Development and validation of a simplified two-point sampling strategy for predicting in-hospital mortality in elderly ICU sepsis patients: a multicenter retrospective cohort study

Abstract Background Sepsis is a leading cause of death in the intensive care unit (ICU), with significantly higher rates in elderly patients due to immunosenescence and greater comorbidity burden. Existing machine learning models for ICU mortality prediction predominantly rely on fixed-time-window feature extraction without evaluating how temporal sampling density influences performance. This study aimed to compare the effects of different temporal sampling density strategies on in-hospital mortality prediction in elderly sepsis patients and identify the critical sampling density at which information saturates. Methods This retrospective cohort study used the Medical Information Mart for Intensive Care IV database as the development set and the Electronic ICU Collaborative Research Database as the external validation set. Patients aged 65 years or older who met the Sepsis-3 criteria and had an ICU stay of at least 24 hours were included. A fixed-architecture long short-term memory network was used to compare six temporal sampling strategies (1-point, 2-point, 3-point, 4-point, 6-point, and 24-point) under a unified feature set and training strategy. Model performance was evaluated in terms of the area under the receiver operating characteristic curve, the area under the curve, the F1 score, and the Brier score. Interpretability analysis was performed to identify the core drivers of the optimal model. Results Of the 546028 screened records, 20163 elderly sepsis patients were included. The 2-point group (admission and 24-hour) achieved optimal discrimination (area under the receiver operating characteristic curve 0.814, 95% confidence interval 0.809–0.818), with the largest improvement over the single-point group (0.122; P < 0.001). The difference versus the 24-point hourly group was not significant (P = 0.961), indicating predictive equivalence. Adding intermediate sampling points beyond 2 points yielded no additional gain. Sensitivity analysis confirmed that the 24-hour endpoint contains richer prognostic information than earlier timepoints do. External validation (n = 11039) confirmed these findings. The 24-hour Glasgow Coma Scale score, blood urea nitrogen, and lactate level were the core predictive indicators. Conclusions In elderly ICU sepsis patients, a simplified two-point sampling strategy (admission baseline and 24-hour endpoint) achieves predictive equivalence to hourly dense monitoring. Intermediate sampling points provide no additional gain, suggesting that simplified sampling is sufficient under ICU data sparsity constraints.