Robust morris screening method (RMSM) for complex physiological models

Morris screening sensitivity analysis (MSM) comes forth as the method needing the minimum number of model simulations to qualify the impact of input parameter variations on outputs of complex, nonlinear and overparametrized models. However, the reliability of MSM indices (mean and standard deviation) and the reproducibility of their results are rarely explored despite the input parameter tuning/identification needs. In fact, these models, such those used in medical applications as digital twins, often lie in this category and need efficient and robust tools to assess both sensitivity and reliability of the outputs to numerous input model parameters.

In this study, a new Robust Morris Screening Method (RMSM) is proposed and based on new indices: the absolute median (χ^*) and the median absolute deviation (ρ). The proposed RMSM approach is evaluated on a complex multi-scales neuromuscular electrophysiological model simulating HD-sEMG (high density surface electromyography) signals at the skin surface. The reliability and stability of new RMSM indicators are evaluated at different trajectories within the parameter space and compared to classical MSM results. For this purpose, We propose a new methodology for parameter screening based on the ratio ρ/χ^* as a graphic indicator of (non)linearity and (non)monotonicity of parameter effects.

Firstly, the results demonstrated that the computed elementary effects (EE) of inputs are not normally distributed using MSM indices contrary to the proposed RMSM indices. Secondly, the ranking stability of RMSM indices was earlier obtained from 20 trajectories (T=20), while MSM ranking remained unstable until T = 100. Thirdly, The screening separation between influential and negligible input model parameters was more distinct and interpretable with RMSM than MSM.

The proposed RMSM approach ensures a fast, reliable and stable ranking of parameters for complex and overparametrized models compared to classical MSM. this allows a more precise exploration of the model parameter influence space for future application in parameter tuning and identification.