Highlights Anatomically Faithful but Temporally Blind: Auditing Attribution for Left-Ventricular Ejection-Fraction Estimation from Echocardiography Hyunkyung Han,Min Jung Kim • Attribution for echocardiographic EF is audited for spatial and temporal faithfulness. • Two architectures are anatomically faithful: IoR 2.91× and 1.98× chance. • Temporal localization of the ES/ED frames is indistinguishable from chance. • Occlusion shows the models themselves ignore the decisive ES/ED frames.
arXiv:2607.13738v1 [cs.CV] 15 Jul 2026
• Spatial faithfulness does not imply temporal faithfulness in video XAI.
Anatomically Faithful but Temporally Blind: Auditing Attribution for Left-Ventricular Ejection-Fraction Estimation from Echocardiography Hyunkyung Hana , Min Jung Kimb,∗ a School of Integrated Medicine, Yonsei University, Seoul, Republic of Korea b Department of Radiology, Research Institute of Radiologic Science, Yonsei University College of Medicine, Seoul, Republic of Korea
ARTICLE INFO
ABSTRACT
Keywords: Explainable AI Faithfulness Echocardiography Ejection fraction Vision transformer Attribution
Background and Objective: Deep video models estimate left-ventricular ejection fraction (EF) from echocardiography with near-expert accuracy, and post-hoc attribution (e.g., Chefer relevance for transformers, Grad-CAM for CNNs) is increasingly used to certify that such models “look at the right place.” Yet whether these explanations are faithful both spatially and temporally has not been audited. Because EF is defined by the end-systolic (ES) and end-diastolic (ED) frames, a faithful explanation must localize not only the left ventricle (space) but also the decisive frames (time). We audit attribution faithfulness for video EF estimation along three axes: spatial grounding, perturbation, and temporal localization. Methods: We fine-tune two architecturally distinct EF regressors on EchoNet-Dynamic—a selfsupervised VideoMAE transformer and a Kinetics-pretrained R(2+1)D CNN—and audit each with architecture-matched attribution (Chefer relevance; Grad-CAM). Faithfulness is quantified by (i) intersection-over-relevance (IoR) against LV segmentation masks relative to a chance baseline, (ii) deletion AUC against a random-deletion baseline, and (iii) a temporal localization index measuring attribution mass on ES/ED frames relative to chance, with per-case 95% confidence intervals over 50 test studies and attention-rollout and random-attribution controls. A tubelet-occlusion probe disentangles whether temporal behavior reflects an attribution failure or a property of the model itself. Results: Both models are anatomically faithful—IoR 2.91× (VideoMAE) and 1.98× (R(2+1)D) above chance—yet temporally blind: temporal localization is indistinguishable from chance (0.97–1.00) and no better than random attribution. Occlusion shows the models themselves do not preferentially rely on ES/ED (0.90× chance), so temporal blindness is an honest reflection of model behavior rather than an attribution artifact. Conclusions: Spatial faithfulness does not imply temporal faithfulness. Attribution can certify anatomical grounding while masking that a model ignores the clinically decisive frames—a caution for XAI-based validation of video diagnostic models and a call for temporally-aware training and evaluation.
1. Introduction Left-ventricular ejection fraction (LVEF) is among the most consequential quantitative biomarkers in cardiology, guiding decisions from heart-failure staging to chemotherapy cardiotoxicity monitoring and device eligibility. Clinically, it is obtained from echocardiography by contouring the left ventricle at end-diastole (ED) and end-systole (ES) and comparing the resulting volumes—an inherently spatiotemporal measurement that depends on identifying both the correct anatomy and the correct moments in the cardiac cycle. Deep-learning models now estimate LVEF directly from echocardiogram video with expert-level accuracy [1], and video transformers have become competitive backbones for the task [2, 3]. Expert-level accuracy alone, however, is insufficient for clinical adoption: clinicians must be able to interrogate why a model produced a given estimate. Attribution maps—spatial or spatiotemporal heatmaps of input importance—have become the dominant explanation format and are increasingly reported alongside medical-imaging predictions. Their utility rests on an assumption that is rarely tested: that the attribution is faithful, meaning the highlighted regions are ∗ Corresponding author.
