Risk factors for complications associated with peripherally inserted central venous catheters for parenteral nutrition: Machine learning and survival analysis based on deep learning.

BACKGROUND & AIMS: Peripherally inserted central catheters (PICCs) are widely used in patients receiving total parenteral nutrition (TPN), and identifying risk factors for PICC-related complications is essential to improve patient outcomes. This study aims to develop artificial intelligence (AI)-based survival analysis and machine learning models to predict PICC complications and identify significant risk factors.

METHODS

This study was designed as a retrospective medical record analysis. Data were collected from 218 patients who underwent PICC insertion. Logistic regression, support vector machine, random forest, and extreme gradient boosting were used to develop discrete complication prediction models, whereas survival analysis models, including random survival forest, DeepSurv, and DeepHit, were used to create time-varying complication prediction models. Model performance was evaluated using accuracy for complication occurrence and the concordance index (C-index) and integrated Brier score (IBS) for catheter use.

RESULTS

Complication prediction achieved a mean accuracy of 0.92. Among the survival models, DeepSurv exhibited the best C-index (0.61) but a relatively higher IBS (0.170). Significant complication risk factors included the catheter insertion site, catheter diameter, gender, cancer diagnosis, and timing of the PICC insertion decision. Left-arm insertion and larger catheter diameters were associated with higher complication risks.

CONCLUSIONS

This study is significant in developing a PICC complication prediction model to support clinical decision-making and explaining the model's functioning using explainable AI (XAI) techniques.