Machine Learning Models for Prediction of Response to Therapeutic Plasma Exchange in Pediatric Acute Liver Failure

Therapeutic plasma exchange (TPE) artificially lowers prognostic markers such as INR and bilirubin, complicating timely decisions regarding liver transplantation (LT) in acute liver failure (ALF).

We aimed to utilize machine learning (ML) models for predicting response to TPE and to create an online clinical decision support tool.

Children aged 2-18 years with ALF who underwent at least one TPE session were included. Data were retrieved from a prospectively maintained database which included clinical data and serial biochemical variables. Multiple classifiers (random forest, logistic regression [LR], gradient boosted trees, extreme gradient boosted trees [XGBoost], support vector machine [SVM]) were trained and evaluated.

Of the 511 pediatric ALF patients admitted during the study period, 139 underwent TPE. After excluding LT, 110 were included in the primary analysis. The ML classifiers had maximum accuracy at T2 (12-18 h post-first TPE) and T4 (12-18 h post the second TPE). Among the various ML classifiers evaluated, LR, SVM, and XGBoost demonstrated highest discrimination at T2 (area under the curve [AUC]: 0.802-0.812; sensitivity: 79.6-80.8%, specificity: 70.2-73.4%, and positive predictive value: 83-84.1%) with limited improvement after data augmentation using synthetic minority oversampling technique. The AUCs of these three ML models were slightly lower at T4 (0.722-0.783). These sets of models, trained separately for real-time prediction of TPE response at baseline, T2, and T4, were incorporated into a web-based tool to guide decision-making in pediatric ALF.

ML models based on LR, SVM, and XGBoost trained on serial clinical and biochemical data, particularly those acquired 12-18 h after each TPE session, can provide clinically meaningful and timely predictions of response to TPE, thus assisting clinical decision-making in pediatric ALF.