---
title: "Rule-learning explainable AI for MRI process optimization: global rule models from scanner logs"
id: "plos-one-20-process-optimization-with-rule-learning-explainable-ai-and-its-application-to"
canonical_url: "https://medichelpline.com/clinical-feed/plos-one-20-process-optimization-with-rule-learning-explainable-ai-and-its-application-to"
content_type: "clinical_feed_article"
specialty: "Radiology"
source_name: "PLOS ONE (Medicine)"
source_url: "https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0358070"
published_at: "2026-09-16T14:00:00.000Z"
evidence_level: "Journal Feed"
license: "CC-BY-NC-4.0 / Informational Use"
---
# Rule-learning explainable AI for MRI process optimization: global rule models from scanner logs
## Provenance & Clinical Metadata
- **Canonical URL:** https://medichelpline.com/clinical-feed/plos-one-20-process-optimization-with-rule-learning-explainable-ai-and-its-application-to
- **Specialty:** [Radiology](https://medichelpline.com/clinical-feed/radiology.md)
- **Primary Source:** PLOS ONE (Medicine)
- **Source URL:** [Original Journal Publication](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0358070)
- **Published At:** 2026-09-16T14:00:00.000Z
- **Evidence Rating:** Journal Feed
## Executive GIST (TL;DR)
- The study demonstrates a practical methodology using **rule-learning explainable AI** (RL-XAI) to discover concise, interpretable process optimization rules from rich device log data. - Researchers applied RL-XAI to 3,994 contrast-enhanced liver and prostate MRI exams performed on five 3T scanners across three outpatient sites between January 2019 and January 2024. - Exams longer than 25 minutes were labeled as requiring optimization; this applied to 35.4% of liver and 52.6% of prostate cases in the dataset. - Modeling criteria prioritized: finding **globally-optimal** solutions for a fixed number of features, producing multiple top models (multi-model optimization), and using process-driven evaluation metrics suited to operational goals. - Scanner logs supplied granular timestamps and device interaction events (table movements, acquisition starts/ends, UI interactions), which were aggregated into features for RL-XAI modeling. - The RL-XAI pipeline discovered the top N = 1,000 rules by F1 score; a secondary operational-impact metric filtered these to 20 high-impact rules estimating each rule’s effect on average scan duration. - Changing liver MRI protocols guided by the top rule produced a 10.9% reduction in median scan time and decreased the proportion of long liver exams from 35.4% to 23.9% (p < 0.001) as measured by a permutation test. - Prostate MRI optimizations informed by the rules led to the introduction of a new scanning sequence intended to improve exam quality; details on quantitative time reduction for prostate were not reported. - The authors conclude that globally-optimal, multi-model RL-XAI can convert feature-rich **scanner logs** into actionable, human-interpretable rules that, combined with domain expertise, support measurable workflow improvements. - Data access restrictions apply because of protected health information; the study was exempt from IRB oversight under protocol 2022P002693 and covered data access dates from 10/01/2022 to 12/10/2025. - Funding: no specific funding. Competing interests: authors declared none.
