Deterministic and stochastic interventions in reducing drug–drug interactions in inappropriate prescribing: A systematic review
TL;DR
Current evidence does not demonstrate a proportional improvement in clinically reliable decision support in stochastic and generative models for DDI detection, and future interventions must prioritize hybrid architectures that integrate explainable rule-based guardrails with rigorously validated stochastic models.
Abstract
Background Drug–drug interactions (DDIs) remain a major contributor to preventable patient harm, particularly in the context of polypharmacy. Over the past two decades, interventions to mitigate inappropriate prescribing have evolved from deterministic, rule-based clinical decision support toward increasingly complex data-driven and stochastic models. However, the extent to which these methodological advances translate into improved clinical safety remains unclear. Methods We conducted a systematic review in accordance with PRISMA 2020 guidelines, guided by the SPIDER framework. PubMed, Scopus, ScienceDirect, and IEEE Xplore were searched from inception to October 2025 for primary studies evaluating computational or clinical decision support interventions aimed at reducing DDIs or inappropriate prescribing. Eligible studies included deterministic rule-based systems, ontological frameworks, and artificial intelligence-driven predictive models. Risk of bias was assessed using the Prediction Model Risk of Bias Assessment Tool, extended with artificial intelligence-specific considerations (PROBAST+AI). Due to heterogeneity in study designs and outcome measures, findings were synthesized narratively. Results Ten studies met the inclusion criteria. Earlier interventions predominantly employed deterministic approaches focused on workflow optimization, alert management, and policy enforcement, demonstrating modest improvements in prescribing processes but inconsistent links to patient-level outcomes. More recent studies applied stochastic and generative models using high-dimensional clinical datasets to predict DDIs, reporting strong internal performance metrics. However, PROBAST+AI assessment identified a consistently high risk of bias in the analysis domain for AI-driven studies, primarily due to limited external validation, insufficient calibration reporting, and unclear handling of overfitting and data leakage. Conclusions While stochastic and generative models offer enhanced predictive capacity for DDI detection, current evidence does not demonstrate a proportional improvement in clinically reliable decision support. Deterministic systems provide transparency and safety constraints but lack adaptability to patient-specific contexts. Future interventions must prioritize hybrid architectures that integrate explainable rule-based guardrails with rigorously validated stochastic models to ensure that methodological complexity yields reproducible gains in patient safety.
