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

OBJECTIVE: Heterogeneous treatment effect (HTE) is an important method for studying treatment effects between features in physiological measurements. This study proposes a new method for estimating HTE in physiological measurements, namely slide-windowbased double machine learning (slide-DML). HTE exhibit covariate dependence, therefore demonstrate different treatment effect under varying conditions. Current non-parametric methods for estimating HTE have complex computational mechanisms, limiting their applicability in realistic scenarios. METHODS AND PROCEDURES: Slide-DML estimate HTE by sliding windows to capture local similarity in heterogeneous features, combined with machine learning algorithms. RESULTS: We investigate the feasibility of the proposed method through simulations using synthetic data generated from statistical models. The simulation results indicate a mean square error of 0.006 in statistical models data, which is superior to machine learning-based HTE estimation methods. In the semi-synthetic data experiments, the proposed method achieved a root mean squared error of 3.526, outperforming other machine learning-based and deep learning-based approaches. Furthermore, real-life measurements demonstrate that the predictive outcomes of this method are more aligned with clinical reasoning and possess better interpretability. CONCLUSION: Slide-DML can effectively and accurately predict HTE. Clinical and Translational Impact Statement: We propose a new method for estimating HTE using physiological measurement devices, which enables a better analysis of the impact of heart rate on blood pressure estimation in photoplethysmogram. This approach can assist doctors in making more effective health plans in home healthcare settings.