Developing a nomogram for postoperative delirium in elderly patients with hip fractures

Abstract

Background: Postoperative delirium (POD) is a prevalent complication, particularly in elderly patients with hip fractures (HFs). It significantly affects recovery, length of hospital stay, healthcare costs, and long-term outcomes. Existing studies have investigated risk factors for POD, but most are limited by single-factor analyses or small sample sizes. This study systematically determines independent risk factors using large-scale data and machine learning techniques and develops a validated nomogram model to support early prediction and management of POD.

Aim: To investigate POD incidence in elderly patients with HF and the independent risk factors, according to which a nomogram prediction model was developed and validated.

Methods: This retrospective study included elderly patients with HF who were surgically treated in Dongying People's Hospital from April 2018 to April 2022. The endpoint event includes POD. They were categorized into the modeling and validation cohorts in a 7:3 ratio by randomization. Both cohorts were further classified into the delirium and normal (non-delirium) groups according to the presence or absence of the endpoint event. The incidence of POD was calculated, and logistic multivariate analysis was conducted to determine the independent risk factors. The calibration curve and the Hosmer-Lemeshow test as well as the net benefit threshold probability interval by the decision curve were utilized to statistically validate the accuracy of the nomogram prediction model, developed according to each factor's influence intensity.

Results: This study included 532 elderly patients with HF, with an overall POD incidence of 14.85%. The comparison of baseline data with perioperative indicators revealed statistical differences in age (P < 0.001), number of comorbidities (P = 0.042), American Society of Anesthesiologists grading (P = 0.004), preoperative red blood cell (RBC) count (P < 0.001), preoperative albumin (P < 0.001), preoperative hemoglobin (P < 0.001), preoperative platelet count (P < 0.001), intraoperative blood loss (P < 0.001), RBC transfusion of ≥ 2 units (P = 0.001), and postoperative intensive care unit care (P < 0.001) between the delirium and non-delirium groups. The participants were randomized to a training group (n = 372) and a validation group (n = 160). A score-risk nomogram prediction model was developed after screening key POD features using Lasso regression, support vector machine, and the random forest method. The nomogram showed excellent discriminatory capacity with area under the curve of 0.833 [95% confidence interval (CI) interval: 0.774-0.888] in the training group and 0.850 (95%CI: 0.718-0.982) in the validation group. Calibration curves demonstrated good agreement between predicted and actual probabilities, and decision curve analysis confirmed clinical net benefits within risk thresholds of 0%-30% and 0%-36%, respectively. The model has strong accuracy and clinical utility for predicting the risk of POD.

Conclusion: This study reveals cognitive impairment history, American Society of Anesthesiologists grade of > 2, RBC transfusion of ≥ 2 units, postoperative intensive care unit care, and preoperative hemoglobin level as independent risk factors for POD in elderly patients with HF. The developed nomogram model demonstrates excellent accuracy and stability in predicting the risk of POD, which is recommended to be applied in clinical practice to optimize postoperative management and reduce delirium incidence.

Keywords: Hip fracture; Nomogram; Postoperative delirium; Retrospective study; Risk factor.

Figure 1

Feature screening and Venn diagram analysis based on Lasso regression, support vector machine, and random forest. A: Important features and their coefficients screened out by Lasso regression; B: Important features and their importance scores screened out by support vector machine; C: Important features and their impact on model accuracy screened out by random forest; D: The Venn diagram shows the features commonly screened out by the three methods, including age, previous history of cognitive impairment, preoperative red blood cell count, preoperative albumin, preoperative hemoglobin, preoperative platelet count, and intraoperative blood loss. ASA: American Society of Anesthesiologists; SVM: support vector machine; COPD: Chronic obstructive pulmonary disease; BMI: Body mass index.

Figure 2

Univariate logistic regression for screening risk factors of postoperative delirium. HR: Hazard ratio; CI: Confidence interval; OR: Odds ratio; RBC: Red blood cell; PLT: Preoperative platelet count.

Figure 3

The value of 7 characteristic factors in predicting postoperative delirium in patients. A: Age; B: History of cognitive impairment; C: Preoperative red blood cells; D: Preoperative albumin; E: Preoperative hemoglobin; F: Preoperative platelet; G: Intraoperative blood loss. AUC: Area under the curve; RBC: Red blood cell; PLT: Preoperative platelet count.

Figure 4

Multivariate logistic regression for screening risk factors of postoperative delirium. HR: Hazard ratio; CI: Confidence interval; OR: Odds ratio; NA: Not available; RBC: Red blood cell; PLT: Preoperative platelet count.

Figure 5

Construction process of the risk prediction model for delirium based on six characteristic factors. RBC: Red blood cell; PLT: Preoperative platelet count.

Figure 6

Risk model validation. A: The receiver operating characteristic curve shows the discriminatory capacity of the model in the training group, with an area under the curve of 0.833; B: The precision-recall curve presents the model’s precision at different precision-recall rates in the training group; C: The calibration curve presents the relationship between the predicted probability of the model and the actual incidence rate in the training group; D: The decision curve analysis curve indicates that the model has clinical benefits within the risk threshold range of 0%-30% in the training group; E: The receiver operating characteristic curve shows the discriminative ability of the model in the validation group, with an area under the curve of 0.850; F: The precision-recall curve displays the accuracy of the model at different precision-recall rates in the validation group; G: The calibration curve displays the relationship between the predicted probability of the model and the actual occurrence rate in the validation group; H: The decision curve analysis curve indicates that the model has clinical benefits within the 0%-36% risk threshold range in the validation group. TPR: True positive rate; FPR: False positive rate; AUC: Area under the curve; CI: Confidence interval.