## Clinical Analysis & Structured Key Points
Process optimization with rule-learning explainable AI, and its application to MRI scanning | PLOS One Browse Subject Areas ? Click through the PLOS taxonomy to find articles in your field. For more information about PLOS Subject Areas, click here . Article Authors Metrics Comments Media Coverage Reader Comments Figures Figures Abstract Purpose High complexity of modern processes renders manual process optimization impossible. The main goal of our work was to formalize and to demonstrate the practical efficiency of explainable rule-learning AI in complex process optimization. Materials and methods Process optimization methodology was based on globally-optimal Boolean rule-learning AI models, capable of identifying multiple inefficiency patterns, and supporting process-oriented optimization metrics. To illustrate this approach in real-word environment, the study used 3994 liver and prostate Magnetic Resonance Imaging (MRI) exams performed on five 3T scanners at three outpatient facilities from January 2019 to January 2024. Imaging exams longer than 25 minutes were labeled as requiring optimization, which applied to 35.4% of liver and 52.6% of prostate cases. Rule-learning AI was applied to the MRI scanner log data to discover short and interpretable process optimization rules. The selected Boolean rules were used to implement improved exam protocols, and a permutation test was used to measure the statistical significance of the resulting change in average exam duration. Results N=1000 top rules, identifying the most significant processing delay patterns, were discovered by rule-learning AI from the scanner log data based on F1 score. A smaller set of 20 top rules was selected using the secondary operational impact metric, an estimate of each rule’s impact on average scan duration. A change in liver MRI protocols based on findings from the top rule resulted in a 10.9% reduction of median scan time. The proportion of long liver exams was reduced from 35.4% to 23.9% (p < 0.001). Similar optimizations in prostate protocols were used to add a new scanning sequence, improving exam quality. Conclusions Globally optimal, multi-model rule-learning AI can transform feature-rich device logs into concise, interpretable, and operationally meaningful rules. When combined with domain expertise, this approach can support measurable and sustainable improvements in complex clinical workflows. Citation: Hartmann S, Sharp A, Johnston H, Gee MS, Huang SY, Harisinghani MG, et al. (2026) Process optimization with rule-learning explainable AI, and its application to MRI scanning. PLoS One 21(9): e0358070. https://doi.org/10.1371/journal.pone.0358070 Editor: Arvind Mahindru, DAV University, INDIA Received: February 4, 2026; Accepted: August 26, 2026; Published: September 16, 2026 Copyright: © 2026 Hartmann et al. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability: The individual-level data underlying this study cannot be shared publicly because it contains or can be linked to protected health information. These restrictions were imposed by the Mass General Brigham Institutional Review Board/Human Research Affairs under Protocol No. 2022P002693 and Mass General Brigham policies governing protected health information. Qualified researchers may contact Mass General Brigham IRB at IRB@mgb.org ; any access would require prior institutional and IRB approval. Funding: The author(s) received no specific funding for this work. Competing interests: The authors have declared that no competing interests exist. 1. Introduction Modern workflows, driven by complex technology and narrowing resource constraints, are becoming increasingly challenging for human decision-making and optimization. Medical imaging (radiology) exemplifies this challenge: to optimize the duration of patient scanning, one must consider clinical needs, patient characteristics, types of imaging acquisitions, staffing, scanner settings, scheduling constraints, and more. This degree of optimization cannot be achieved manually, and cannot be based on hospital patient records alone, requiring more process-specific data and analyses [ 1 , 2 ]. Previous work has recognized the need for better process optimization algorithms and more detailed process data. On the methodological side, this has given rise to a new line of research exploring the use of classical explainable AI (XAI) methods, ranging from linear regression to rule lists, to uncover basic, human-interpretable processing logic [ 3 – 9 ]. On the applied side, automatically generated, feature-rich process log data has attracted growing attention for its potential to support human understanding of complex processes, including applications in medical imaging [ 10 – 13 ]. However, this work has just started – particularly in healthcare, where only a few recent studies have suggested using imaging scanner logs to extract more accurate measurements of patient movement, scanner utilization, and scheduling. Even in these studies, scanner log use was limited to serving as a source of exam and measurement sequence start-end times [ 2 , 14 – 17 ]. Consequently, the use of scanner logs in radiology has remained limited, largely due to the proprietary barriers in acquiring primary scanner log data, and difficulties in processing the large quantities of information contained