Clinical characteristics and establishment of clinical prediction model for traumatic cholecystitis in older patients with hip fracture: a retrospective machine learning models and SHAP value-based interpretation model

Abstract Background Traumatic cholecystitis after hip fracture is rare, but its prognosis is poor. However, studies that explore the risk factors for traumatic cholecystitis in older patients with hip fractures and establish effective predictive models were limited. Therefore, our study focused on the following three questions: (1)What were the characteristics of traumatic cholecystitis in older hip fracture patients? (2) What were the risk factors for traumatic cholecystitis in older hip fracture patients? (3) What was the best model for predicting traumatic cholecystitis in older hip fracture patients and analyze it globally and locally interpretatively using SHAP values? Methods We retrospectively collected all clinical parameters of older hip fracture patients admitted to the Third Hospital of Hebei Medical University from January 2018 to December 2022. We developed five machine learning models for traumatic cholecystitis, which included logistic regression model(LR), random forest model(RF), lightGBM, XGBoost and KNN. We evaluate the predictive performance and clinical applicability of the LR model using the area under the ROC curve, calibration curve, and clinical decision curve. We evaluated the machine learning models by calculating AUC-ROC, PR-AUC, accuracy, sensitivity, specificity, precision and F1 score. SHAP values were used to interpret the models, enhancing clinicians’ understanding and application of the models. Results This study included 2376 patients, with an incidence rate of traumatic cholecystitis of 1.3% (30/2376). Independent risk factors for traumatic cholecystitis in older hip fracture patients included the C-reactive protein-to-albumin ratio (CAR), history of biliary disease, haemoglobin (HB), albumin, hypokalaemia, and CRP. RF showed the highest PR-AUC (0.940, 95% CI: 0.835–0.997) among the five models. The ROC-AUC for the five models were 0.999(95% CI: 0.998,1.000) (RF), 0.741(95% CI: 0.649,0.839) (XGBoost), 0.852(95% CI: 0.785,0.915) (LR), 0.971(95% CI: 0.955,0.984) (LightGBM), and 0.980(95% CI: 0.975,0.985) (KNN), with sensitivities ranging from 26.7% (LR) to 100% (RF). The accuracy of these models ranged from 91.8%% (LightGBM) to 99.7% (RF). The random forest model performed the best. Discussion Among the five machine learning models, the random forest model showed the best discriminative performance in the present dataset, but this exploratory finding may overestimate its true predictive capability because of the small number of positive cases and lack of external validation. SHAP values were used to improve model interpretability. In the absence of large-scale prospective validation, this random forest algorithm can serve as an auxiliary perioperative comprehensive risk stratification tool, helping clinicians identify potential high-risk older patients as early as possible and implement closer monitoring of abdominal signs.