A preprint study of a medical AI model that combines chest CT scans with selected electronic health-record data found a small edge over an imaging-only comparator in classifying ILD. The model’s mean AUC, the score used here to compare classification performance, was 0.8750±0.0443, versus 0.8646±0.0467 for imaging-only REN. That is an absolute difference of 0.0104. But the paired comparison across five patient-level folds was not statistically significant, with t(4)=0.265 and p=0.804. The comparison with another imaging model, SwinUNETR, was significant at p<0.001. The result is therefore a performance signal, not a decisive win over REN.
How the model splits the evidence
The analysis drew on a retrospective registry cohort containing 597 patients and 1,898 longitudinal chest CT scans collected from 2001 to 2023. ILD was confirmed in 365 patients, or 61.1%. Of the patients, 489, or 81.9%, were female, and the mean age was 63.7 years with a standard deviation of 12.7 years. Those figures describe the population on which the reported internal test was run.
At the center of the design is a hierarchical multimodal mixture-of-experts model, or MoE. It has two levels of gating, a routing mechanism that gives different specialist outputs more or less weight for a prediction. The imaging branch uses five lobe-specific experts built on 3D SwinUNETR backbones. Lobe masks and fixed radiomics-derived lobe-importance weights are used to aggregate the lobe signals. On the record side, EHR variables are organized into groups within the hierarchy.
The smaller record set scored higher
The selective EHR setup used three groups containing 12 variables, covering pulmonary function, disease-specific demographic information and biomarkers. The full hierarchy used seven groups containing 69 variables, and its mean AUC was 0.8496, below the selective model’s 0.8750. The comparison does not show that a smaller variable set would always produce a higher score. It shows an association within this tested arrangement.
The complete hierarchical model had the highest mean AUC among the evaluated fusion strategies. A simple concatenation of imaging and EHR inputs followed by logistic regression had AUC 0.8595±0.0438. Concatenation followed by an MLP scored 0.8337±0.0551, while EHR-only logistic regression scored 0.8280±0.0517. These comparisons favor the complete hierarchy within this experiment, but they did not independently test the modality gate and the EHR sub-gate. The contribution of each gate therefore remains unresolved.
Confidence did not always follow correctness
Timing of the record data made little difference in the test. Across forward EHR linkage windows of 0, 7, 14, 30, 60 and 90 days, AUCs were 0.8746, 0.8746, 0.8749, 0.8750, 0.8750 and 0.8750, respectively, with standard deviations of 0.0443, 0.0443, 0.0442, 0.0443, 0.0443 and 0.0441. The reported changes were negligible. This timing check was not a substitute for separate external validation, which was not reported.
Repeated scans also exposed variation in individual predictions. Among 192 patients with more than one CT scan, the mean within-patient prediction standard deviation was 0.0812±0.0964, while the mean range between the lowest and highest predicted probabilities was 0.2004±0.2406. In the subgroup breakdown, 138 ILD-positive patients had an average prediction standard deviation of 0.0756±0.0922 and range of 0.1901±0.2344. The 54 ILD-negative patients had higher reported averages, with standard deviation 0.0953±0.1060 and range 0.2265±0.2562. The analysis was restricted to patients with repeated scans.
The architecture also records how it routes information. Average modality weights were nearly even, with imaging at 0.498±0.122 and EHR at 0.502±0.122. At the prediction level, imaging gate weights were higher for ILD-negative predictions, while EHR gate weights were higher for ILD-positive predictions. Those activations describe model routing, not causal feature importance. A larger gate weight does not by itself establish independent clinical importance for a particular feature or record group.
The model was not error-free. Across five folds, expected calibration error, a measure of how closely confidence tracked correctness, was 0.131±0.038, while the Brier score, another calibration measure, was 0.135±0.037. Of 960 test samples, 167, or 17.4%, were misclassified. Seventy-one of those errors, 42.5%, occurred at confidence of at least 0.8. Misclassified cases had a lower mean EHR gate weight than correct predictions, 0.383 versus 0.527, and false positives were more imaging-driven.
A promising internal test, with clear gaps
The evaluation kept all longitudinal scans from each patient in the same fold, used mean±SD AUC reporting and applied two-sided paired t-tests to identical patient partitions. Training proceeded in three stages, with the imaging expert frozen during joint multimodal optimization. This patient-level setup was designed to keep repeated scans from the same person together during evaluation.
The data pipeline also carried important caveats. EHR variables were linked using the latest prior measurement, with a 30-day subsequent window allowed when needed. Missing values were then replaced with zero and standardized using training-partition statistics, without explicit missingness indicators. The authors say the numerical improvement over imaging-only REN should be confirmed in a larger evaluation, and no separate external validation was reported.
The findings support conclusions about discrimination, calibration and routing behavior in this dataset, not about patient outcomes or clinical benefit. The model’s higher average score over REN was not statistically significant, and the routing weights do not establish causal clinical importance. The main unanswered question is whether the same performance pattern holds beyond the evaluated registry cohort.
Paper data and sources
Original title: Hierarchical MoE for Multi-Modal ILD Diagnosis
Authors: Alec K. Peltekian, Gorkem Durak, Halil Ertugrul Aktas et al.
Journal/Repository: MICCAI Machine Learning in Medical Imaging (MLMI 2026)
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text