
Abstract Objectives Hospital rapid diagnostic units (RDUs) assess outpatients with suspected serious disease, yet no validated tool estimates malignancy risk at the first RDU visit. We developed and internally validated UDR-PREDICT, a logistic prediction model with a derived integer score. Methods Retrospective analysis of a prospectively maintained registry of first RDU visits (2022–2025) in a Spanish tertiary hospital. The outcome was a final diagnosis of solid or haematologic malignancy. Candidate predictors available at referral were selected by least absolute shrinkage and selection operator (LASSO), refitted by logistic regression and shrunk. An integer score was derived with the Framingham points method. Both were validated by bootstrapping the full procedure and by internal–external cross-validation across calendar periods; 2025 served as an exploratory temporal test. Results Of 975 analysable patients, 147 (15.1 %) had malignancy. The model retained age, sex, abdominal mass, lymphadenopathy, pulmonary nodule and bone lesions. Optimism-corrected area under the curve (AUC) was 0.719 for the model and 0.721 for the score (calibration slopes 0.857 and 0.859). Across periods the pooled AUC was 0.693 (95 % CI 0.641–0.741) and the pooled calibration slope 0.737 (0.550–0.924). In 2025 (19 events) the AUC was 0.731 (0.618–0.826). Predicted risk rose from 3.9 % at 0 points to 64.9 % at 14 points; a score ≥9 identified 9.9 % of patients, 48.5 % of whom had malignancy. Conclusions UDR-PREDICT showed modest discrimination and some overfitting. It may help order first-visit priorities but requires external validation and recalibration before clinical use.
Abstract Objectives Diagnostic errors arise through complex interactions between diagnostic process failures, contextual factors, and cognitive pitfalls. However, most previous studies examined these factors individually, providing limited insights into how diagnostic breakdowns co-occur in clinically challenging cases. We aimed to identify recurrent diagnostic failure modes using a hierarchical cluster analysis of published case reports of diagnostic errors. Methods We performed a secondary analysis of previously published systematic reviews of diagnostic-error case reports. Diagnostic process failures were coded using the diagnostic error evaluation and research (DEER) taxonomy. Hierarchical clustering using Ward’s method and the Euclidean distance was applied to identify the patterns of co-occurring DEER failure points. The Caliński–Harabasz pseudo-F index was used to evaluate cluster structures. Reliable Diagnosis Challenges and Generic Diagnostic Pitfalls frameworks were used to characterize the contextual and cognitive contributors within clusters. Results A total of 563 cases of diagnostic errors were analyzed. Cluster analysis revealed two main categories of failure modes: (1) test ordering and diagnostic prioritization failure, in which mimicking presentations were more common, and (2) testing interpretation and diagnostic synthesis failure. The latter category was further divided into three clinically relevant subtypes: laboratory/radiological reading failure, test interpretation failure, and diagnostic synthesis failure despite preserved information acquisition. Conclusions This study identified exploratory patterns of co-occurring diagnostic process failures in published case reports. These patterns require prospective validation in independent multicenter clinical datasets before they can be used to guide diagnostic safety interventions.