[email protected] (H. Han); [email protected] (M.J. Kim)
ORCID (s): 0009-0006-2672-384X (H. Han); 0000-0003-4949-1237 (M.J. Kim)
H. Han and M. J. Kim: Preprint submitted to Elsevier
Page 1 of 11
Anatomically faithful but temporally blind
both anatomically meaningful and causally responsible for the output. For video models this assumption is doubly demanding, because a faithful explanation must localize correctly in space (the left ventricle) and in time (the ES/ED frames that define EF). Existing evaluations fall short of testing this. Most report qualitative overlays or apply perturbation-based checks to 2D image classifiers, and studies that do examine echocardiography attribution seldom separate spatial from temporal faithfulness—even though, for EF, temporal grounding is arguably the more clinically natural expectation. It therefore remains unknown whether video attribution that looks convincing in space is also selective in time, or whether apparently spatiotemporal explanations conceal a temporal blind spot. We address this gap with a quantitative audit of attribution for LVEF estimation, applied to two backbones with distinct inductive biases: the spatiotemporal-convolutional R(2+1)D [3], audited with Grad-CAM, and the selfsupervised VideoMAE transformer [2], audited with the transformer-native Chefer relevance-propagation method [4]. Our contributions are threefold. First, we define a three-axis faithfulness protocol for echocardiography XAI— spatial grounding against expert LV masks, deletion-based causal faithfulness, and temporal localization against ES/ED frames—each measured against an explicit chance baseline. Second, we reveal a striking dissociation that holds across both architectures: attribution is strongly grounded in the left ventricle (IoR 0.430 vs. chance 0.148, a 2.91× enrichment for VideoMAE; 1.98× for R(2+1)D), yet its temporal localization is indistinguishable from chance and no better than random attribution. Third, we argue that this spatial/temporal dissociation is an actionable, quantified caveat for clinical deployment: the attribution reliably indicates where the model looks but not when, and clinicians should not over-read temporal cues from current video XAI.
2. Related Work Deep learning for echocardiographic function. Automated cardiac function assessment from echocardiography was established at scale by EchoNet-Dynamic [1], which trained spatiotemporal convolutional networks on beat-level clips and matched expert variability in EF estimation; the paradigm has since been extended to pediatric cohorts [5] and to open segmentation benchmarks such as CAMUS [6]. Architecturally, EF estimators have progressed from factorized 3D convolutions [3] to video transformers with self-supervised pretraining [2] and, most recently, to dedicated echo video-regression transformers [7]. These systems achieve strong accuracy but remain predominantly black boxes: they output an EF value without exposing which anatomy or which cardiac phase drives the estimate.
Video representation backbones. General-purpose video models provide the backbones adapted for echocardiogra-
phy. Inflated 3D convolutions [8], factorized spatiotemporal convolutions [3], and two-pathway networks [9] dominated the convolutional era, while attention-based models—TimeSformer [10], ViViT [11], and the masked-autoencoding VideoMAE [2]—now lead on action recognition. We audit one representative from each family (R(2+1)D and VideoMAE) to test whether our findings depend on inductive bias.
Attribution methods. Post-hoc attribution scores input importance to explain a prediction. Gradient- and perturbationbased methods for convolutional networks include saliency maps [12], guided backpropagation [13], Grad-CAM [14], Integrated Gradients [15], SmoothGrad [16], layer-wise relevance propagation [17], and model-agnostic surrogates such as LIME [18] and SHAP [19]. For transformers [20, 21], raw attention is an unreliable explanation—attention weights can be uncorrelated with feature importance and freely manipulated [22, 23]—motivating attention-flow aggregation [24] and relevance-propagation methods that combine attention with gradients [4, 25]. Generic LRP rule sets require access to internal linear maps that fused attention implementations do not expose, so we adopt the transformer-native Chefer formulation for VideoMAE and Grad-CAM for R(2+1)D.
Evaluating faithfulness, and the temporal gap. Whether an attribution is faithful—reflecting the model’s true
decision process rather than merely looking plausible—is assessed with perturbation metrics such as deletion/insertion [26], remove-and-retrain [27], and model-randomization sanity checks [28], several of which reveal that popular methods can fail basic tests. In medical imaging these evaluations are largely confined to 2D classifiers and to spatial grounding [29], and a broader critique warns that current explanations rarely certify that a model reached its decision for a clinically valid reason [30]. The temporal dimension of faithfulness—intrinsic to video and clinically central for EF, which is defined by the ES/ED phases—is rarely isolated or quantified. We fill this gap by auditing spatial and temporal faithfulness separately, and by diagnosing, through frame occlusion, whether an observed temporal null originates in the explanation or in the model.
