Development of a Risk Prediction Model for Moderate-to-Severe Complications Associated With Peripheral Intravenous Catheters in Older Patients

Objective To develop and validate a risk prediction model for moderate-to-severe complications associated with peripheral venous catheters (PVC) in older patients, and to provide evidence for precision-based clinical prevention and management.

Methods In this prospective cohort study, the clinical data of 750 older patients who underwent PVC placement for intravenous therapy between November 2023 and September 2024 were collected. Patients were stratified and randomly assigned to a training set (n = 526) and a test set (n = 224) at a ratio of 7∶3. Independent risk factors for moderate-to-severe PVC-related complications were identified by univariate and multivariate logistic regression analyses. A total of 4 predictive models were established—binary logistic regression (BLR), support vector machine (SVM), random forest (RF), and naive Bayes (NB). Model discrimination was assessed using the area under the receiver operating characteristic curve (AUC) and DeLong test. Model calibration was evaluated by calibration curves. Clinical net benefit was determined using decision curve analysis (DCA). Shapley additive explanations (SHAP) were used to visualize and interpret feature importance in the models.

Results According to multivariate logistic regression analysis, functional status (mild dependence: odds ratio OR = 1.361, 95% CI: 1.031-1.802; moderate dependence: OR = 1.992, 95% CI: 1.291-3.061; severe dependence: OR = 3.060, 95% CI: 1.700-5.512), infusion extravasation (OR = 8.998, 95% CI: 4.286-18.887), the clinical department the patient was admitted to (Gastrointestinal Surgery: OR = 28.396, 95% CI: 7.537-106.990; Pancreatic Surgery: OR = 5.258, 95% CI: 1.177-23.499; Pancreatitis Center: OR = 5.974, 95% CI: 1.287-27.725), primary school or less formal education (OR = 2.525, 95% CI: 1.158-5.506), drinking history (OR = 1.970, 95% CI: 1.011-3.838), the edges of the dressing being curled up (OR = 1.458, 95% CI: 1.150-1.848), and the number of comorbidities (OR = 0.893, 95% CI: 0.821-0.972) were identified as independent risk factors for moderate-to-severe PVC-related complications in older patients, while 16G-22G catheters were identified as an independent protective factor (P < 0.05). Validation in the test set revealed that the RF model yielded the highest AUC value (0.891, 95% CI: 0.840-0.942). According to the DeLong test, the discrimination performance of the RF model was significantly superior to those of the BLR model (AUC = 0.808, 95% CI: 0.738-0.879) and the SVM model (AUC = 0.842, 95% CI: 0.774-0.909) (P < 0.05). SHAP analysis revealed that infusion extravasation, functional status, and the clinical department into which the patient was admitted were the top 3 features of the RF model. DCA showed that the RF model provided significant net clinical benefit within the threshold probability range of 0-0.8.

Conclusion In this study, an RF prediction model for moderate-to-severe PVC-related complications in older patients was successfully developed and validated. The RF prediction model demonstrates excellent discrimination and high clinical applicability, enabling early identification of high-risk patients and informing precision-based clinical interventions.