Machine Learning-based Prediction of Registry-defined Trial Completion Status in Mesenchymal Stem Cell and Extracellular Vesicle Therapies: An Evidence-mapping Study Using ClinicalTrials.gov
Indira Azamatova
A
BSTRACT
Background:
Regenerative medicine remains limited by a persistent gap between early clinical investigation and progression to late-stage confirmatory trials. Mesenchymal stem/stromal cell (MSC) therapies are widely investigated, while extracellular vesicle (EV)-based approaches have emerged as potential cell-free therapeutic platforms. However, registry-level factors associated with completion or discontinuation of MSC- and EV-related clinical trials remain insufficiently characterized.
Objective:
To systematically characterize the global clinical trial landscape of MSC- and EV-related interventional studies registered in ClinicalTrials.gov and to identify trial-level factors associated with registry-defined trial completion status using machine learning (ML)-based approaches.
Materials and Methods:
A retrospective, registry-based observational study was conducted using ClinicalTrials.gov data. ClinicalTrials.gov was queried, and the dataset was downloaded on April 17, 2026 using predefined MSC- and EV-related search terms. Interventional trials involving MSC-based, EV-based, or combined MSC + EV therapeutic approaches were included. The evidence-mapping dataset included 1105 unique trial records. The primary modeling endpoint was registry-defined trial completion status. Trials listed as completed were classified as completed, whereas trials listed as terminated, withdrawn, or suspended were classified as discontinued. Trials with non-final registry statuses were retained for descriptive evidence mapping but excluded from supervised modeling. To reduce target leakage, study status, recruitment status, completion/discontinuation labels, completion date, termination-related information, and other status-derived variables were excluded from model inputs. Logistic Regression, Random Forest, and XGBoost classifiers were evaluated using receiver operating characteristic-area under the curve (ROC-AUC), precision-recall (PR)-AUC, accuracy, precision, recall, F1 score, Brier score, confusion matrices, and bootstrap-derived 95% confidence intervals (CIs).
Results:
The evidence-mapping dataset included 1105 trials with a median planned enrollment of 20 participants (interquartile range, 10–50), reflecting a predominance of small and early-phase studies. Final registry status was available for 707 trials, including 531 completed trials and 176 discontinued trials. For conservative supervised modeling, 630 clearly classified MSC-, EV-, or combined MSC + EV final-status trials were included after exclusion of other/unclassified therapeutic-platform records. MSC-based interventions accounted for 825 trials (74.7%), EV-only interventions for 36 trials (3.3%), and combined MSC + EV approaches for 101 trials (9.1%). Random Forest achieved the highest ROC-AUC (0.887; 95% CI: 0.794–0.961) and PR-AUC (0.940; 95% CI: 0.880–0.985), whereas XGBoost achieved the highest accuracy (0.913; 95% CI: 0.857–0.960), precision (0.919; 95% CI: 0.861–0.970), and F1 score (0.943; 95% CI, 0.905–0.974). Planned enrollment size was the strongest predictor associated with registry-defined completion status. EV-only trials showed a numerically higher observed completion proportion (14/16, 87.5%; 95% CI: 64.0–96.5), but this finding should be interpreted cautiously because of the small final-status subgroup.
Conclusion:
MSC- and EV-related clinical trials remain concentrated in early development, with limited progression to late-phase confirmatory studies and continued predominance of MSC-based platforms. ML analysis of registry data may help identify trial-level factors associated with completion versus discontinuation and may support hypothesis generation for future regenerative medicine trial design. However, registry-defined completion status is an operational endpoint and should not be interpreted as evidence of clinical efficacy, safety, regulatory approval, publication impact, or patient benefit.