H. Han and M. J. Kim: Preprint submitted to Elsevier
Page 2 of 11
Anatomically faithful but temporally blind
3. Methods 3.1. Task and data We frame left-ventricular ejection fraction (LVEF) as a video regression problem: given a B-mode echocardiogram clip 𝑥 ∈ ℝ𝑇 ×𝐻×𝑊 of 𝑇 frames, a model predicts a scalar 𝑦̂ ∈ [0, 100]. We use the EchoNet-Dynamic dataset [1], which provides apical four-chamber videos with expert EF labels and, for each study, annotated end-diastolic (ED) and end-systolic (ES) frames together with left-ventricular (LV) tracings at those frames. We adopt the standard train/validation/test split and sample clips of 𝑇 frames at 112 × 112 resolution.1
3.2. Models To test whether our findings depend on architecture, we audit two backbones with deliberately different inductive biases. The first, R(2+1)D [3], factorizes 3D convolution into separate spatial and temporal operators and serves as a strong convolutional baseline. The second, VideoMAE [2], is a self-supervised video vision transformer pretrained by masked autoencoding and then fine-tuned for EF regression. Both are trained with an 𝐿1 regression objective and AdamW; the transformer is fine-tuned at a reduced learning rate to preserve the pretrained encoder. A single linear head maps the pooled representation to the EF scalar.
3.3. Architecture-matched attribution We explain VideoMAE predictions with the relevance-propagation method of Chefer et al. [4], which combines attention weights with their gradients and propagates relevance layer by layer to yield a per-token relevance map. We adopt this transformer-native formulation for a concrete implementation reason: VideoMAE’s attention is realized through PyTorch’s fused nn.MultiheadAttention, which does not expose per-head attention matrices or the intermediate linear maps that generic layer-wise relevance-propagation rule sets (e.g. 𝜖/𝛾/𝑧+ rules) require, making those methods inapplicable without re-implementing the block. The Chefer method instead hooks the attention and its gradient directly. For the convolutional R(2+1)D backbone, which exposes no attention, we use Grad-CAM at the final spatiotemporal convolutional block as an architecture-matched attribution and apply the identical three-axis protocol. In both cases the resulting relevance 𝑅 ∈ ℝ𝑇 ×ℎ×𝑤 is defined over the spatiotemporal token (or feature) grid; ∑ we bilinearly upsample it to the input resolution for spatial analysis and sum over space, 𝑟(𝑡) = 𝑝 𝑅𝑡 (𝑝), for temporal analysis.
3.4. Three-axis faithfulness protocol We evaluate attribution along three complementary axes, each compared against an explicit chance baseline so that “above chance” is well defined.
Spatial grounding (IoR). Within the ES/ED windows, let 𝑀𝑡 be the expert LV mask. We define the intersectionover-relevance ∑ ∑ IoR =
𝑡
𝑝∈𝑀𝑡 𝑅𝑡 (𝑝)
∑ ∑ 𝑡
𝑝 𝑅𝑡 (𝑝)
,
(1)
i.e. the fraction of relevance mass that falls inside the LV. The spatial chance baseline is the mask area fraction 𝔼[|𝑀𝑡 |∕|⋅|], the IoR expected under spatially uniform relevance; we report IoR and its ratio to this baseline. All IoR values in this paper use an identical protocol: relevance and mask are evaluated on the 7×7 token grid, and IoR is the fraction of the mask covered by the top-30% most-relevant grid cells, so that every model and attribution method is compared on the same footing.
Causal faithfulness (deletion). We progressively remove the most-relevant tokens—replacing them with a zero-fill
baseline—and re-evaluate the model, tracing the change in the EF prediction as a function of the fraction of tokens deleted; we summarize the resulting curve by its normalized area (deletion AUC). To calibrate this metric we also delete tokens in a random order, so that the attribution’s deletion AUC can be read against a random-deletion baseline rather than in absolute terms. 1 The EchoNet-Dynamic split comprises 7,465 training, 1,288 validation, and 1,277 test videos; we sample clips of 𝑇 = 32 frames at 112 × 112 resolution with an ES/ED-covering sampler.