in the logs. As a result, leveraging the full set of device log features with explainable AI models to discover human-understandable process optimization rules still presents a compelling but unexplored opportunity for improving operational efficiency. Additionally, theoretical advances in this area must still be translated into real-world applications to demonstrate that XAI-derived insights can lead to tangible process improvements. Motivated by this challenge in our daily clinical work, we set two principal goals of our study as follows: a) to develop practical criteria for using rule-learning XAI (RL-XAI) in complex process optimization, and b) to demonstrate how this methodology can lead to successful process optimization when RL-XAI is applied to complex device log data. Beyond achieving specific optimization gains, we also sought to better understand the mechanisms through which rule-learning XAI supports process improvement. 2. Methods Our project consisted of two parts: first, we applied RL-XAI models to identify the most significant and interpretable patterns of processing delays, and second, we implemented model findings to improve scan efficiency, and to verify results in a real-world clinical setting. This retrospective study was conducted in a large academic medical center. The historical data collected for this study was strictly operational, therefore ethical review and approval were not required, and the study was exempt by our institution from IRB oversight in accordance with applicable regulations; the exemption was granted under protocol 2022P002693. The data was accessed for research purposes from 10/01/2022–12/10/2025. Two paper authors had industry affiliations, but were involved only in providing and interpreting the scanner log data, and did not impact study model development, analyses, or conclusions. 2.1. Modeling criteria Aiming at practical gains, we formulated the following criteria as the most essential for translating the XAI results into tangible process improvements: Globally-optimal solutions : Finding the very best model(s), subject to a fixed number of features. When optimizing complex processes with many interconnected variables, it is critical to identify the most significant patterns of processing inefficiencies. Classical XAI algorithms, relying on faster greedy methods, do not guarantee finding the best solutions, which may result in inaccurate process interpretations. Multi-model optimization : Finding the top N best models rather than a single model. For many problems, disparate explanatory models may provide similarly accurate predictions, but some may offer more actionable optimization targets than others. Most current XAI methods, including ensemble approaches, are designed to produce only a single final model. Process-driven evaluation metric: The metric used to determine the best model should be adaptable to the problem in question. Conventional loss functions and metrics, traditionally used to optimize AI models, may not be useful or intuitive for process management. Supporting process-specific metrics extends model optimization to a broader range of real-world problems. Together, these criteria define a general rule-learning optimization methodology that can be applied across various processes, and distinguish our approach from prior work. To illustrate their practical value, we used Magnetic Resonance Imaging (MRI) scan optimization, as one of the most complex medical imaging workflows, characterized by long and highly variable examination times [ 1 , 16 , 18 , 19 ]. 2.2 Operational data, target, and features This project used detailed MRI scanner logs, providing a granular and feature-rich representation of the imaging processes for contrast-enhanced liver and prostate exams – two common MRI scan types, associated with especially long and highly variable scan times. The scanner logs contained precise timestamps for all events performed during imaging exams, including scanner table movements, acquisition types, and user interface interactions ( Table 1 ). Download: PNG larger image TIFF original image Table 1. Properties of liver and prostate examinations used in process optimization analysis. Each exam may be defined by a set of selected imaging sequences and protocols. Each exam instance results in multiple scanner log events, including imaging measurements (events associated with MRI image acquisition), which are aggregated at the exam level to create variables for the RL-XAI models. https://doi.org/10.1371/journal.pone.0358070.t001 Because scanner logs record only device-specific information, we had to merge them with Electronic Health Record (EHR) data to establish clinical context, including examination and appointment-scheduling information. Since the two data sources shared no patient or examination identifiers, the scanner-to-EHR data mapping had to be done based on the scanner IDs and examination times, with each exam in the scanner log matched by timestamp to the nearest EHR exam performed on the same scanner. Not all exam records matched successfully: 28% of liver and 22% of prostate exams present in EHR were missing from the scanner logs, because the log data failed to save correctly during random time intervals due to networking issues. Additionally, 5.3% of exams labeled as abdominal in the scanner logs were not mapped to EHR