Diagnostic models that combine a CT scan with a patient's health record are now more accurate than either input alone. They are still hard to trust, and the reason is specific: we do not know which parts of their reasoning will hold at the next hospital. A pulmonary embolism (PE) model trained at one site learns two kinds of regularity at once. Some are physiology: a clot in a pulmonary artery strains the right ventricle and lowers blood oxygen. Others are local habit: a D-dimer blood test is ordered mainly when the local triage policy already suspects PE. Current multimodal models squeeze both kinds of regularity into one feature vector or one similarity graph, so when the model is moved, nobody can tell which kind broke.
We propose an interpretable PE model whose reasoning runs through a small set of named clinical findings, called a concept bottleneck. The links between findings are kept as separate, typed layers: anatomical, physiological, risk-factor, treatment, and test-ordering/measurement. They are not merged into one graph. The claim is that keeping relation types separate does more than tidy up the explanation. It makes the model's failure under external validation diagnosable. When the model is evaluated, locked, at an external hospital, one can measure which relation layer changed. We predict the change will concentrate in the ordering/measurement layer while the physiological layer stays stable. We also predict that a model pre-committed to rely less on that layer will transport better than both a flat concept model and a black-box fusion model.
Core Claim and What Is New
The closest existing work has each piece but not the combination. PathMoE fuses slides, reports and cell graphs and validates externally on TCGA. Its graph has a single edge type, though, and its interpretability is limited to gating weights over modalities. Graph Concept Bottleneck Models adds a learned graph between concepts, but with one homogeneous edge type, image-only input, and no cross-site test. Relational Concept Bottleneck Models supports typed relations between concepts but was tested on puzzles and citation graphs. PECon is the strongest CT-plus-records PE model, but it explains itself only with heatmaps and was evaluated only on internal splits.
What is new here is treating explanation transport as the quantity being validated, and using relation-typed structure as the instrument that makes it measurable. Nobody in the file has asked whether a concept graph's edges, as opposed to its output score, stay the same across hospitals. Gap 3 in the research file states this directly.
Two prior lines motivate the design. Work on un-flattening networks (On the separability of aggregated multilayer networks) shows that once layers are summed, recovering which edge came from which layer becomes unreliable when layers overlap heavily. A merged clinical concept graph would therefore hide the very distinction a validator needs. MuSDAC shows that treating different relation types as separate channels during domain transfer avoids washing out discriminative signal. It reports 0.781 average cross-domain accuracy against 0.715 for HAN on citation networks. We borrow its channel-separation idea but apply it to interpretable concepts in a clinical setting, not to black-box node classification.
Research Program
The model has three stages. First, image encoders predict imaging findings and record encoders predict clinical findings. The imaging encoder starts from a pretrained PENet (PENet). Imaging findings include filling defect by location (central, lobar, segmental), right-ventricular strain, infarct and effusion. Clinical findings include tachycardia, hypoxemia, elevated D-dimer, active malignancy, recent surgery, prior thrombosis and current anticoagulation. Separately, the record encoder emits measurement-status concepts such as "D-dimer was ordered" and "troponin was ordered." These are treated as findings, not hidden inside a missing-value imputation.
Second, the concept vector $c \in [0,1]^K$ over $K$ findings is refined by message passing over $R$ typed adjacency layers $A_r$. Each layer is seeded from clinical knowledge and given learned, sparsity-penalized weights. The diagnosis is a linear readout:
$$\hat{y} = \sigma\Big(w_0^\top c + \sum_{r=1}^{R} \alpha_r\, w_r^\top (A_r c)\Big),$$
Here $w_0$ weights findings directly, $w_r$ weights the relation-$r$ evidence each finding receives from its neighbours, and $\alpha_r \ge 0$ is a per-relation gate. Because the readout is linear in $c$ and in each relation channel, the logit splits exactly into a direct part plus one contribution per relation type. That split is the explanation we validate.
Third, the gate on the measurement/ordering layer, $\alpha_{\text{ord}}$, is fixed before seeing any external data. It is chosen by a rule decided in advance on internal data alone: the smallest $\alpha_{\text{ord}}$ that costs at most one AUROC point internally. This keeps the external test locked, following the derive, lock, then validate discipline described in the prostate radiomics review.
Two leakage risks are handled explicitly. Concept scores are scalar probabilities rather than embeddings, which addresses the leakage concern raised for concept bottlenecks. A residual "unexplained" channel is also allowed, but its share of the logit is reported as a measured leakage budget instead of being hidden.
For the external audit, we compute three quantities per relation type $r$ at the derivation site and the external site. The first is the shift in the concept co-occurrence statistics that layer $r$ relies on. The second is the shift in that channel's average logit contribution. The third is the performance lost when channel $r$ is zeroed. The central prediction is a contrast. Physiological and anatomical channels should keep their contributions under shift, measured by rank correlation of per-patient contributions across matched subgroups. The ordering channel should not. Merged graphs cannot produce this contrast at all, and the separability results suggest one cannot reliably reconstruct it after the fact.
Data and Its Honest Limits
Derivation uses RadFusion (RadFusion): 1,837 CT studies with 12-month EHR summaries from Stanford.
The binding constraint is external multimodal data. PENet's Intermountain cohort of 200 studies is image-only as described in the file, and RadFusion names external extension as an open problem. The project therefore depends on obtaining linked records for an external CT cohort. Reconstructing the Intermountain EHR link is the natural first target. A second external hospital with a different PE testing protocol would be needed for the ordering-layer hypothesis to be tested rather than assumed.