H. Han and M. J. Kim: Preprint submitted to Elsevier
Page 3 of 11
Anatomically faithful but temporally blind
Algorithm 1 Three-axis faithfulness audit for video EF attribution Require: clip 𝑥 ∈ ℝ𝑇 ×𝐻×𝑊 ; model 𝑓 ; attribution ; LV masks {𝑀𝑡 }; ES/ED indices (𝑡𝑒𝑠 , 𝑡𝑒𝑑 ) 1: 𝑅 ← (𝑓 , 𝑥) ⊳ relevance over grid 𝑇 × ℎ × 𝑤 ̃ ← upsample(𝑅); 𝑟(𝑡) ← ∑ 𝑅̃ 𝑡 (𝑝) 2: 𝑅 𝑝 Axis 1 — grounding ( ∑spatial )/( ∑ ∑ ) ∑ ̃ ̃ 3: IoR ← 𝑡 𝑝∈𝑀𝑡 𝑅𝑡 (𝑝) 𝑡 𝑝 𝑅𝑡 (𝑝) 4: chance𝑠𝑝 ← 𝔼𝑡 [ |𝑀𝑡 |∕(𝐻 ⋅𝑊 ) ]; report IoR∕chance𝑠𝑝 Axis 2 — causal faithfulness (deletion) ̃ desc.; for 𝑘 = 1..𝑁 delete top-𝑘, record |𝑓 (𝑥⧵𝑘 ) − 𝑓 (𝑥)|; compare to a random deletion order 5: order tokens by 𝑅 6: delAUC ← AUC of deletion curve Axis 3 — temporal localization ( ) 7: TLI ← align 𝑟(𝑡), ES∕ED proximity , normalized s.t. 1.0 = chance Diagnosis — model property vs attribution failure 8: for each tubelet 𝑡 do 𝑢(𝑡) ← 𝑓 (𝑥) − 𝑓 (𝑥occlude(𝑡) ) 9: end for 10: conc𝑢 ← mass of |𝑢| at ES/ED ∕ chance; 𝜌 ← Spearman(𝑢, 𝑟) 11: if conc𝑢 ≈ 1 then temporal null is a model property (B) else attribution failure (A) 12: return IoR∕chance𝑠𝑝 , delAUC, TLI, conc𝑢 , 𝜌 Table 1 Audited EF regressors on the EchoNet-Dynamic test split. R(2+1)D is the stronger of the two; VideoMAE is weaker but its encoder still grounds on the LV. The temporal blindness reported below holds for both—including the stronger R(2+1)D—so it is not an artifact of the weaker backbone. Neither reaches the best accuracy reported for this dataset; see Limitations. Model
Backbone
VideoMAE R(2+1)D-18
ViT (transformer) CNN (spatiotemporal)
Pretraining
Test MAE (%) ↓
Test 𝑟 ↑
Self-supervised (MAE) Kinetics-400
14.1 7.10
0.53 0.75
Temporal localization. Because EF is computed from the ES/ED volumes, a temporally faithful explanation should
concentrate relevance 𝑟(𝑡) near the annotated ES/ED frames. We measure the alignment between 𝑟(𝑡) and ES/ED proximity, normalized so that a value of 1.00 corresponds to the temporal chance baseline (relevance uniform over frames) and higher values indicate greater ES/ED selectivity; Pearson 𝑟 between 𝑟(𝑡) and ES/ED proximity is reported as an auxiliary correlate.
4. Results 4.1. Predictive performance On the EchoNet-Dynamic test set, the Kinetics-pretrained R(2+1)D-18 is a competent EF regressor (MAE 7.10 EF points, Pearson 𝑟 = 0.75), operating in a clinically usable range (Table 1). The self-supervised VideoMAE is a weaker regressor (MAE 14.1, 𝑟 = 0.53); its encoder nonetheless localizes the left ventricle (Sec. 4.2) and its EF estimate responds to LV tubelet occlusion (Sec. 5), confirming that it computes an LV-based EF signal. Crucially, the temporal blindness we report below holds for both backbones, including the stronger R(2+1)D, so it is not an artifact of the weaker regressor; and because our audit concerns where a model attends rather than its absolute accuracy, VideoMAE’s modest MAE does not affect the conclusions, which replicate on the stronger backbone. Neither regressor matches the best accuracy reported for this dataset, and we set out in Sec. 5 what that does and does not license.
