[Paper Review] Using Explainable AI to Cross-Validate Socio-economic Disparities Among Covid-19 Patient Mortality

Shi Li, Redoan Rahman+4 more|PubMed|Feb 16, 2023

Machine Learning in Healthcare15 references4 citations

TL;DR

This study uses explainable AI (XAI) methods—SHAP and LIME—on an XGBoost model trained on de-identified EHR data from Austin hospitals to investigate socioeconomic disparities in COVID-19 mortality. It confirms that Medicare status, age, and gender are top predictors of mortality, with consistent feature importance across XAI methods, validating their use in cross-verifying model interpretability and health disparities research.

ABSTRACT

This paper applies eXplainable Artificial Intelligence (XAI) methods to investigate the socioeconomic disparities in COVID-19 patient mortality. An Extreme Gradient Boosting (XGBoost) prediction model is built based on a de-identified Austin area hospital dataset to predict the mortality of COVID-19 patients. We apply two XAI methods, Shapley Additive exPlanations (SHAP) and Locally Interpretable Model Agnostic Explanations (LIME), to compare the global and local interpretation of feature importance. This paper demonstrates the advantages of using XAI which shows the feature importance and decisive capability. Furthermore, we use the XAI methods to cross-validate their interpretations for individual patients. The XAI models reveal that Medicare financial class, older age, and gender have high impact on the mortality prediction. We find that LIME's local interpretation does not show significant differences in feature importance comparing to SHAP, which suggests pattern confirmation. This paper demonstrates the importance of XAI methods in cross-validation of feature attributions.

Motivation & Objective

• To investigate socioeconomic disparities in COVID-19 patient mortality using machine learning and explainable AI.
• To evaluate the consistency and reliability of feature importance across multiple XAI methods (SHAP and LIME).
• To validate model interpretations through cross-verification using multiple explainability techniques.
• To identify key clinical and social determinants influencing mortality outcomes in a real-world hospital dataset.
• To demonstrate the utility of XAI in uncovering and validating health disparities in clinical prediction models.

Proposed method

• An XGBoost model is trained on de-identified electronic health record (EHR) data from a single hospital system in Austin, Texas.
• SHAP (Shapley Additive Explanations) is used to compute global and local feature importance scores for model predictions.
• LIME (Locally Interpretable Model-agnostic Explanations) is applied to generate local interpretability by approximating the model locally around individual predictions.
• Feature importance rankings from SHAP and LIME are compared to assess consistency and cross-validate interpretations.
• The model predicts in-hospital mortality for COVID-19 patients using demographic, clinical, and socioeconomic variables.
• Cross-validation of XAI interpretations is performed to ensure robustness and reliability of feature attribution.

Experimental results

Research questions

• RQ1 To what extent do SHAP and LIME produce consistent feature importance rankings in predicting COVID-19 mortality?
• RQ2 Which socioeconomic and clinical features most significantly influence the model’s prediction of patient mortality?
• RQ3 How can XAI methods be used to validate and cross-verify model interpretations in health disparities research?
• RQ4 Are there measurable disparities in mortality prediction linked to financial class (e.g., Medicare) and age?
• RQ5 Can XAI methods reliably identify and confirm patterns in health outcome disparities across individual patient cases?

Key findings

• Medicare financial class emerged as a top predictor of mortality, indicating significant socioeconomic disparities in outcomes.
• Older age was consistently identified as a major factor influencing mortality prediction across both SHAP and LIME interpretations.
• Gender also showed high impact on mortality prediction, with differences observed in model explanations.
• LIME and SHAP produced highly similar local feature importance rankings, suggesting strong consistency and pattern confirmation between methods.
• The cross-verification of XAI interpretations enhanced confidence in the model’s feature attributions and their clinical relevance.
• The study demonstrates that XAI can be effectively used to validate and uncover disparities in clinical prediction models.