# An interpretable machine learning model with SHAP explanations predicts spontaneous bleeding in pediatric acute liver failure

## TL;DR

A machine learning model with SHAP explanations was developed to predict spontaneous bleeding in children with acute liver failure, offering improved accuracy and interpretability for clinical decision-making.

## Contribution

The novel use of SHAP explanations in a Gradient Boosting Machine model to predict and interpret spontaneous bleeding risk in pediatric acute liver failure.

## Key findings

- The GBM model achieved an AUC of 0.858 in internal validation and 0.839 in external validation for predicting spontaneous bleeding.
- Key predictors included platelet count, infection, MODS, HRS, D-dimer, total protein, and lactic acid levels.
- SHAP analysis showed infection, MODS, and HRS increase bleeding risk, while higher platelet and protein levels are protective.

## Abstract

Pediatric acute liver failure (PALF) is a severe clinical syndrome associated with a high risk of spontaneous bleeding, leading to increased mortality and poor outcomes. Traditional methods for predicting bleeding risk in PALF are limited, highlighting the need for more accurate and interpretable models. This study aimed to develop and validate a machine learning (ML) model for predicting spontaneous bleeding in pediatric patients with PALF, leveraging the SHapley Additive exPlanations (SHAP) method to enhance interpretability.

A retrospective observational cohort study was conducted using data from the Clinical Science Research Big Data Platform at the Children’s Hospital of Chongqing Medical University. Data from 501 patients with PALF were used for model training and internal validation, and an independent cohort of 153 patients was used for external validation. Thirty-four clinical variables were selected based on expert input and prior research. Feature selection was performed using the Boruta algorithm and least absolute shrinkage and selection operator (LASSO) regression. Ten ML algorithms were assessed, and the Gradient Boosting Machine (GBM) model was selected for its superior performance. Model evaluation metrics included the area under the curve (AUC), accuracy, recall, specificity, precision, F1 score, Brier score, calibration curves, and decision curve analysis (DCA). SHAP values were employed to interpret the model’s predictions.

The GBM model achieved an AUC of 0.858 (95% CI, 0.778–0.899) in internal validation and 0.839 (95% CI, 0.774–0.904) in external validation. Key predictors of spontaneous bleeding included platelet count, infection, multiple organ dysfunction syndrome (MODS), hepatorenal syndrome (HRS), D-dimer, total protein, and lactic acid levels. SHAP analysis demonstrated that infection, MODS, and HRS were positively associated with bleeding risk, while higher platelet counts, total protein, and fibrinogen levels were protective. Calibration curves and DCA confirmed the model’s clinical utility and generalizability.

The proposed ML model exhibits strong predictive performance and interpretability for spontaneous bleeding in pediatric patients with PALF. This tool may aid clinicians in identifying high-risk patients and guiding clinical interventions. Future research should focus on validating the model with more diverse datasets and exploring predictions of bleeding severity and specific complications.