4.2. Attribution is spatially grounded Attribution overlaps the expert LV mask far above chance (VideoMAE IoR = 0.430 vs. 0.148, 2.91×). The same pattern holds for the convolutional R(2+1)D backbone under Grad-CAM (IoR = 0.361±0.088, 1.98× chance; Table 2, Fig. 1–2), showing that spatial grounding is architecture-independent rather than a quirk of one backbone or one attribution method. H. Han and M. J. Kim: Preprint submitted to Elsevier
Page 4 of 11
Anatomically faithful but temporally blind Table 2 Spatial grounding (intersection-over-relevance, IoR, against LV masks). Both architectures concentrate relevance on the left ventricle far above chance—attribution is anatomically faithful regardless of backbone. Model
Attribution
IoR ↑
Chance
Ratio ↑
VideoMAE R(2+1)D-18
Chefer Grad-CAM
0.430 0.361 ± 0.088
0.148 0.182
2.91× 1.98×
Deletion provides a weaker, more equivocal signal. Deleting the most-relevant tokens first perturbs the EF prediction (deletion AUC = 0.180±0.098) only marginally more than deleting tokens in a random order (0.161±0.092), and insertion shows a similarly small gap (0.150 vs. 0.146). This limited separation is consistent with the known outof-distribution confound of token-removal metrics—replacing patches with a baseline perturbs the model regardless of which patches are removed. We therefore report deletion/insertion for transparency but base our faithfulness conclusions on the spatial-grounding and temporal-localization analyses, which are not subject to this confound.
Figure 1: Spatial grounding is architecture-independent: relevance concentrates on the LV far above chance for both the VideoMAE transformer under Chefer (2.91×) and the R(2+1)D CNN under Grad-CAM (1.98×).
Figure 2: Per-case spatial grounding for R(2+1)D Grad-CAM (50 studies). The IoR distribution (mean 0.361) lies well above the spatial chance baseline (0.182), confirming anatomical faithfulness for the convolutional backbone.
H. Han and M. J. Kim: Preprint submitted to Elsevier
Page 5 of 11
Anatomically faithful but temporally blind Table 3 Temporal localization index (attribution mass on ES/ED frames, normalized so that 1.00 = chance; higher = more temporal selectivity), mean ± 95% CI over 50 studies. No method—across either architecture— exceeds chance, and Chefer is statistically indistinguishable from random attribution: the explanations are temporally blind. Model
Attribution
Temporal localization (1.00 = chance)
VideoMAE VideoMAE VideoMAE
Chefer Attention-rollout Random
0.973 ± 0.027 1.007 ± 0.007 0.997 ± 0.007
R(2+1)D-18
Grad-CAM
1.000 ± 0.000
4.3. Temporal localization is at chance In contrast, temporal alignment with the ES/ED frames is at chance (0.973 ± 0.027, 95% CI; Pearson 𝑟 = 0.075 as an auxiliary measure), showing the attribution does not identify when the model attends. This temporal null is not specific to Chefer or to the transformer: attention-rollout and random attribution are statistically indistinguishable from it, and R(2+1)D Grad-CAM likewise shows no ES/ED selectivity (Table 3, Fig. 3). Fig. 4 makes the dissociation concrete on a single representative study: relevance sits inside the expert LV tracing at both the end-systolic and end-diastolic frames (panels a,b), the deletion curve separates only marginally from its random baseline (panel c), and the per-frame relevance profile shows no peak at the annotated ES/ED frames (panel d). The three axes are summarised in Table 4: attribution certifies where the model looks but not when, and the pattern is reproduced across both a transformer and a CNN.
Figure 3: Per-case temporal localization index over 50 test studies (index 1.0 = chance). Chefer, attention-rollout, and random attribution all sit at chance: no method identifies when the model attends.
H. Han and M. J. Kim: Preprint submitted to Elsevier
Page 6 of 11
Anatomically faithful but temporally blind
Figure 4: Three-axis attribution audit on a representative test case (predicted EF = 40.7%). (a,b) Chefer relevance overlaid on the end-systolic and end-diastolic frames concentrates inside the expert LV tracing (above-chance spatial grounding; dataset IoR = 0.430 vs. chance 0.148, 2.91×). (c) Deleting the most-relevant tokens shifts the EF prediction (dataset deletion AUC = 0.180), only marginally above a random-deletion baseline (0.161); we report this for transparency. (d) Per-frame relevance mass is not concentrated at the ES/ED frames (dashed lines) that define EF (at-chance temporal localization; dataset score 0.973). The attribution is thus spatially faithful but temporally blind.
Table 4 Summary of the three-axis faithfulness audit. Attribution certifies where (anatomy) but not when (the ES/ED frames that define EF); the dissociation holds across both a transformer and a CNN. Axis
Metric
Spatial grounding Spatial grounding
Result
IoR ratio (Chefer, VideoMAE) 2.91× chance IoR ratio (Grad-CAM, 1.98× chance R(2+1)D) Perturbation Deletion vs. random 0.180 vs. 0.161 (marginal) Temporal localiza- Loc. index (all methods) 0.97–1.00 tion AttributionChefer vs. Random (temporal) indistinguishable specific?
