Assessing robust prediction models without test datasets: A causal discovery approach on near-infrared spectra

Abstract

Machine learning prediction models calibrated with spectral data use correlations between variables without considering causation. The absence of genuine cause–effect relations hinders the ability to ensure methodical prediction reproducibility. Therefore, tools supporting causal-based discovery are essential in spectroscopy and chemometrics to enhance robustness. Accordingly, this study invokes causal inference theory to establish the causal discovery index (CDI) to distinguish datasets with reliable causal structures from those prone to spurious correlations. This framework was applied to seven simulated near-infrared spectral causal structures. Simulated near-infrared spectra were utilized to ensure that the framework performance was optimized and verified appropriately in a generalized methodology. Reliable structures were confirmed to be differentiated by the differences in the mean and standard deviation of bootstrapped CDI indices. Distinctive thresholds for the mean and standard deviation were established at the sample size of 1000 and 10,000. The framework consistently performed well with multiple spectral preprocessing methods such as derivation and dimension reduction. It was also robust with variations, surpassing the conventional test-set validation method without the use of additional independent datasets. This would benefit the applicability of the novel framework in practical situations where dataset collection can be limited. Moreover, it can be extended to various sensor-based data, encompassing only seven possible causal structures.