JMIR Publications

ABSTRACT

Background:

Given the increasing number of peritoneal dialysis patients and the importance of optimal resource allocation, an effective length of hospital stay (LOS) prediction model is needed by physicians.

Objective:

We aimed to construct an LOS prediction model for peritoneal dialysis patients using stacked generalization (stacking), and compare the prediction performance of the stacking model with the benchmark logistics regression (LR) model and its three base models: support vector machine (SVM), random forest (RF) and K-nearest neighbor (KNN).

Methods:

Patient data collected using the Hospital Quality Monitoring System (HQMS) in China were used to develop and validate the models. We constructed the stacking model using three machine learning methods: SVM, RF, and KNN as its base models and LR as its meta-model. In addition, a LR model was constructed as benchmark model. The prediction performance of the ensemble model was compared with each base model and the benchmark model. Five-fold cross-validation was employed to develop and validate the models. In performance comparison, the prediction performance of each model was measured using Brier score, area under the receiver operating characteristic curve (AUROC), estimated calibration index (ECI), and geometric mean (Gm). Further, calibration plot was employed to visually demonstrate the calibration power of each model.

Results:

The final cohort in our study consisted of 22,859 eligible peritoneal dialysis patients, among which 25.2% had a prolonged LOS (pLOS). The results showed that the stacking model achieved the best overall performance (Brier score, 0.157), discrimination (AUROC, 0.756), and calibration (ECI, 7.846); the second-best balanced accuracy (Gm, 0.688). Compared with the benchmark LR model, the stacking model outperformed in terms of all performance measures.

Conclusions:

The stacking-based LOS prediction model has superior prediction performance compared to its base models and traditional LR model. The stacking model has great potential to assist physicians to do LOS prediction for peritoneal dialysis patients, and it may be helpful in assisting physicians to optimally allocate medical resources and achieve better clinical outcomes. However, a prospective study is still needed to evaluate the stacked generalization model before its use in clinical practice.