnostic instrument. The authors emphasize that single-center retrospective data with internal validation only cannot support clinical deployment: multicenter prospective external validation, incorporating systematic cognitive screening that this cohort lacked, is needed before the model earns a place in emergency departments.

Subject of Research: Prediction of postoperative delirium in elderly hip fracture patients using a bedside model based on age, NRS-2002, and ADL scores

Article Title: A simplified bedside model using age, NRS-2002, and ADL to predict postoperative delirium in elderly hip fracture patients: a retrospective cohort study

Article References: Chen, Q., Chen, S., Guan, L., Wu, Y., & Pei, Z. (2026). A simplified bedside model using age, NRS-2002, and ADL to predict postoperative delirium in elderly hip fracture patients: a retrospective cohort study. Frontiers in Psychiatry, 17, Article 1963389. https://doi.org/10.3389/fpsyt.2026.1963389

Image Credits: AI Generated

DOI: 10.3389/fpsyt.2026.1963389

Keywords: postoperative delirium, hip fracture, elderly patients, NRS-2002, ADL score, prediction model, nutritional risk, geriatric surgery, retrospective cohort study, delirium screening, cross-validation, emergency care