Development and Validation of an Explainable Machine Learning Model for Predicting Invasive Fungal Infection in
Fei-Xiang Xiong1, Jian-Guo Yan2, Xue-Jie Zhang1
1Center of Integrative Medicine, Beijing Ditan Hospital, Capital Medical University, Beijing, China.
Background And Objective:
Acute-on-chronic liver failure (ACLF) is associated with significantly higher short-term mortality, and the presence of invasive fungal infection (IFI) further increases this risk. This study aims to develop a ML model that predicts the risk of IFI in ACLF patients.
Methods:
This study included 1112 patients divided into a training set and a validation set, with another 188 patients serving as an external validation cohort. The Recursive Feature Elimination (RFE) method was used to select the most significant variables for model development. Four machine learning algorithms were compared to identify the optimal model. The models were evaluated and compared using C-index, time-dependent ROC curves, decision curve analysis (DCA), and calibration curves. The LIME (Local Interpretable Model-Agnostic Explanations) method was used to identify the high-risk populations utilised by the model.
Results:
778 patients were included in the training set, 334 in the internal validation set, and 188 in the external validation set. The study found that Random Forest (RF) was the best-performing ML algorithm. In the training set, the RF model achieved an AUROC of 0.922 (0.911-0.933), significantly higher than MELD (0.854, 0.835-0.873, p < 0.001), CLIF-C OF (0.753, 0.724-0.783, p < 0.001), and CLIF-C ACLF (0.879, 0.863-0.896, p = 0.020). The same trend was observed in both the internal and external validation sets. The time-dependent ROC curve showed that the RF model outperformed the other scores for predicting the risk of IFI in 28 days. DCA and calibration curves also demonstrated superior clinical benefits for the RF model across all datasets. LIME revealed bacterial infection (BI), Na < 136 mmol/L, CRP (C-reactive protein) > 20.1 g/L, and TBIL(Total Bilirubin) > 196.7 μmol/L as the high-risk groups.
Conclusion:
The RF model effectively predicts the risk of IFI in ACLF patients. The application of LIME enables the identification of high-risk populations, providing clinical value for patient management.
Insights
This study developed a Random Forest model to predict invasive fungal infection (IFI) risk in acute-on-chronic liver failure (ACLF) patients, outperforming existing scores. The model identifies high-risk factors like bacterial infection and specific lab values for better patient management.
Area of Science:
- Machine Learning in Clinical Medicine
- Hepatology and Critical Care
Background:
- Acute-on-chronic liver failure (ACLF) carries a high short-term mortality risk.
- Invasive fungal infection (IFI) significantly exacerbates mortality in ACLF patients.
- Accurate prediction of IFI risk is crucial for timely intervention and improved outcomes.
Purpose of the Study:
- To develop and validate a machine learning (ML) model for predicting the risk of IFI in ACLF patients.
- To compare the performance of the developed ML model against established clinical scores.
- To identify key predictors of IFI in ACLF patients using interpretable ML techniques.
Main Methods:
- A cohort of 1112 ACLF patients was divided into training, internal validation, and external validation sets.
- Recursive Feature Elimination (RFE) was employed for optimal variable selection.
- Four ML algorithms were evaluated, with Random Forest (RF) selected as the best performer. Model performance was assessed using C-index, time-dependent ROC curves, decision curve analysis (DCA), and calibration curves. Local Interpretable Model-Agnostic Explanations (LIME) was used for risk factor identification.
Main Results:
- The RF model demonstrated superior predictive performance for IFI risk compared to MELD, CLIF-C OF, and CLIF-C ACLF scores across all validation sets (AUROC 0.922 in training set).
- Time-dependent ROC curves confirmed the RF model's advantage in predicting 28-day IFI risk.
- LIME analysis identified bacterial infection, low sodium levels (Na < 136 mmol/L), high C-reactive protein (CRP > 20.1 g/L), and high total bilirubin (TBIL > 196.7 μmol/L) as significant risk factors.
Conclusions:
- The developed Random Forest model provides an effective tool for predicting IFI risk in ACLF patients.
- The model's interpretability through LIME allows for the identification of high-risk patient subgroups.
- This approach offers significant clinical value for optimizing patient management and potentially reducing IFI-related mortality in ACLF.