records, because their scanner log timestamps did not overlap with EHR timestamps, reflecting random timing errors in either data source. In both cases, our investigation showed that missing records were uncorrelated with exam attributes. Therefore, the unmatched exams were excluded from our dataset, with no evidence that their exclusion introduced bias into our analysis ( Fig 1 ). Download: PNG larger image TIFF original image Fig 1. Study data and modeling flow. Scanner log data was joined with the EHR records, resulting in a large feature-rich dataset. This data was processed by the short RL-XAI engine, which identified the best most concise logical expressions, explaining scanning delays. https://doi.org/10.1371/journal.pone.0358070.g001 The resulting log-to-EHR data mapping was instrumental for binding scanner log events to the overall scanning scheduling and efficiency. In particular, all patient MRI appointments at our facilities were scheduled for 30-minute slots, including 5 minutes of patient preparation time. Consequently, examinations with scan times exceeding 25 minutes left insufficient time to remove the patient from the scanner and begin the next examination on schedule. Liver and prostate MRIs were among the most frequent exam types exceeding the expected 25-minute scan limit and were thus selected as the focus of our study. The study included 1857 liver and 2137 prostate exams performed on five 3T MRI scanners from a major scanner vendor, at three outpatient facilities, from January 2019 to January 2024. 35.4% (239/676) of liver exams and 52.6% (631/1199) of prostate exams from January 2019 to September 2022, before any interventions were implemented, were too long for their scheduled time slots. 20% of exams from this period, stratified by class, were randomly assigned to the test set, and separate models were trained and evaluated for liver and prostate. Data from 1482 liver exams performed on three other MRI scanners at the same outpatient facilities that did not have their protocols optimized during the study period was used for comparison. The mapping of device logs and EHR data also enabled us to collect and engineer 540 exam-level features, capturing as much process-related information as possible. Chosen in collaboration with the scanner engineering team, these features included counts of various MRI scanning events such as scanner table movements, coil changes, specific device alerts or processing failures; time-based information such as the day of the week and the number of exams performed on the scanner earlier that day; and exam-based information such as different exam types. The exam date was also included as a feature to identify whether differences in duration were more correlated with temporal trends than other exam properties. Each MRI scan included multiple imaging sequences (image acquisition steps). In order to reflect their interdependencies, we added not only their individual counts, but also groupings by two or three consecutive sequences, and their relative frequencies. 2.3 Model choice and feature extraction The principal goal of this project was to use multi-model explainable AI (XAI) to achieve the maximum possible process time reduction; therefore, we chose the globally-optimal rule-learning XAI (RL-XAI) approach of “human knowledge models” (HKM) developed in [ 20 ]. As suggested by vast volume of cognitive research, humans can understand and apply decision models with at most four (Boolean) variables, where each variable is used at most once [ 21 – 23 ]. Therefore, the HKM method employs logic-based XAI learners that find the most accurate combinations of short Boolean rules, where both features and feature value thresholds are discovered to achieve the best explanation of a binary outcome from a complex multidimensional dataset. This approach of combining short models with simple Boolean operators (“not”, “and”, and “or”) provides several key advantages essential for process optimization tasks. First, unlike conventional XAI approaches that produce numerical feature-importance rankings (SHAP, p-values, stepwise selection, model-specific importance algorithms), RL-XAI learns the best features together with the corresponding best rules, thereby capturing the best sets of optimal features with their actual optimal predictive logic . These simple and direct rules can be easily interpreted and evaluated by humans – compared to other XAI approaches such as regressions or Bayesian networks, which typically incorporate more complex sets of variables, math, and explanations. Second, while current feature importance techniques evaluate the contribution of individual features, RL-XAI is truly multivariate, identifying optimal combinations of features that operate jointly within a single model. Third, RL-XAI does not require feature preselection, which helps reduce concerns about multicollinearity: features included in the top RL-XAI models are unlikely to be redundant because these models outperform alternative models of the same size. Fourth, unlike conventional AI optimization which relies on suboptimal greedy algorithms, short RL-XAI learners enable globally-optimal optimization (computationally unfeasible for large models): considering all possible 2- or 3-feature subsets even from hundreds of original features may be done in a reasonable amount of time. Fifth, RL-XAI