Faithful? ✓ ✓ ∼ × ×
5. Discussion Our audit dissociates two facets of faithfulness that are usually conflated: the attribution is trustworthy about where the model looks (spatial grounding well above chance) but uninformative about when (temporal localization at chance).
H. Han and M. J. Kim: Preprint submitted to Elsevier
Page 7 of 11
Anatomically faithful but temporally blind Table 5 Tubelet-occlusion diagnosis (VideoMAE). 𝑢(𝑡) is the causal EF change when tubelet 𝑡 is occluded (what the model relies on); 𝑠(𝑡) is the Chefer temporal profile (what the explanation claims). Because the model itself does not preferentially rely on ES/ED (0.90× chance), temporal blindness is an honest property of the model, not a failure of the attribution. Quantity conc𝑢 @ ES/ED (model reliance) conc𝑠 @ ES/ED (attribution mass) 𝜌(𝑢, 𝑠)
Value 0.90× chance 0.97× chance +0.245
Interpretation model does not over-rely on ES/ED explanation is also flat in time attribution weakly tracks causal importance
Verdict: (B) model property — temporal blindness is honest; the fix is temporally-aware modeling.
Because LVEF is clinically defined from the ES/ED volumes, a naive reading would treat the temporal null as a failure of the explanation. We argue the opposite.
Is the temporal null an attribution artifact or a model property? To disambiguate, we run a frame-occlusion analysis on 50 test studies: for each clip we occlude each tubelet in turn and record the induced change in the EF prediction, yielding a per-frame causal-importance profile 𝑢(𝑡) that is independent of any attribution method. Two findings emerge (Table 5, Fig. 5). First, occluding the ES/ED tubelets perturbs the prediction no more than occluding arbitrary tubelets (causal mass at ES/ED = 0.90× chance), indicating the model does not preferentially rely on the clinically decisive frames. Second, the Chefer relevance profile weakly but positively tracks this causal importance (Spearman 𝜌 = 0.245) while itself being unconcentrated at ES/ED (0.97× chance). Together these show that the temporal null is a property of the model, not a failure of the attribution: the network estimates EF from left-ventricular geometry available throughout the clip rather than from the ES/ED contrast, and the attribution faithfully reflects this temporal agnosticism. Read this way, “temporally blind” is not an unfaithful explanation but a faithful explanation of a temporally-agnostic model.
Figure 5: Frame-occlusion diagnosis (50 studies). Left: causal reliance on ES/ED (conc𝑢 ) is centered below chance (mean 0.90×): the model does not preferentially use the clinically decisive frames. Right: the Chefer profile weakly tracks true causal importance (𝜌 = 0.245). Together these establish that temporal blindness is a property of the model, not a failure of the attribution (Verdict B).
Why might the model be temporally agnostic? EF is a ratio of end-diastolic and end-systolic volumes, yet the instantaneous LV geometry that determines those volumes is visible in every frame of an apical four-chamber view. A regressor can therefore approximate EF from aggregate chamber geometry across the clip without explicitly localizing the two extremal phases—consistent with the flat causal-importance profile we observe. This does not imply the phases are clinically irrelevant; it implies the learned solution does not route the prediction through them.
H. Han and M. J. Kim: Preprint submitted to Elsevier
Page 8 of 11
Anatomically faithful but temporally blind
Faithfulness is not clinical validity. We stress that a faithful explanation is not the same as a clinically valid one.
Our results show the attribution honestly reflects the model’s decision process; but that process itself does not route the EF estimate through the ES/ED frames a cardiologist would use. A faithful explanation of a temporally-agnostic model therefore does not license clinical trust—on the contrary, it surfaces that the model reaches its answer for a reason that departs from clinical practice. A spatially convincing heatmap presented without a temporal check can thus give a false sense of clinical grounding.
Implications for clinical video XAI. Our results argue for reporting spatial and temporal faithfulness separately.
A spatially convincing heatmap can coexist with a complete absence of temporal selectivity, so clinicians should not adopt individual frames highlighted by current video XAI as clinical evidence of when a model attended; a temporal-localization check against task-defining frames (here, ES/ED) should accompany any spatial heatmap before an explanation is used to justify a per-patient decision. More broadly, the spatial/temporal dissociation is a concrete, quantified caveat for deploying video attribution as a trust signal.
