Objective: To preregister a confirmatory secondary analysis that derives and externally validates simple, interpretable personalisation rules for adaptive virtual reality exposure therapy (VRET) based on individual differences in timing-sensitive defensive responding and awareness. We leverage two existing fear-conditioning datasets collected in a physical laboratory and immersive VR to: (i) predict extinction success from pre-acquisition and early-acquisition features, (ii) quantify the incremental value of awareness (trial-wise expectancy in lab; interoceptive awareness in VR), and (iii) translate results into practical decision rules that recommend timing/predictability parameters for VRET. Design and data: Parallel human conditioning experiments with identical core manipulations of CS–US lead interval (short ≤200 ms; long ≥1000 ms) and predictability (paired vs unpaired), conducted in a physical lab (N=59 analysed) and in VR (N=81 acquisition; N=70 extinction). Eyeblink startle indexed defensive responding; the lab included trial-wise expectancy; VR included MAIA-2 interoceptive awareness. Analyses focus on the common-interval set (200/1000/4000 ms); 100 ms (VR) is treated as exploratory due to prepulse inhibition (PPI). Outcome and predictors: The primary outcome is standardized extinction success at long intervals (≥1000 ms), defined a priori as the within-subject reduction in standardized CS+ startle from acquisition to extinction (and a CS+–CS− contrast variant), estimated as a slope and as an early–late contrast; outcomes are z-standardized within context. Predictors (restricted to pre-acquisition and early-acquisition features to enable prospective use) include: baseline startle/habituation, US intensity, sex, short-interval reactivity (≤200 ms) by predictability, long-interval discrimination during acquisition, and awareness measures (lab: expectancy indices; VR: preregistered MAIA composites with FDR control). Analysis plan: We will develop a base, context-agnostic model using common features and evaluate performance via nested cross-validation and external validation by context (train on lab, test on VR; and vice versa). We will then quantify the incremental predictive value of awareness measures (ΔR2/RMSE for continuous outcomes; AUC/ΔBrier if a preregistered dichotomous “success” threshold is applied), assess calibration (intercept/slope, calibration curves), and estimate clinical utility via decision curve analysis. Models prioritise interpretability (regularised regression and shallow decision trees) and preclude information leakage (predictors available before extinction). Sensitivity analyses include alternative outcome definitions (slope vs contrast; CS+ vs CS+–CS−), standardized vs raw units, and exclusion of 100 ms. Multiplicity for MAIA is controlled by Benjamini–Hochberg FDR. The primary hypotheses are: H1, base timing features predict extinction success; H2, awareness adds incremental value; H3, models generalise across contexts with acceptable calibration; H4, derived decision rules yield positive net benefit. All code, de-identified data, and a simple scorecard/nomogram implementation will be shared.
Bergsnev et al. (Fri,) studied this question.