Patterns of CDSS adoption in primary care: a cluster analysis and predictive modelling study from a stepped wedge trial

Abstract Background Despite the potential of eHealth solutions to enhance medication management and patient safety, the integration of clinical decision support systems (CDSSs) into primary care remains challenging for healthcare systems worldwide. In Germany, the Digital Healthcare Act created a legal framework to accelerate eHealth adoption, but implementation lags behind international standards. We reanalysed data from a completed cluster-randomised stepped wedge trial (SW-CRT) that implemented a CDSS for medication management in German primary care to identify and characterise distinct implementation patterns and their predictive factors. Methods We linked routine health insurance records, practice structural data, pseudonymised CDSS logbook entries, and cross-sectional postal survey data from general practitioners (GPs) participating in the SW-CRT ( n = 736 practices). We used hierarchical cluster analysis (Ward’s method) on five implementation outcomes to identify distinct adoption patterns. Random Forest models were developed to assess how well structural, patient-level, and attitudinal variables could classify practices into these patterns. Results Of 736 participating practices, 356 (48%) performed at least one medication review. Hierarchical cluster analysis of these practices based on five implementation outcomes identified three distinct adoption patterns. The remaining 380 practices that did not perform medication reviews constituted a fourth pattern. CDSS usage intensity did not align with cluster-specific intervention effect across patterns. The pattern with the lowest usage intensity and fidelity showed the largest cluster-specific intervention effect on the combined endpoint of hospitalisation and mortality. Practices in this pattern reported significantly higher change commitment, change efficacy, and cognitive participation. Random Forest models using structural variables alone showed limited discrimination (AUC 0.56–0.66). Including a binary indicator of GP survey participation improved discrimination (AUC 0.61–0.82). Conclusions Within this single trial context, higher CDSS usage intensity did not correspond to larger cluster-specific intervention effects, and adoption behaviour was heterogeneous across practices. Structural variables alone were insufficient to distinguish adoption patterns; differences were instead associated with attitudinal factors such as change commitment, change efficacy, and willingness to engage with the intervention. Because these attitudinal measures were collected after practices had reached intervention status, they cannot be interpreted as antecedents of adoption. These findings nonetheless underscore the value of assessing implementer engagement during implementation and of tailoring implementation strategies to distinct adoption patterns rather than pursuing uniform approaches. Trial registration AdAM: ClinicalTrials.gov (NCT03430336), 6 February 2018; eHealth COMPATH: Open Science Framework (osf.io/gau5w), 29 December 2023.