Prediction of key abnormal biochemical parameters in femoral neck fractures: an AI approach

Abstract

Purpose:

This study aims to develop machine learning models to predict perioperative biochemical abnormalities in femoral neck fracture patients, optimising treatment strategies and enhancing outcomes.

Methods:

A retrospective analysis was performed on a local clinical registry dataset, which included patients undergoing femoral neck fracture surgery from 2023 to 2024. The study focused on analysing preoperative and postoperative potassium, haemoglobin, and albumin concentrations. 6 ML algorithms were developed for prediction. Model interpretability was revealed using the control variable method, and robustness was enhanced through external data validation.

Results:

A total of 220 patients who completed the questionnaire and clinical tests were included in the study. Additionally, external data validation was performed on 15 patients beyond the initial cohort. Among the 6 ML algorithms used to predict biochemical indicators in patients with femoral neck fractures, SVR achieved the best performance in predicting preoperative potassium concentration K*, with an R² of 0.792 and an MAE of 0.335 mmol/L. Additionally, XGBoost showed good performance in predicting K, HGB*, HGB, ALB*, and ALB, with particularly excellent results in predicting HGB, achieving an R² of 0.943 and an MAE of only 0.478 g/L [* preoperative concentration].

Conclusions:

This study developed several ML-based predictive models that effectively assess changes in perioperative biochemical parameters in patients with femoral neck fractures. The interpretability heatmap clearly indicated the clinical features most influential on each biochemical parameter, such as the close relationship between K* and creatinine, which aligns with kidney regulation mechanisms and existing physiological knowledge. External data validation further demonstrated the model’s robustness, suggesting that the model is applicable not only to the existing dataset but also to a broader clinical population. Overall, the proposed model provides an effective tool for perioperative management, with promising potential for clinical practice to help optimise treatment strategies and improve patient outcomes and quality of life.