learners can accommodate application-specific loss functions, allowing optimization using the metric most relevant to the process being studied. Finally, exhaustive RL-XAI easily extends to a multi-model approach, producing the list of the first N best models instead of a single-model solution and thereby meeting our criteria for practical process optimization. Interestingly, past research has shown that many complex systems may be governed by a very small subset of their parameters [ 24 ]. This justifies the use of small RL-XAI models, for which exhaustive (globally-optimal) model search is attainable despite having polynomial time complexity. Note that although many common statistical techniques, such as principal component analysis, can significantly reduce the original data dimensionality, they do not preserve the original set of variables [ 25 ]. In operational improvements, optimizing within the original feature set is imperative to identify the key process drivers that can be targeted with interventions to improve workflow efficiency. 2.4 Statistical methods The second part of our study was to implement changes to protocols based on our findings from RL-XAI models. Changing scanner settings is a complex task which cannot be randomized in a busy outpatient clinic; therefore, we had to implement scanning protocol changes with a pre-post design. In collaboration with clinical and engineering experts, and using the top RL-XAI models identified by our globally-optimal search, we selected the best “unbottlenecking” logical rules that could be implemented without sacrificing diagnostic quality. We then compared the difference in exam durations before and after these protocol changes. During the time period of our analysis, scanner log data was available for 72% of liver exams and 78% of prostate exams performed at the institution, due to previously discussed networking errors. When evaluating the liver exam intervention, we analyzed data from 72% of a finite population of exams, so instead of conventional parametric statistical testing, we performed a two-sided two-sample permutation test to determine whether the reduction in average exam duration resulting from the intervention was statistically significant. The pre and post intervention group labels were reassigned 100,000 times using random permutations. The observed reduction in average exam durations was compared to the differences in average exam durations generated by these 100,000 permutations. Confidence intervals for model performance metrics on the train and test sets were computed using percentile bootstrapping with 10,000 samples. Statistical analysis was done using SciPy 1.10.1 (2023) and Python 3.8.19 (2024), and p-values below 0.05 were considered significant. 3 Results The application of RL-XAI models to MRI scanner log data led to several important results identifying and eliminating major scanning inefficiencies, and illustrating the importance of the model selection criteria stated above. 3.1 Optimizing operational impact metric The models selected as the best will depend on the choice of the optimization metric, and this choice should be operationally meaningful. Therefore, we used the following two metrics for evaluating each model’s quality on our MRI datasets: F1 score: a standard metric for measuring the quality of binary classification models in cases where classes may be imbalanced, and Operational impact: The product of the fraction of exams the model classifies as long and the average increase in duration among those exams. Operational impact represents the average time savings achieved for the entir
## Related Clinical Research

- [Reliability of ultrasound measurement of medial knee joint space width during valgus stress](https://medichelpline.com/clinical-feed/plos-one-4-relative-and-absolute-reliability-of-medial-knee-joint-space-width-measurement.md)
- [Effects of Retrospective Lipid Suppression on Metabolite Quantification in Preclinical Proton MRSI](https://medichelpline.com/clinical-feed/biorxiv-5-impacts-of-retrospective-lipid-suppression-on-metabolite-quantification-in.md)
- [SNR-efficient Tensor-Valued Diffusion Encoding Protocol for High-Resolution Brain Microstructure I](https://medichelpline.com/clinical-feed/medrxiv-12-developing-an-snr-efficient-tensor-valued-diffusion-encoding-protocol-for.md)
- [Two-Photon–Driven RESOLFT Microscopy Enables Super-Resolution Imaging Deep in Tissue](https://medichelpline.com/clinical-feed/biorxiv-13-fully-two-photon-driven-resolft-microscopy-for-super-resolution-imaging-inside.md)
- [PILR‑U‑Net: Physics‑Informed Latent‑Regularized 3D MRI Segmentation of Pathological Muscle](https://medichelpline.com/clinical-feed/medrxiv-18-3d-segmentation-of-pathological-muscle-with-a-physics-informed-latent.md)

## Navigation
- [← Back to Radiology Feed](https://medichelpline.com/clinical-feed/radiology.md)
- [← All Clinical Specialties](https://medichelpline.com/clinical-feed.md)
## Medical & Regulatory Disclaimer

> [!CAUTION]
> MedicHelpline content is structured for research, educational, and professional discovery purposes. It does not constitute individual medical advice, clinical diagnosis, or treatment recommendations.
> Always verify dosing, contraindications, and regulatory alerts against official product labeling and primary regulatory sources before clinical decision-making.