Towards proactive palliative care in oncology: developing an explainable EHR-based machine learning model for...

Area of Science:

• Oncology
• Medical Informatics
• Machine Learning

Background:

• Early identification of end-of-life is crucial for proactive palliative care.
• Machine learning (ML) models using electronic health records (EHR) show promise for cancer prognostication.
• Existing models often lack performance transparency, clinical alignment, and explainability, hindering adoption.

Purpose of the Study:

• To develop an explainable ML model using EHR data to predict 365-day mortality risk in advanced cancer patients.
• To facilitate the early integration of palliative care services in outpatient settings.
• To enhance clinical trust and adoption through model transparency and explainability.

Main Methods:

• A cohort of 5,926 advanced cancer patients (Stage 3-4 solid organ) was analyzed.
• Extreme Gradient Boosting (XGBoost) was employed to predict 365-day mortality from outpatient EHR data.
• Shapley Additive Explanations (SHAP) were used for model interpretability, with performance assessed by AUROC, AUPRC, and Brier score.

Main Results:

• The model achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.861 and an Area Under the Precision-Recall Curve (AUPRC) of 0.771.
• A Brier score of 0.147 indicated slight overestimation of mortality risk.
• SHAP values provided global and individual feature impact visualizations.

Conclusions:

• The developed ML model effectively predicts 365-day mortality in advanced cancer patients.
• The model demonstrates strong discrimination and precision-recall, supporting personalized prognostication.
• This tool can aid in the earlier integration of palliative care, improving patient outcomes.