Efficient Data Collection for Establishing Practical Identifiability via Active Learning

Practical identifiability analysis (PIA) plays a crucial role in model development by determining whether available data are sufficient to yield reliable parameter estimates. In bioengineering applications, identifying the minimal experimental design that ensures parameter identifiability is essential in order to reduce cost, time, and resource consumption. In this paper, we introduce E-ALPIPE, a sequential active learning algorithm that recommends new data collection points most likely to establish practical identifiability given the current data, mathematical model and noise assumptions. We empirically evaluate E-ALPIPE against both a benchmark algorithm from the literature and random sampling over three synthetic experiments. Our results show that E-ALPIPE requires up to 50% fewer observations on average to achieve practical identifiability, compared to the strongest competitor, while producing comparable or narrower confidence intervals and more accurate point estimates of system dynamics. Author summary Scientists and engineers often use computer models to understand complex biological systems, such as how cells grow or how chemicals behave in a reactor. To make these models useful, they need to determine the correct values for various model parameters, based on data from experiments. However, experiments are usually expensive, time-consuming, and often produce noisy data. If the available data is insufficient, our confidence in the fitted model parameters may be low and thus the model’s predictions might be unreliable. This is known as a problem of practical identifiability. This paper introduces a new method called E-ALPIPE, which helps scientists decide which experiment is likely to produce the most useful information for improving their model. Using E-ALPIPE they can determine the model parameter values more efficiently, saving time and resources. We tested E-ALPIPE on several example problems and found that it could reach the same level of reliability using up to 50% fewer experiments than standard methods. It also produced more precise and accurate results. Overall, E-ALPIPE is a smarter, more efficient way to design experiments for tuning the parameters of computer models.