Confidence-Gated Cloud-Edge Cascade Triage via Variational Risk Minimization for Medical Imaging

• VRM treats missing reports as latent clinical interpretations and marginalizes LVLM-generated text samples into uncertainty-aware supervision.Rather than distilling from one report-conditioned teacher target, the student learns from a distribution of teacher predictions.Source: §1 · p. 1
• The compact deployment couples a 75MB, 14ms edge student (RTX 2060) with confidence-gated cloud escalation.The cascade provides explicit latency-accuracy control.Source: §1 · pp. 1–2
• Across matched encoder families, VRM consistently outperforms direct fine-tuning baselines, with gains confirmed by paired DeLong tests.The comparison uses direct fine-tuning baselines from the same encoder families.Source: §1 · p. 1
• Marginalization improves with larger K, reduces teacher-target variance and report-selection volatility, and supports calibration and hallucination recovery.Source: §1 · p. 1

2.1. Structural Modality Censorship in Medical AI

• Reports arrive after emergency triage, making joint image-text availability a workflow-driven mismatch rather than random missingness.Source: §2.1 · p. 2

2.2. Distillation Under Missing-Modality Constraints

• Unlike feature alignment, modality dropout, or deterministic imputation, VRM distills a variationally marginalized teacher distribution over LVLM-sampled reports.Source: §2.2 · p. 2

2.3. Uncertainty, Hallucination, and Reliability-Centered Deployment

• Edge-focused models address compute constraints but not supervision reliability; the paper couples marginalization with confidence-gated escalation for uncertainty handling.Source: §2.3 · p. 2

3.1. Problem Formulation

• At inference, the student predicts urgency from the chest X-ray alone, although training includes the image, report, and urgency label.The student is formulated to approximate the multimodal teacher’s marginalized predictive distribution.Source: §3.1 · p. 2

3.2. Latent Query Alignment (LQA)

• LQA uses learnable queries that first contextualize on text and then ground to image evidence, with repeated transformer-style interactions preserving clinical visual signals.The fused representation is intended to suppress background noise as well.Source: §3.2 · p. 2

3.3.1. Latent Clinical State Marginalization

• The framework treats the unobserved clinical state underlying each image and report as latent, using clinical text as a proxy for that state.An LVLM proposal distribution q(t|x) supplies plausible interpretations over which the predictive risk is marginalized.Source: §3.3.1 · pp. 2–3

3.3.2. Monte Carlo Approximation

• The teacher target is estimated by averaging K approximately independent LVLM samples; under the i.i.d. idealization, estimator variance decays as O(1/K).The estimator is unbiased with respect to q(t|x) as K approaches infinity.Source: §3.3.2 · p. 3

3.4. LVLM Sampling Implementation

• For diversity in Monte Carlo samples, MedGemma-27B is used with nucleus sampling: temperature 0.9, top-p 0.95, top-k 50, and repetition penalty 1.1.Source: §3.4 · p. 3

3.5. Distillation Objectives

• The multi-task VRM loss combines response-level distillation, intermediate-feature alignment, and weighted cross-entropy against ground-truth labels.Teacher embeddings are averaged across samples for the feature-alignment objective.Source: §3.5 · p. 4
• Soft-label distillation uses KL divergence to align the student’s predictive distribution with the teacher’s marginalized target.The target logits arise from marginalization over sampled reports; the alignment uses a temperature hyperparameter.Source: §3.5 · p. 4

3.6. Confidence-Gated Cascade Triage

• The cascade accepts cases at the edge only when confidence is high and uncertainty is low; otherwise, it escalates them to a cloud teacher for confirmation.Escalated cases may be de-identified at the edge or routed to a hospital secure private cloud, while the design allocates multimodal reasoning to ambiguous cases.Source: §3.6 · p. 4

4.1. Experimental Setup

• The experiments use MIMIC-CXR and four external cohorts, with a patient-level 8:1:1 train/validation/test split for MIMIC-CXR.Urgency labels use radiologist consensus; NLP resolves context-dependent findings by acute-versus-chronic indicators, with κ = 0.86 agreement against manual annotations on 200 cases.Source: §4.1 · p. 4

4.2. External Benchmark Reference

• 0.842 vs. 0.748 AUC: VRM closes the distillation gap over Standard KD, and direct fine-tuning baselines remain below VRM in this task setting.Paired DeLong tests found VRM significantly better than Standard KD, DINOv2-RadBench, CXR Foundation, and EVA-X Direct FT on MIMIC-CXR.Source: §4.2 · p. 5, Table 4 · p. 6