Limitations. Our audit covers a single dataset (EchoNet-Dynamic). Our findings indicate that the model does not rely on the ES/ED frames, not that it is incapable of using temporal information in principle; establishing the latter would require interventional training and is left to future work. For R(2+1)D, Grad-CAM’s coarse temporal resolution at the final convolutional block may reinforce the temporal null; however, the same null appears for VideoMAE under the higher-resolution Chefer method, so it is not attributable to resolution alone. A further limitation concerns the absolute accuracy of the audited regressors. Our R(2+1)D-18 reaches a test MAE of 7.10 EF points, whereas the same architecture has been reported on this dataset at 4.10, and recent methods reach 3.90–4.07 [31, 32, 33]; our backbones are therefore not fully converged relative to the published state of the art, in part because we evaluate a single clip per study while those results average predictions over clips or beats. This does not affect the internal logic of the audit, which compares where a given model attends against chance rather than against another model’s accuracy, and the temporal null is reproduced across two architectures and by an attribution-independent occlusion probe. It does, however, bound the generality of the claim: we show that these regressors do not route their EF estimate through the ES/ED frames, not that a fully converged regressor would fail to do so. Whether ES/ED reliance emerges with additional training, larger pretraining corpora, or multi-clip aggregation is an open question, and one we regard as the most informative single follow-up to this work. Extending the three-axis protocol to additional attribution methods, architectures, tasks, and external cohorts is important future work, as is a clinician study of how temporal cues in video XAI are actually read.
6. Conclusion Attribution for echocardiographic LVEF estimation is anatomically faithful in space but temporally blind—a dissociation that holds across both a transformer (Chefer) and a CNN (Grad-CAM), and which a frame-occlusion probe attributes to the model rather than to the explanation. We recommend reporting spatial and temporal faithfulness separately when deploying video XAI in cardiology.
Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Data availability This study uses the publicly available EchoNet-Dynamic dataset [1], available at https://echonet.github.io/ dynamic/ (access requires acceptance of the Stanford AIMI data use agreement). Code to reproduce the three-axis faithfulness audit will be released upon publication.
H. Han and M. J. Kim: Preprint submitted to Elsevier
Page 9 of 11
Anatomically faithful but temporally blind
Acknowledgements References [1] D. Ouyang, B. He, A. Ghorbani, N. Yuan, J. Ebinger, C. P. Langlotz, et al., Video-based AI for beat-to-beat assessment of cardiac function, Nature 580 (7802) (2020) 252–256. [2] Z. Tong, Y. Song, J. Wang, L. Wang, VideoMAE: Masked autoencoders are data-efficient learners for self-supervised video pre-training, in: Advances in Neural Information Processing Systems (NeurIPS), Vol. 35, 2022. [3] D. Tran, H. Wang, L. Torresani, J. Ray, Y. LeCun, M. Paluri, A closer look at spatiotemporal convolutions for action recognition, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 6450–6459. [4] H. Chefer, S. Gur, L. Wolf, Transformer interpretability beyond attention visualization, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 782–791. [5] C. S. Reddy, L. Lopez, D. Ouyang, J. Y. Zou, B. He, Video-based deep learning for automated assessment of left ventricular ejection fraction in pediatric patients, Journal of the American Society of Echocardiography 36 (2023) 482–489. [6] S. Leclerc, E. Smistad, J. Pedrosa, A. Østvik, et al., Deep learning for segmentation using an open large-scale dataset in 2D echocardiography, IEEE Transactions on Medical Imaging 38 (9) (2019) 2198–2210. [7] T. Akan, S. Alp, M. S. Bhuiyan, T. Helmy, A. W. Orr, M. M. R. Bhuiyan, S. A. Conrad, J. A. Vanchiere, C. G. Kevil, M. A. N. Bhuiyan, ViViEchoformer: Deep video regressor predicting ejection fraction, Journal of Imaging Informatics in Medicine 38 (2024) 2041–2052. [8] J. Carreira, A. Zisserman, Quo vadis, action recognition? A new model and the kinetics dataset, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 6299–6308. [9] C. Feichtenhofer, H. Fan, J. Malik, K. He, SlowFast networks for video