On the interpretability of machine learning-based model for predicting hypertension.

Area of Science:

• Cardiology
• Machine Learning
• Medical Informatics

Background:

• Complex machine learning models often outperform traditional ones but lack clinical trust due to poor interpretability.
• Clinicians require understandable explanations for machine learning predictions to ensure trust and effective use in medical decision-making.

Purpose of the Study:

• To demonstrate the utility of various model-agnostic explanation techniques for machine learning models.
• To analyze the outcomes of a random forest model predicting hypertension risk using cardiorespiratory fitness data.

Main Methods:

• Applied five global interpretability techniques (Feature Importance, Partial Dependence Plot, Individual Conditional Expectation, Feature Interaction, Global Surrogate Models) and two local interpretability techniques (Local Surrogate Models, Shapley Value).
• Utilized a dataset of 23,095 patients with exercise treadmill stress testing data and 10-year follow-up.

Main Results:

• Different interpretability techniques offer distinct insights into model behavior.
• Global interpretations provide insights into the entire conditional distribution, aiding understanding of the overall model.
• Local interpretations focus on specific instances, enhancing understanding of individual predictions.

Conclusions:

• Both global and local interpretability techniques are valuable for assisting clinicians in medical decision-making.
• Global techniques generalize across populations, while local techniques focus on instance-level explanations.
• Clinicians retain final decision-making authority, integrating model explanations with their domain expertise.