4.3. Clinical Utility Under Safety-Constrained Operation

• At fixed 90% sensitivity, VRM specificity is 65.2% versus 60.3% for EVA-X direct FT and 45.5% for Standard KD.The corresponding reductions in unnecessary urgent flags are 4.9 and 19.7 percentage points; VRM also improves NPV over Standard KD (0.924 vs 0.826).Source: §4.3 · p. 5, Table 6 · p. 6, Table 7 · p. 6

4.4. Cascade Triage Evaluation

• A confidence-gated cascade reaches AUC 0.941 on MIMIC and OOD average 0.892 at 103ms average latency, escalating 20.3% of difficult cases.Edge-only inference is 14ms; the cascade uses edge predictions for accepted cases and sends escalated cases to the cloud teacher.Source: §4.4 · p. 6, Table 8 · p. 7

4.5.1. E

• K=5 improves AUC by 6.2% over K=1, while teacher prediction variance falls from 0.034 to 0.012 as K increases.K=10 adds only 0.3% AUC beyond K=5; training time scales linearly with K, and K=5 is selected as the performance-cost trade-off.Source: §4.5.1 · p. 7, Figure 3 · p. 7

4.5.2. Component Analysis

• Adding feature distillation raises AUC from 0.725 with classification loss alone to 0.785, and the full VRM objective adds a further 5.7% absolute gain.The full objective includes KL divergence with marginalized samples, which the analysis identifies as the largest incremental gain.Source: §4.5.2 · p. 7, Table 9 · p. 7

4.6.1. Generated Text Quality; 4.6.3. Monte Carlo Sample Quality Examples

• 88% of LVLM-generated reports were rated clinically acceptable or better, with κ = 0.78 inter-rater agreement and BERTScore 0.731.On cases with hallucinated reports, recovery reaches ∼66% at K=5 and ∼80% at K=10; the example analysis notes that marginalizing samples reduces dependence on a single overcalling report.Source: §4.6.1 · p. 7, Figure 4 · p. 8, §4.6.3 · p. 8

4.6.2. Calibration

• VRM has ECE=0.092 versus Standard KD’s ECE=0.119, and its Brier score is 0.142.The passage presents the lower calibration error as support for safer thresholded triage decisions under deployment constraints.Source: §4.6.2 · p. 7

4.7. Robustness to Hallucinated Supervision

• 14.6% error at K = 5 on hallucinated cases, down from 42.5% on MIMIC-CXR and 44.8% on RIH-CXR at K = 1.At K = 3, error drops to 20.1%, a 52.7% relative reduction; residual error is 8.5% at K = 10.Source: §4.7 · p. 8, Table 10 · p. 8

4.8.1. Individual Report vs. Marginalized Ensemble

• 0.842 marginalized AUC exceeds the 0.808 mean and 0.832 best single-report AUC, approaching the 0.856 per-case Oracle bound.The accompanying figure reports a 53% reduction in per-sample prediction standard deviation, from 0.095 to 0.045.Source: §4.8.1 · p. 9, Figure 5 · p. 9, Table 11 · p. 9

4.8.2. Report Selection Stability Analysis

• K = 5 yields mean AUC 0.842 with σ = 0.005 and range 0.026, versus K = 1 at 0.808, σ = 0.019, and range 0.097.K = 3 is intermediate at σ = 0.011; the bootstrap figure shows tighter K = 5 distributions and faster running-mean convergence.Source: §4.8.2 · pp. 9–10, Table 12 · p. 9, Figure 6 · p. 10

4.9. Sensitivity to LVLM Sampler Choice

• 0.842 and 0.839 AUC for medical samplers exceed 0.786 and 0.779 for general-purpose samplers, a mean gap of approximately 5.8% AUC.General-purpose samplers still exceed Standard KD (0.748) by 3.1–3.8%, but fall below direct FT baselines (DINOv2: 0.821; EVA-X: 0.828).Source: §4.9 · p. 10, Table 13 · p. 11

4.10.1. Per-Pathology Performance Breakdown

• Urgent tension pneumothorax (0.912 F1) and pulmonary edema (0.875 F1) are classified most reliably, while acute pneumonia reaches 0.724 F1.“No Finding” reaches 0.891 F1, whereas small effusions reach 0.658 F1 and are often misclassified as urgent.Source: §4.10.1 · p. 10, Table 14 · p. 11

