Interpretable machine-learning risk prediction of unplanned extubation among cancer patients with peripherally...

Unplanned extubation of peripherally inserted central catheters (PICC-UE) in patients with cancer has been linked to factors including women, diabetes, thrombosis history, valved catheter, double-lumen catheter, and self-management. However, the effect of patient quality of life has not been explored. Therefore, this study aimed to analyze PICC-UE risk factors using a machine-learning algorithm, focusing on the role of patient quality of life. A total of 212 cancer patients who underwent PICC catheterization were included in this study from February 2021 to June 2022. Patients were categorized into two groups based on PICC-UE occurrence: the PICC-UE group (n = 23) and the non-PICC-UE group (n = 189). Referring to previous reports and professional cognition, data of 30 potential risk factors within one week before extubation were collected, with a focus on incorporating health-related quality of life (HRQOL) and patient self-management scores. PICC-UE risk factors were examined using four machine-learning algorithms with three encoding methods and four data imbalance processing methods. Then, the key factors causing PICC-UE were interpreted using the SHapley Additive exPlanations (SHAP) tool. PICC-UE occurred in 23 of 212 patients (overall incidence: 10.8%). The HRQOL score, which has been underexplored in prior studies, demonstrated a statistically significant difference between the PICC-UE and non-PICC-UE groups (P < 0.001) and exhibited a strong association with patient self-management score, as evidenced by its concentration in the upper right quadrant of the planar scatter plot. As a novel derivative composite metric, weighted quality of life (WQOL), calculated as the product of HRQOL and self-management scores, was identified as the most influential risk factor for PICC-UE, surpassing both individual self-management and HRQOL scores (SHAP-value = 1.02 vs. 0.22 and 0.00). Furthermore, increased upper arm circumference was also found to be a significant predictor (SHAP value = 0.22). This study demonstrates the synergistic effect of patients' quality of life and self-management capacity on the PICC-UE occurrence. The WQOL metric, which integrates both factors, serves as a significant predictor of PICC-UE occurrence and should therefore be considered an essential component in clinical assessment.