Predicting 6-month decisional regret among family surrogate decision-makers of ICU patients: development and internal...

To develop and internally validate prediction models estimating the individual probability of high decisional regret at 6 months among family surrogate decision-makers (FSDMs) of intensive care unit (ICU) patients, and to compare regression-based and machine-learning approaches with respect to discrimination, calibration and clinical utility.

A single-center prospective cohort study for the development and internal validation of prediction models.

This single-center observational prediction-model study was conducted in five ICUs of a tertiary university-affiliated hospital in Hangzhou, China, from October 2023 to October 2024. We enrolled 336 family surrogate decision-makers. Sociodemographic characteristics, health literacy, eHealth literacy, anxiety, depression, trust in physicians, and family functioning were collected within 3 days after ICU discharge. Decisional regret was assessed 6 months after hospital discharge using the Decision. Regret Scale and analyzed as a binary outcome, with a transformed score of ≥25 defining high decisional regret. Candidate predictors were selected using univariate analysis and least absolute shrinkage and selection operator regression. Logistic regression, random forest, support vector machine, and extreme gradient boosting models were developed and internally validated using a 70:30 train-test split. Model performance was assessed by discrimination, calibration, and decision curve analysis. SHAP analysis was used as an exploratory interpretability method.

Among 336 family surrogate decision-makers, 165 (49.11%) had high decisional regret and 171 (50.89%) had low decisional regret. Univariate analysis identified nine variables associated with decisional regret, including health literacy, trust in physicians, patient survival status, relationship to the patient, household income, anxiety, depression, and family functioning (all p < 0.05). In the test set, logistic regression showed the strongest overall performance (AUC 0.875, 95% CI 0.808-0.942; accuracy 0.752; sensitivity 0.702; specificity 0.796; F1 score 0.725). It also showed the most stable calibration and clinical net benefit. The final model comprised five predictors: eHealth literacy, trust in physicians, average monthly household income, anxiety and depression. In a secondary explanatory analysis, each contributed independently to the predicted probability (all p < 0.05); these estimates are reported to aid interpretation of the model and are not intended as causal effect estimates. Exploratory SHAP analysis identified trust in physicians, anxiety, and eHealth literacy as prominent contributors to model predictions.

Decisional regret was common among ICU family surrogate decision-makers. Trust in physicians, eHealth literacy, household income, anxiety, and depression were key predictors. Among the candidate models, logistic regression provided the best balance of discrimination, calibration, and clinical utility, suggesting that a parsimonious and interpretable model may support early risk screening in ICU family-centered care.