4.10.2. Systematic Bias Assessment

• Among 200 misclassified cases, chronic-versus-acute ambiguity accounts for 36%, device/artifact false positives 31%, and small-effusion over-triage 21%.Marginalization reduces stochastic hallucinations but does not eliminate coherent model biases; the authors motivate confidence-gated escalation and review for high-uncertainty cases.Source: §4.10.2 · pp. 10–11

Discussion

• VRM’s marginalized teacher distribution makes supervision less sensitive to stochastic report variation and better aligned with clinical uncertainty.Ablations, individual-versus-ensemble comparisons, and bootstrap tests indicate improved supervision quality and deployment stability.Source: §5 · p. 11
• VRM adds a marginal +1.4% AUC over the strongest direct FT baseline, alongside reported gains in calibration, deployment stability, and hallucination robustness.The discussion describes marginalization as a hedge against the “report lottery” and notes that high-quality LVLM samplers are needed for peak performance.Source: §5 · p. 11
• Under the same 90% sensitivity constraint, VRM outperforms direct FT and standard KD; confidence-gated routing handles routine cases at the edge and selectively escalates ambiguous cases.The paper presents this routing as preserving low overall latency.Source: §5 · p. 11
• VRM operates locally with deterministic edge compute while distilling cloud-level reasoning, addressing edge memory and latency constraints and cloud privacy and transmission burdens.The discussion contrasts impractical direct deployment of massive open-source LVLMs with the VRM approach.Source: §5 · p. 11
• The student’s ceiling is bounded by teacher knowledge and LVLM generation quality, while image-only inference struggles with context-dependent findings and consistent hallucinatory biases can remain.The authors identify high-uncertainty review as critical and list adaptive sampling and prospective multi-site evaluation as future directions.Source: §5 · pp. 11–12

Conclusion

• VRM models missing reports as latent clinical interpretations and uses Monte Carlo marginalization rather than single-report teacher targets.The conclusion describes this as uncertainty-aware supervision under structural modality censorship.Source: §6 · p. 12
• Under encoder transfer, VRM surpasses direct FT baselines from the same encoder families, with paired DeLong tests reporting p < 0.01.The conclusion also reports improved calibration, hallucination recovery, and reduced report-selection variance.Source: §6 · p. 12
• Medical-specialized 26B–34B samplers are needed for top performance, while general-purpose samplers still exceed Standard KD.The conclusion characterizes report quality as meaningful but not gating.Source: §6 · p. 12
• The compact edge student is 75MB at 14ms edge latency, while cascade routing reaches AUC 0.941 at 103ms average latency with 20.3% cloud escalation.The conclusion also reports an OOD average AUC of 0.806 across five datasets.Source: §6 · p. 12

Ethical Compliance

• MIMIC-CXR, CheXpert, IU-CXR, and PadChest were used under their stated data-use agreements or licenses.The passage names the PhysioNet credentialed license, Stanford research agreement, public-domain status, and University of Alicante academic-use license, respectively.Source: Ethical Compliance · p. 12
• RIH-CXR was deemed exempt from IRB review as a quality improvement initiative.Source: Ethical Compliance · p. 12
• The experiments complied with institutional guidelines and HIPAA regulations.Source: Ethical Compliance · p. 12

Patient Informed Consent

• RIH-CXR did not require individual patient informed consent because it was deemed exempt from IRB review as a quality improvement initiative.The study used de-identified retrospective chest X-ray data and publicly available datasets governed by their respective data use agreements.Source: Patient Informed Consent · p. 12

Funding

• Four U.S. National Science Foundation grants supported the work in part, and the funders had no role in study design, analysis, writing, or the submission decision.The grants were IIS-2306789, IIS-2306790, IIS-2306791, and IIS-2306792.Source: Funding · p. 12

Broader Impact

• VRM is intended to augment radiologists by surfacing potentially urgent cases for faster expert review, not to replace their decision-making.The authors frame the approach within a reliability-centered paradigm for high-stakes clinical workflows.Source: Broader Impact · p. 12
• Clinical deployment should include appropriate regulatory clearance and ongoing performance monitoring across demographically diverse populations.The authors cite FDA 510(k) or De Novo pathways as examples of regulatory clearance.Source: Broader Impact · p. 12
• Development remains anchored to a limited set of institutional pipelines and labeling practices, which may reduce performance in underrepresented populations.The authors call for prospective validation with equitable representation across age, sex, race, and geographic origin.Source: Broader Impact · p. 12
• The authors state that expert-level triage on consumer-grade hardware has potential to improve outcomes in resource-limited settings where timely specialist radiology access is a bottleneck.Source: Broader Impact · p. 12