recognition, in: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 6202–6211. [10] G. Bertasius, H. Wang, L. Torresani, Is space-time attention all you need for video understanding?, in: International Conference on Machine Learning (ICML), 2021. [11] A. Arnab, M. Dehghani, G. Heigold, C. Sun, M. Lučić, C. Schmid, ViViT: A video vision transformer, in: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 6836–6846. [12] K. Simonyan, A. Vedaldi, A. Zisserman, Deep inside convolutional networks: Visualising image classification models and saliency maps, in: ICLR Workshop, 2014. [13] J. T. Springenberg, A. Dosovitskiy, T. Brox, M. Riedmiller, Striving for simplicity: The all convolutional net, in: ICLR Workshop, 2015. [14] R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, D. Batra, Grad-CAM: Visual explanations from deep networks via gradientbased localization, in: Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017, pp. 618–626. [15] M. Sundararajan, A. Taly, Q. Yan, Axiomatic attribution for deep networks, in: International Conference on Machine Learning (ICML), 2017, pp. 3319–3328. [16] D. Smilkov, N. Thorat, B. Kim, F. Viégas, M. Wattenberg, SmoothGrad: removing noise by adding noise, arXiv preprint arXiv:1706.03825 (2017). [17] S. Bach, A. Binder, G. Montavon, F. Klauschen, K.-R. Müller, W. Samek, On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation, PLoS ONE 10 (7) (2015) e0130140. [18] M. T. Ribeiro, S. Singh, C. Guestrin, “Why should I trust you?”: Explaining the predictions of any classifier, in: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), 2016, pp. 1135–1144. [19] S. M. Lundberg, S.-I. Lee, A unified approach to interpreting model predictions, in: Advances in Neural Information Processing Systems (NeurIPS), 2017. [20] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, I. Polosukhin, Attention is all you need, in: Advances in Neural Information Processing Systems (NeurIPS), 2017. [21] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, et al., An image is worth 16x16 words: Transformers for image recognition at scale, in: International Conference on Learning Representations (ICLR), 2021. [22] S. Jain, B. C. Wallace, Attention is not explanation, in: Proceedings of NAACL-HLT, 2019, pp. 3543–3556. [23] S. Wiegreffe, Y. Pinter, Attention is not not explanation, in: Proceedings of EMNLP-IJCNLP, 2019, pp. 11–20. [24] S. Abnar, W. Zuidema, Quantifying attention flow in transformers, in: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), 2020, pp. 4190–4197. [25] H. Chefer, S. Gur, L. Wolf, Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers, in: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 397–406. [26] V. Petsiuk, A. Das, K. Saenko, RISE: Randomized input sampling for explanation of black-box models, in: British Machine Vision Conference (BMVC), 2018. [27] S. Hooker, D. Erhan, P.-J. Kindermans, B. Kim, A benchmark for interpretability methods in deep neural networks, in: Advances in Neural Information Processing Systems (NeurIPS), 2019. [28] J. Adebayo, J. Gilmer, M. Muelly, I. Goodfellow, M. Hardt, B. Kim, Sanity checks for saliency maps, in: Advances in Neural Information Processing Systems (NeurIPS), 2018. [29] N. Arun, N. Gaw, P. Singh, K. Chang, et al., Assessing the trustworthiness of saliency maps for localizing abnormalities in medical imaging, Radiology: Artificial Intelligence 3 (6) (2021) e200267. [30] M. Ghassemi, L. Oakden-Rayner, A. L. Beam, The false hope of current approaches to explainable artificial intelligence in health care, The Lancet Digital Health 3 (11) (2021) e745–e750. [31] F. A. Maani, N. Saeed, A. Matsun, M. Yaqub, CoReEcho: Continuous representation learning for 2D+time echocardiography analysis, in: Medical Image Computing and Computer-Assisted Intervention (MICCAI), Springer, 2024, pp. 591–601.
H. Han and M. J. Kim: Preprint submitted to Elsevier
Page 10 of 11
Anatomically faithful but temporally blind [32] R. Muhtaseb, M. Yaqub, EchoCoTr: Estimation of the left ventricular ejection fraction from spatiotemporal echocardiography, in: Medical Image Computing and Computer-Assisted Intervention (MICCAI), Springer, 2022, pp. 370–379. [33] Y. Ghamary, V. Wu, H. Vaseli, C. Luong, T. Tsang, S. A. Bigdeli, P. Abolmaesumi, ProtoEFNet: Dynamic prototype learning for inherently interpretable ejection fraction estimation in echocardiography, arXiv preprint arXiv:2512.03339 (2025).
H. Han and M. J. Kim: Preprint submitted to Elsevier
Page 11 of 11