Time series machine learning for idionomic process-based treatment planning: A tutorial on tsBoruta

Abstract

Background This study showcases how three advanced algorithms—iARIMAX, iBoruta, and tsBoruta—identify personalized treatment processes in clinical psychology, using hair-pulling (trichotillomania) as an empirical case. Method We compared these methods with previous findings and assessed their ability to detect linear and nonlinear associations. We predicted the methods to converge on cognitive fixation as a core predictor of hair-pulling and expected substantial heterogeneity in process-outcome associations—heterogeneity that, if systematic rather than random, could inform the design of personalized interventions. We also predicted tsBoruta would outperform iBoruta due to its consideration of time series elements. Results All three methods confirmed cognitive fixation as a key aggregate level predictor of hair-pulling. While iARIMAX initially showed stronger connections, these became more modest—but still meaningful—when accounting for multiple processes. The Boruta methods showed notable differences in idiographic conclusions, with tsBoruta proving more conservative in confirming significant effects. Notably, 61.11% of participants showed unique combinations of relevant process-outcome links. Targeting three key processes of cognitive fixation, valued action, and anxiety could potentially benefit 52 out of 54 individuals in the sample. Conclusions The findings support combining standardized protocols with personalized interventions that may be valuable for trichotillomania treatment. More broadly, this study provides a methodological tutorial and illustrates how tsBoruta offers a powerful, balanced approach to modeling complexity in clinical data for treatment planning.