This study develops and tests a rule-learning explainable AI (RL-XAI) methodology for optimizing complex operational processes, illustrated in Magnetic Resonance Imaging (MRI) workflows. Researchers applied globally-optimal Boolean rule-learning models to device scanner logs from 3,994 contrast-enhanced liver and prostate MRI exams performed on five 3T scanners at three outpatient facilities between January 2019 and January 2024. Exams longer than 25 minutes were labeled as requiring optimization (35.4% of liver and 52.6% of prostate cases). The AI discovered concise, interpretable rules from rich scanner log features; the top rule informed protocol changes that reduced median liver MRI time by 10.9% and lowered the fraction of long liver exams from 35.4% to 23.9% (p < 0.001). Prostate protocol changes supported addition of a new sequence to improve exam quality. The authors conclude that globally-optimal, multi-model RL-XAI can translate feature-rich scanner logs into operationally meaningful rules that enable measurable workflow improvements when paired with domain expertise.
Modern clinical workflows, particularly in radiology, are driven by complex technology and constrained resources, making manual optimization impractical. MRI scan duration depends on many interacting factors — clinical indications, patient characteristics, acquisition types, staffing, scanner settings, and scheduling — which cannot be optimized using patient records alone. Prior work has begun to explore explainable AI methods and the use of device log data for process understanding, but scanning log utilization has often been limited to basic timing information. Proprietary barriers and the scale of log data have constrained broader use of scanner logs to extract process rules. This study aims to (a) establish practical criteria for applying RL-XAI to complex process optimization and (b) show how RL-XAI applied to detailed device logs can enable measurable process improvements.
The project followed two stages: (1) application of RL-XAI models to detect significant, interpretable patterns of processing delays and (2) implementation of model-derived findings into scan protocols with subsequent verification in a clinical setting.
Data and labeling
The retrospective study used operational MRI scanner logs for contrast-enhanced liver and prostate exams. Logs provided detailed timestamps for all events (e.g., table movements, acquisition events, user interface interactions). Exams with duration longer than 25 minutes were labeled as requiring optimization. The dataset included 3,994 exams collected from five 3T scanners at three outpatient facilities over January 2019–January 2024.
Modeling criteria
To ensure practical applicability, the authors specified three methodological criteria:
Feature construction and modeling
Scanner log events were aggregated at the exam level to create multiple features representing measurement counts, timing characteristics, and user interactions. Boolean rule-learning algorithms that search for globally optimal rules were applied to discover interpretable models predicting the long-exam label. An initial ranking used the F1 score to identify models with strong predictive performance; subsequently, an operational-impact metric estimated each rule’s potential change in average scan duration to prioritize actionable rules.
Selection and evaluation
The RL-XAI pipeline produced the top 1,000 rules by F1 score. A secondary selection based on estimated operational impact reduced this to 20 top rules for clinical review. Selected rules were translated into protocol changes in conjunction with domain experts. The effect of protocol modifications on average and median exam durations was tested using permutation tests to assess statistical significance.
Rule discovery and selection
From the scanner logs, RL-XAI discovered N = 1,000 top rules by F1 score that identified significant processing delay patterns. Using an operational-impact metric, the team selected 20 high-impact rules for domain review and implementation.
Protocol changes and measured impact
Implementation of the highest-priority rule in liver MRI protocols led to a 10.9% reduction in median scan time. The proportion of liver exams exceeding the 25-minute threshold decreased from 35.4% before intervention to 23.9% afterward; this change was statistically significant (p < 0.001) based on permutation testing. For prostate MRI, rules guided protocol alterations that included adding a new scanning sequence intended to improve image quality; the article reports protocol-driven quality improvements but does not report a specific quantitative time reduction for prostate exams.
Operational insights
The RL-XAI approach converted high-dimensional scanner log data into a compact set of interpretable Boolean rules that highlighted specific device interactions and timing conditions associated with prolonged exams. Combining these rules with clinical expertise enabled targeted protocol modifications that produced measurable and sustainable improvements.
The authors argue that the three modeling criteria — globally-optimal rule search, multi-model output, and process-driven metrics — are essential for translating explainable AI outputs into operational actions. Globally-optimal models avoid pitfalls of greedy algorithms and ensure the most significant delay patterns are detected. Producing multiple top models allows teams to choose rules that are actionable within local operational constraints. Process-focused evaluation aligns model selection with real-world goals, such as reducing median exam duration or lowering the proportion of long exams.
Limitations and governance
The study used retrospective operational data and was exempt from IRB oversight under protocol 2022P002693. Individual-level data cannot be publicly shared because it contains or can be linked to protected health information; qualified researchers may seek access through the institution subject to approvals. Two coauthors had industry affiliations limited to supplying and interpreting scanner log data and did not influence model development or conclusions. The authors reported no specific funding and declared no competing interests.
Globally-optimal, multi-model rule-learning explainable AI applied to detailed scanner logs can yield concise, interpretable, and operationally meaningful rules. When combined with domain expertise and applied to MRI protocols, these rules supported measurable reductions in scan times and protocol-driven improvements in image quality. The authors present this RL-XAI framework as a general approach for process optimization in complex, feature-rich operational environments such as medical imaging.