SMILE: Self-Explainable Multimodal Information Bottleneck for Medical Diagnosis
Yuqing Yang, Alexander Schmatz, Zhaozhao Ma, Changkyu Choi, Robert Jenssen, Shujian Yu
cs.CV · cs.LG
Quick verdict
Proposes IB-based self-explainable multimodal diagnosis framework; leads headline metrics on five datasets but not all AUCs; some internal inconsistencies.
Summary
The paper proposes SMILE, a self-explainable multimodal diagnosis framework that formulates per-sample feature selection as an information bottleneck problem. Modality-specific explainers produce binary masks over inputs (patches for images, genes for omics, edges for brain graphs) via a Gumbel-Softmax top-k relaxation; encoders map masked inputs to latent spaces where mutual information is estimated non-parametrically with a matrix-based Renyi's alpha-order entropy functional, avoiding auxiliary MI estimators. The objective extends INVASE (recovered at beta=0) with a compression term, and a generalization bound extending Kawaguchi et al.'s IB theory suggests compression reduces generalization error. Experiments span five datasets (BRCA, ROSMAP, iCTCF, Glaucoma Grading, REST-meta-MDD). SMILE reports the best ACC/F1 on BRCA (87.3%) and ROSMAP (85.1%), iCTCF accuracy of 92.4% (+9.1 points over HUST-19), best Kappa on Glaucoma, and best ACC/MCC on REST-meta-MDD (67.4%/35.0%), though two baselines retain higher AUC there and DMIB keeps best AUC on ROSMAP. Ablations show learned masks outperform random masks, and built-in explanations beat LIME/SHAP applied to a late-fusion MIB backbone on fidelity and consistency. Limitations: no statistical testing, single-number iCTCF baselines, an inconsistent iCTCF selection budget (k=30 in main text vs k=60 in the supplement), reliance on the sufficient-encoder assumption, and no code release.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- SMILE is the first IB-based self-explainable framework for multimodal medical diagnosis, learning predictive representations and instance-wise, modality-specific explanations jointly rather than post hoc.
- The IB objective (Eq. 2) generalizes INVASE as a special case at beta=0, and the added compression term improves rather than trades off generalization (Proposition 2).
- SMILE achieves state-of-the-art diagnostic performance across five multimodal medical datasets, including a 9.1-point absolute accuracy gain on iCTCF (92.4% vs HUST-19's 83.3%).
- Built-in explanations are more faithful (lower confidence drop) and more stable (lower maximum sensitivity) than LIME/SHAP applied to the same encoders within a late-fusion MIB backbone (Table 6).
Strengths
Clean theoretical unification of IB feature selection with INVASE
Proposition 1 (proof in App. A) correctly shows the SMILE objective reduces to INVASE at beta=0 via the decomposition E[DKL] = I(X;Y) - I(Y;X-tilde); the 2D toy example (Fig. 2) gives a concrete motivation for why entropy minimization adds value beyond the l0 budget.
Broad, repetition-disciplined empirical validation
Five datasets covering omics (BRCA/ROSMAP), CT+clinical (iCTCF), fundus+OCT (Glaucoma), rs-fMRI+sMRI graphs (REST-meta-MDD); all SMILE results are means +/- std over 10 runs (Sec. 4.1), and the iCTCF gain is large: 92.4+/-0.5 ACC vs HUST-19's 83.3 (+9.1 points), AUC 95.6 (Table 3).
Explanations evaluated quantitatively, not only visually
Random-mask control in the ablation
Epistemic honesty in the MDD neuroimaging analysis
Weaknesses
Internal numerical inconsistencies between main text and supplement
'Consistently state-of-the-art' outruns the evidence
No statistical hypothesis testing; mixed-quality baseline reporting
Theory-practice gap in MI estimation and unreported hyperparameters
Weakest evaluation cell: Glaucoma Grading
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
SMILE (arXiv 2609.05174) proposes a self-explainable multimodal diagnosis framework built on the information bottleneck (IB). Claims: (1) it is the first IB-based self-explainable framework for multimodal medical diagnosis; (2) its objective (Eq. 2) generalizes INVASE at beta=0 (Prop. 1) and the compression term improves generalization (Prop. 2); (3) state-of-the-art performance across five datasets; (4) built-in explanations are more faithful than post-hoc ones. The scope spans omics (BRCA/ROSMAP), CT montage + clinical features (iCTCF), fundus + OCT (Glaucoma Grading), and rs-fMRI graphs + sMRI volumes (REST-meta-MDD).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is SMILE in medical AI?
SMILE is a proposed self-explainable multimodal diagnosis framework that uses the information bottleneck principle to select, per patient and per modality, the k most informative input elements (image patches, genes, brain-graph edges) driving the diagnosis; the selected features both serve as the explanation and are used directly for prediction.
How much does SMILE improve accuracy on iCTCF?
The paper reports 92.4% accuracy on the iCTCF COVID-19 prognosis dataset, a 9.1 percentage-point absolute improvement over the previous best reported model HUST-19 (83.3%), with AUC of 95.6% and the best weighted Youden indices among compared methods.
Does SMILE outperform every baseline on every dataset?
No. It leads on headline accuracy/F1/Kappa metrics across the five datasets, but DMIB keeps the best AUC on ROSMAP (91.6% vs 90.3%) and MEIB/MIB report higher AUC on REST-meta-MDD (82.2%/77.7% vs 73.5%); the paper argues accuracy and MCC are the more decision-relevant metrics there.
Is code available for SMILE?
The provided paper text contains no code or data availability statement. Reproduction would require reimplementing from the appendix's architecture and training details, which are fairly complete but leave the Gumbel temperature and kernel bandwidth unspecified for most datasets.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure contents (heatmaps, montage overlays, connectomes in Figs. 2 and 4-10) cannot be assessed from extracted text, so qualitative claims about visualizations rest on captions and surrounding prose only.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.