Slide-DML: Sliding-window-based Estimation of Heterogeneous Treatment Effects

Area of Science:

• Biomedical Engineering
• Machine Learning
• Data Science

Background:

• Heterogeneous treatment effect (HTE) analysis is crucial for understanding variable treatment effects in physiological measurements.
• Current non-parametric HTE methods are computationally complex and limited in real-world applications.
• Covariate dependence in HTE necessitates advanced estimation techniques for accurate physiological assessments.

Purpose of the Study:

• To propose a novel method, slide-window-based double machine learning (slide-DML), for estimating HTE in physiological measurements.
• To address the limitations of existing non-parametric HTE methods with a computationally efficient approach.
• To enhance the accuracy and interpretability of HTE estimation in physiological data.

Main Methods:

• Slide-DML utilizes sliding windows to identify local similarities in heterogeneous features.
• Combines sliding window approach with machine learning algorithms for HTE estimation.
• Validated through simulations on synthetic and semi-synthetic physiological datasets.

Main Results:

• Slide-DML achieved a mean squared error of 0.006 on statistical model data, outperforming other machine learning methods.
• Demonstrated a root mean squared error of 3.526 on semi-synthetic data, surpassing deep learning approaches.
• Real-world physiological measurements showed improved clinical alignment and interpretability with slide-DML.

Conclusions:

• Slide-DML provides an effective and accurate method for predicting heterogeneous treatment effects.
• The method enhances analysis of physiological measurements, such as heart rate's impact on blood pressure from photoplethysmogram.
• Facilitates improved clinical decision-making and personalized healthcare planning, particularly in home settings.