[Paper Review] Machine Learning Model Interpretability for Precision Medicine

Gajendra J. Katuwal, Robert F. Chen|arXiv (Cornell University)|Oct 28, 2016

Explainable Artificial Intelligence (XAI)2 references61 citations

TL;DR

This paper proposes using Model-Agnostic Explanations (SHAP) to enhance interpretability of complex machine learning models in precision medicine. Using the MIMIC-II dataset, the authors achieve 80% balanced accuracy in predicting ICU mortality while providing individual-level feature importance, enabling clinicians to understand model decisions transparently.

ABSTRACT

Interpretability of machine learning models is critical for data-driven precision medicine efforts. However, highly predictive models are generally complex and are difficult to interpret. Here using Model-Agnostic Explanations algorithm, we show that complex models such as random forest can be made interpretable. Using MIMIC-II dataset, we successfully predicted ICU mortality with 80% balanced accuracy and were also were able to interpret the relative effect of the features on prediction at individual level.

Motivation & Objective

• To improve interpretability of complex machine learning models used in precision medicine.
• To enable clinicians to understand the contribution of individual features to model predictions at the individual patient level.
• To demonstrate that high-performance models can be made interpretable without sacrificing predictive accuracy.
• To validate the approach on a real-world critical care dataset (MIMIC-II).
• To bridge the gap between high predictive performance and clinical trust in AI-driven decision support.

Proposed method

• The authors apply the Model-Agnostic Explanations (SHAP) framework to interpret predictions from a trained random forest model.
• SHAP values are computed to quantify the contribution of each clinical feature to the prediction for each individual patient.
• The method is model-agnostic, allowing interpretation of any black-box model, including ensemble methods like random forests.
• The MIMIC-II critical care dataset is used, containing clinical features such as vital signs, lab values, and comorbidities.
• The model is trained to predict ICU mortality with a focus on both accuracy and interpretability.
• Feature importance is visualized at both global and individual levels to support clinical insight.

Experimental results

Research questions

• RQ1 Can complex, high-performing machine learning models used in precision medicine be made interpretable without compromising performance?
• RQ2 What is the relative contribution of individual clinical features to ICU mortality predictions at the individual patient level?
• RQ3 How can model interpretability be systematically applied to clinical decision support systems?
• RQ4 Can SHAP-based explanations enhance trust and clinical utility of predictive models in intensive care settings?
• RQ5 What level of predictive accuracy can be achieved while maintaining sufficient interpretability for clinical deployment?

Key findings

• The random forest model achieved 80% balanced accuracy in predicting ICU mortality on the MIMIC-II dataset.
• SHAP-based explanations successfully identified the most influential clinical features for individual patient predictions.
• The method enabled interpretation of model decisions at the individual level, revealing patient-specific feature impacts.
• The interpretability framework was applied to a black-box model without requiring architectural changes.
• The results demonstrate that high-performing models can be made transparent and clinically actionable through post-hoc explanation techniques.
• The approach supports clinical decision-making by providing traceable, feature-level insights into model predictions.