Skip to main content
Search

# Explainable AI

17 related articles

Subscribe to updates
  1. Hybrid quantum medical models get two-level explanations: exact attribution over 8 semantic dims, and CNOT pruning that raises accuracy under noise

    Hybrid quantum-classical medical models can predict but cannot say which semantic factors or which quantum gates drove a decision. FQXAI compresses inputs into an eight-dimensional proxy-supervised semantic bottleneck, turns it into IF-THEN rationales through Gaussian fuzzy membership states, and computes exact Shapley values over all 256 semantic coalitions without Monte Carlo sampling. On MIT-BIH Arrhythmia, PneumoniaMNIST and PathMNIST the models reached accuracies of 0.890, 0.768 and 0.726, and under 5% depolarizing noise SVQX-guided pruning cut CNOT counts from 14 to 11, 12 and 10 while raising accuracy by 2.20, 1.74 and 1.51 percentage points.

  2. Nature Medicine

    A medical vision-language model that scores 0.895 on diagnosis and points back to the lesion

    Medical imaging foundation models have traded generality against pixel-level precision, and most return a label without evidence. EVLF-FM feeds both disease-level and pixel-level visual features into a language model, first with supervised fine-tuning and then with reinforcement learning on verifiable rewards, spanning six specialties and 10 imaging modalities. On internal diagnostic validation it reached the highest mean accuracy among evaluated models, 0.895, and averaged 0.842 Acc@0.5 for medical visual grounding across nine modalities.

  3. A CRISPRi guide-efficacy model nearly fails on a new library; swapping the within-gene percentile label lifts transfer correlation from 0.008 to 0.136

    A CRISPRi guide-efficacy model trained in K562 cells barely transferred to the independent Dolcetto library from Sanson et al. 2018 in HT29 and A375 cells, reaching Spearman correlations of only 0.008 and 0.012. Keeping the same 101 features, architecture and hyperparameters but replacing the within-gene percentile label with a continuous cross-study z-score raised transfer to 0.136 in HT29 and 0.158 in A375, and leave-one-chromosome-out performance in K562 from 0.375 to 0.429. The transfer failure therefore traces mainly to label design rather than missing chromatin features.

  4. ECG models break on filter phase, not filtering: a causal all-pass shift drops conduction-block AUROC from 0.91 to 0.62

    Real-time ECG monitoring must use causal filters, whose nonlinear phase misaligns the frequency components of the cardiac waveform. Across 36,583 twelve-lead records from three independent clinical databases, changing phase alone while holding spectral magnitude fixed cut ResNet-18 macro-AUROC by 0.1244 on PTB-XL, with conduction-block detection falling from above 0.91 to about 0.62. A training-time regularizer that randomly applies all-pass phase perturbations recovered 86.9% to 90.9% of that loss across cohorts without changing the architecture.

  5. arXiv

    Where a model looks is not where people look: gaze heatmaps as a benchmark show Mamba localizes chest X-rays well but misses human intensity

    A cross-architecture comparison of CNNs, ViTs and Vision Mamba uses human eye-tracking heatmaps as an independent benchmark to quantify how closely model saliency matches human gaze, finding that Mamba reaches the highest localization scores on chest X-rays while its saliency intensity is nearly uncorrelated with human fixations.

  6. Ground beef lymph-node adulteration moves past visual inspection: four-excitation fluorescence plus LSTM holds RPD 9.3 at unseen slaughterhouses

    Once lymph-node tissue is mixed into finely comminuted ground beef, visual inspection can no longer find it, while DNA and protein methods are specific but destructive and need wet-chemistry preparation. Four-excitation front-face fluorescence spectroscopy with an explainable LSTM was trained on 824 samples from five slaughterhouses and three breeds, then tested on two unseen slaughterhouses, reaching R2P 0.988–0.993, RMSEP 2.60–3.35% w/w and RPD 9.3–12.0. An excitation ablation attributes the specificity to multi-excitation sampling rather than the network, with a 1.8% w/w detection limit and a 6.0% w/w quantification limit.

  7. Thousands of uncertain monogenic diabetes variants get a ranked list, and an ablation shows how much of the accuracy is curator logic rebuilt

    Many variants found in monogenic diabetes genes remain variants of uncertain significance because evidence is insufficient or conflicting. A pipeline trained on 7,706 confidently classified single nucleotide variants across 22 such genes reached ROC-AUC 0.9989 and AUPRC 0.9970 on 771 held-out variants, with recall 1.00 on the pathogenic class. Feature ablation showed that part of this accuracy reflects overlap with the ACMG criteria used to assign ClinVar labels, yet on non-missense variants, where computational evidence could not have been directly invoked, the model still beat a majority-class baseline, 99.0% versus 87.0%; it then ranked 6,306 uncertain variants with per-variant explanations.

  8. A wrist-worn single-lead ECG now flags 13 structural heart conditions, supervised by routine echo reports

    A Peking University-led team aligned routine single-lead ECG with physician-authored echocardiography reports to train AnyECG-Echo, a screening framework that resolves 13 fine-grained structural heart disease subtypes at once. Earlier AI-ECG models mostly required 12-lead input, were validated within a single health system, or needed per-condition manual labels. Across internal, temporal and geographically independent external cohorts totaling 25,222 patients, the framework reached high AUROC for high-impact subtypes such as reduced left ventricular systolic function, global heart enlargement and mitral stenosis, and retained usable discrimination in the external cohort.

  9. bioRxiv

    Skip the cell-type labels: NetMap reads regulatory networks and cell identity together from mixed single-cell data

    Regulatory network inference no longer needs cell types called first. Earlier pipelines clustered cells, annotated types with marker genes, then inferred one network per cluster, so clustering and annotation errors propagated into the network. NetMap infers a gene regulatory network per cell before any clustering, uses GradientShap to attribute network edges to input genes, and reaches clustering scores of 1.0, 0.99 and 0.99 across three PBMC multiome evaluation cases.

  10. arXiv

    Polytopal Neural Networks constrain latent representations to learned polytopes, retaining 96.9% of unconstrained accuracy on ImageNet-1k

    The authors propose Polytopal Neural Networks (PNNs), which constrain each layer's latent representation during training to a convex combination of learned archetypes so that the explanatory coordinates are the same coordinates the next layer receives; across image classification and reconstruction benchmarks, PNNs approach unconstrained networks with small accuracy loss, outperform VQ-VAE and Dirichlet VAE in unsupervised reconstruction, and offer a VQ training route without EMA updates, commitment loss, or straight-through estimator.