Imaging-finding labels will be drawn from radiology reports where available, plus a radiologist-annotated subset. Clot-location concepts are additionally checked against CAD-PE. CAD-PE contains 91 PE-positive scans from six Madrid hospitals on Siemens scanners, with embolus-level ground truth, so it serves as an external test of the imaging concepts only.
As a second, cheaper testbed, we use the pediatric brain tumor (PBT) and TCGA pair from PathMoE. We replace its single-relation cell graph with cell-type-typed edges, which allows a check of whether the same "typed layers localize shift" behaviour appears in pathology, where an external cohort already exists.
Baselines
The naive reference point is RadFusion's late fusion: an image model and an EHR model trained separately, with their probabilities averaged. It reported 89.0% test accuracy. We rerun it on our splits and also lock it for external testing.
The strongest existing method is PECon: contrastive alignment of PENet CT features with an EHR encoder, explained with Grad-CAM. It reports F1 0.913, accuracy 0.90 and AUROC 0.943 on RadFusion, the best number in the file for this task. We retrain it on identical splits with the same PENet initialisation and the same hyperparameter search budget, measured in GPU-hours and number of configurations. We then evaluate it locked at the external site. For an image-only external anchor, PENet reports AUROC 0.84 internal and 0.85 at Intermountain.
There is no reported result found on Lacuna for any concept bottleneck model, flat or graph-based, on RadFusion or an external PE cohort. We therefore build two bottleneck comparators ourselves with the same concept set and encoders. The first is a flat bottleneck with independent concepts. The second is a Graph CBM-style single merged graph, constructed by summing our $R$ layers.
The ablation that isolates the key ingredient is the identical model with all typed layers summed into one. Its best-tuned gate on the merged layer plays the role of $\alpha_{\text{ord}}$.
We do not expect the proposed model to beat PECon internally. Matching it within about two AUROC points would be acceptable. Beating PECon means a smaller internal-to-external AUROC drop with overlapping or better external AUROC, at matched calibration. Beating the merged-graph ablation means two things: a statistically larger external gain from the pre-registered ordering gate, and a correct a priori prediction of which channel shifts.
Why not just…?
Why not just keep PECon and inspect its Grad-CAM maps at the new hospital? Heatmaps show where in the CT the model looked. The suspected failure is in the record side, in how test-ordering patterns combine with image evidence, and a CT heatmap has nothing to say about that. Heatmaps also offer no way to intervene. Graph CBMs report clear accuracy gains from correcting concepts at test time, especially at low intervention ratios, and heatmaps cannot support that. Still, PECon may simply transport well. If its external drop is as small as PENet's was, the accuracy case for our model weakens to an interpretability-only case, and we will say so.
Why not just a flat concept bottleneck, checking each concept's distribution at the new site? This is the most serious threat. Per-concept checks catch shifts in how often each finding occurs. They miss shifts in how findings co-occur, for example D-dimer being ordered for everyone at one hospital and only for suspected cases at another, even when the positive rate looks similar. The experiment settles this directly. If per-concept shift scores predict external degradation as well as per-relation scores do, and the flat model transports as well, then the typed structure adds nothing beyond bookkeeping.
Why not just delete missingness and ordering features from the records model? Ordering is genuinely informative: a clinician's suspicion is real diagnostic evidence. Deleting it may cost internal accuracy that does not need to be lost at sites with similar triage. We include this as a baseline. If deleting the features matches our gated model externally and loses no more than it internally, the gate is unnecessary. In that case, the remaining contribution is the audit, not the model.
Proposed Experiments and Decisive Checks
The first check runs before any clinical data. We simulate PE-like patients with known physiological links plus a site-dependent ordering policy, varying how strongly the layers overlap in the way the separability benchmark varies its overlap parameter. This tests whether the per-relation audit identifies the shifted layer and whether merged-graph models can recover it. If identification fails at realistic overlap, the clinical study is redesigned.
The second check is the internal comparison on RadFusion. We report AUROC, F1, calibration error, concept-prediction AUROC and the leakage budget.
The third and decisive check is locked external evaluation. We report the external AUROC drop for every model and the per-channel shift and contribution stability. We also report the gain from test-time concept correction at the external site, and fairness as true-positive-rate gaps, comparing against RadFusion's 0.059 largest gap for fusion.
The idea is falsified if any of three things holds externally. The ordering channel does not shift more than the physiological channels. The pre-committed gate does not reduce the external drop relative to the merged ablation. Or the simple flat-concept or feature-deletion baselines match it. On CAD-PE, clot-location concepts should localize emboli without any retraining. On PBT and TCGA, typed cell-graph layers should show the same localization of shift. A replication there would indicate that the finding is about relational explanations under site change, not only about PE.
References
- PathMoE: Interpretable Multimodal Interaction Experts for Pediatric Brain Tumor Classification
- Graph Concept Bottleneck Models
- Relational Concept Bottleneck Models
- PECon: Contrastive Pretraining to Enhance Feature Alignment Between CT and EHR Data for Improved Pulmonary Embolism Diagnosis
- On the separability of aggregated multilayer networks
- Domain Adaptive Classification on Heterogeneous Information Networks
- PENet - a scalable deep-learning model for automated diagnosis of pulmonary embolism using volumetric CT imaging
- Radiomics in prostate cancer imaging for a personalized treatment approach - current aspects of methodology and a systematic review on validated studies
- RadFusion: Benchmarking Performance and Fairness for Multimodal Pulmonary Embolism Detection from CT and EHR
- Computer Aided Detection for Pulmonary Embolism Challenge (CAD-PE)