Concordance-Based Validation of Electronic Health Records and Modality Log Files to Improve MRI Exam Duration Prediction and Scheduling Performance.
Authors
Affiliations (5)
Affiliations (5)
- Department of Industrial & Systems Engineering, University of Washington, Seattle, USA. [email protected].
- Department of Industrial & Systems Engineering, University of Washington, Seattle, USA.
- Department of Radiology, University of Washington, Seattle, USA.
- Department of Radiology, University of Alabama at Birmingham, Birmingham, USA.
- Philips Healthcare, Bothell, USA.
Abstract
Integrating multi-source healthcare data for predictive modeling requires rigorous data quality validation, yet concordance between data systems is rarely assessed prior to model development.This study evaluated inter-system concordance between magnetic resonance imaging (MRI) exam duration data derived from modality log files (MLFs) and electronic health records (EHRs), developed a concordance-based data cleaning framework, and determined whether machine learning models trained on validated multi-source data improve MRI scheduling accuracy compared with template-based scheduling. Exam duration data from February 2022 through February 2024 were extracted from MLFs and EHRs. After fuzzy merging, concordance was assessed using Bland-Altman analysis and concordance correlation coefficients. Outliers were removed based on inter-system agreement thresholds. A Random Forest regression model was trained on the cleaned dataset to predict exam durations and compared with current template-based scheduling across MRI procedure codes. A total of 52,112 records were extracted from MLFs and 46,570 from EHRs. After exclusions and fuzzy merging, 30,275 records were retained; restricting to procedure codes with ≥ 400 records yielded 22,737 records, and 16,297 remained after concordance-based outlier removal. Bland-Altman analysis revealed discordance between MLF and EHR duration measurements, and the concordance-based filtering improved the concordance correlation coefficient from 0.33 to 0.87. The Random Forest model outperformed template-based scheduling for 11 of 12 procedure codes, with mean absolute error reductions ranging from 2.0% to 57.0%. For high-variability procedures, the proportion of exams completed within ± 10 min of the scheduled duration increased from as low as 29% to over 79%. These findings demonstrate that concordance-based validation is critical when integrating multi-source healthcare data, and that machine learning models trained on validated data substantially improved MRI scheduling accuracy, particularly for procedures with high intrinsic variability.