Leveraging machine learning for precision medicine: a predictive model for cognitive impairment in cholestasis patients

Abstract

Background: Cholestasis, characterized by impaired bile flow, impacts cognitive function through systemic mechanisms, including inflammation and metabolic dysregulation. Despite its significance, targeted predictive models for cognitive impairment in cholestasis remain underexplored. This study addresses this gap by developing a machine learning-based predictive model tailored to this population.

Methods: Clinical and biochemical data from Qingyang People's Hospital (2021-2023) were used to train and validate models for predicting cognitive impairment (MoCA ≤ 17). Recursive feature elimination identified critical predictors, while LightGBM and other machine learning models were evaluated. SHAP analysis enhanced model interpretability, and clinical utility was assessed through decision curve analysis (DCA).

Results: LightGBM outperformed other models with an AUC of 0.7955 on the testing dataset. Age, plasma D-dimer, and albumin were key predictors. SHAP analysis revealed non-linear interactions among features, demonstrating the model's clinical alignment. DCA confirmed its utility in improving patient stratification.

Conclusion: The developed LightGBM-based model effectively predicts cognitive impairment in cholestasis patients, providing actionable insights for early intervention. Integrating this tool into clinical workflows can enhance precision medicine and improve outcomes in this high-risk population.

Keywords: Cholestasis; Cognitive impairment; LightGBM; Machine learning; Predictive modeling.

Fig. 1

Baseline characteristics and feature selection results. A Clinical and laboratory characteristics of the study population, stratified by cognitive impairment status (MoCA ≤ 17). Significant group differences are marked. B Feature importance ranking using recursive feature elimination (RFE) with a random forest algorithm. (C) Final selection of 13 features contributing most significantly to the predictive model

Fig. 2

Performance of machine learning models. A, B ROC curves for training (A) and testing (B) datasets comparing all models. LightGBM shows the best balance of performance and generalizability. C, D Predicted probability distributions for training (C) and testing (D) datasets, showing the separation between cognitive impairment and non-impairment groups

Fig. 3

Clinical utility of the LightGBM model. A, B Decision curve analysis (DCA) for training (A) and testing (B) datasets, showing net benefit across risk thresholds. C, D Clinical impact curve (CIC) for training (C) and testing (D) datasets, illustrating the alignment of predicted high-risk cases with true positives

Fig. 4

Model Interpretability Using SHAP Analysis (A) SHAP summary plot showing the overall importance of features in the LightGBM model, with age, plasma D-dimer, and albumin identified as the most influential predictors. The color gradient represents the feature values, where higher or lower values impact predictions in specific directions. B Mean absolute SHAP values ranking feature importance, highlighting the significant contribution of age, plasma D-dimer, and metabolic markers. C SHAP dependence plots illustrating the relationships between feature values and their contributions to the model, revealing nonlinear effects for age, plasma D-dimer, and albumin. D SHAP interaction plot showing the interplay between age and plasma D-dimer, demonstrating how their combined effects influence predictions. E SHAP waterfall plot providing an individual prediction explanation, detailing how specific features cumulatively contribute to the predicted probability of cognitive impairment