{"id":"b1525dc1-1588-4400-a321-7730c211cb5d","arxiv_id":"2505.10231","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Human-AI alignment via attention supervision reduced fairness gaps in 27 of 30 chest X-ray comparisons and improved out-of-domain accuracy, but full or randomized alignment can introduce trade-offs.","lead":"Training chest X-ray models with radiologist attention maps, known as human-AI alignment, generally shrank fairness gaps between demographic groups and improved performance on new hospitals' data. A generalist reader may care because this points to a training strategy that could make medical AI fairer and more portable, though over-aligning to human attention can backfire.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Random-attention control already reduces fairness gaps without human insight, so attributing the effect to human-AI alignment is unproven; Table 2's 27/30 count is also not reproducible from the printed values.","rationale":"The Reader's weakest_assumption concerns whether the cross-attention maps are causally tied to classification and whether penalizing them changes the representation rather than the visualization. My concern is a concrete and stronger version of that: the paper's own random-alignment ablation already shows that fairness-gap reductions occur without any human semantic content. This makes the central attribution to 'human insights' insecure, independent of the interpretability of attention maps. The paper's interpretation of the random condition as 'decorrelation' is plausible but untested. A matched random control with equivalent loss magnitude and training budget would settle whether human alignment has any unique fairness benefit. In addition, the 27/30 summary count is arithmetically inconsistent with Table 2 as printed, so the word 'consistently' is not justified by the displayed results. These issues are addressable in revision: add the matched control, report per-comparison confidence intervals or significance tests, and soften the consistency claim. The Reader's CONDITIONAL verdict remains appropriate, so no verdict change is needed.","tokens_in":8185,"tokens_out":7287,"duration_ms":73472,"concrete_test":"Run a matched-control experiment where random attention maps are generated with the same spatial support, size, location distribution, and loss weight as the radiologist annotations, and compare fairness gaps (AUC, accuracy, sensitivity, F1) across three conditions: w/o alignment, Aligner, and random-alignment, under identical training schedules and the same number of repeats. If Aligner fails to beat random alignment within 95% confidence intervals on the four OOD tasks, the human-insight attribution is unsupported. In the same audit, recompute the claimed 27/30 count from Table 2 directionally and with paired significance tests per metric.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that human semantic guidance, not generic regularization, drives the fairness improvements. The paper's own randomized-alignment ablation (Sec. 2.1, Fig. 4) is the natural control for this, and it reports that randomly generated attention maps also reduce fairness gaps while degrading performance. That means the Aligner's fairness benefit may be a generic decorrelation/regularization effect rather than a consequence of encoding radiologist knowledge, exactly the alternative flagged in the Reader's weakest_assumption. Without a matched comparison of human-aligned versus random guidance under equal loss weight and training budget, the phrase 'incorporating human insights consistently reduces fairness gaps' is not supported. Separately, the headline statistic '27 out of 30 comparisons' is not reproducible from Table 2 as printed: counting the four fairness-gap columns across the six rows gives 18 of 24 directional improvements, and even including Hit Rate gives 24 of 30, not 27. Because this summary count is the paper's primary quantitative evidence, the 'consistently reduces fairness gaps' conclusion needs re-evaluation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies whether Human-AI alignment, implemented as supervision of a Vision Transformer's cross-attention maps with radiologist attention annotations (based on the authors' prior DWARF method), affects demographic fairness gaps in chest X-ray classification. It compares a no-alignment baseline (from KEVL, ref [30]) with an 'Aligner' model on three conditions and two out-of-domain datasets, across four fairness-gap metrics plus Hit Rate, with five repeats per experiment. The paper reports that alignment improves fairness in 27 of 30 comparisons, improves out-of-domain performance, helps in low-data regimes, and that random attention alignment also reduces fairness gaps while degrading performance. The central claim is that incorporating human insights 'consistently' reduces fairness gaps and improves generalization, with diminishing returns under excessive alignment.","tokens_in":8381,"tokens_out":6307,"duration_ms":57026,"significance":"The question is timely and the study is one of the first to connect explanation-guided learning to subgroup fairness in medical imaging. The empirical setup has clear strengths: multi-center training data (NIH, PadChest, VinDr, MIMIC), multiple OOD test sets, five repeats, a random-attention ablation, a low-data analysis, and public code. If the central claim were supported by a reproducible count and by statistical testing, the paper would be a solid contribution to both the fairness and explainability literatures. However, the headline '27/30' figure is not reproducible from Table 2, and the random-attention control suggests that the fairness benefit may not be specific to human knowledge. The significance is therefore conditional on a corrected and tightened analysis.","major_comments":[{"comment":"The claim that Human-AI alignment improved fairness in '27 out of 30 comparisons' is not reproducible from Table 2 as printed. Counting only the four fairness-gap columns (Accuracy, AUC, Sensitivity, F1) across the six dataset-by-demographic-group rows gives 18 improvements and 6 worsening; if Hit Rate is counted as a fifth column, the total is 24 improvements, not 27. Hit Rate measures attention alignment rather than fairness, so it should not be pooled with fairness-gap improvements. This count is the main quantitative support for the 'consistently reduces fairness gaps' conclusion, so the number and the conclusion must be corrected.","section":"Abstract and Section 3, Table 2"},{"comment":"The random-attention ablation is not a matched control for the attribution of the fairness effect to human insight. The text states only that random shapes at random locations are generated each epoch; it does not report the loss weight, the number or size of masks, or the training budget for the random condition, and Figure 4 presents results for the sex group only. Since random alignment also reduces fairness gaps, the current evidence cannot distinguish human semantic guidance from generic attention-map regularization or decorrelation. A matched comparison with equal loss weight and training budget, for both sex and age groups, is needed before the effect can be attributed to Human-AI alignment.","section":"Section 2.1, Figure 4"},{"comment":"No statistical significance tests are reported, although each experiment was repeated five times. Several printed means and standard deviations overlap substantially, e.g., CheXpert Pleural Effusion Gender AUC gap is 0.21±0.18 for w/o and 0.23±0.20 for Aligner, and JSRT Nodule Age accuracy gap is 16.11±2.42 versus 20.09±4.07. Without paired hypothesis tests or confidence intervals, the statement that alignment 'consistently reduces fairness gaps' (Abstract and Section 4) is stronger than the evidence supports. Figure 5 also reports low-data comparisons without visible or described error bars, making it impossible to assess whether the displayed differences are meaningful.","section":"Section 3, Tables and Figures"},{"comment":"Figure 3 and the surrounding text report that fairness improves up to 75% alignment but degrades at 100%, yet Table 2 labels the fully aligned model as 'Aligner' and summarizes it with a universal improvement count. The paper should state explicitly which alignment level is used in Table 2 and in Figures 5 and 6, and it should reconcile the non-monotonic trend in Figure 3 with the 'consistently reduces fairness gaps' phrasing used in the Abstract and Conclusion.","section":"Section 3, Figure 3"}],"minor_comments":[{"comment":"Equation (1) introduces α, ε, and w_FP, but only w_FP = 2.0 is reported; the smoothing-term values for α and ε should be given.","section":"Section 2.2, Equation (1)"},{"comment":"The Introduction cites [21] twice in the list '[5,6,7,21,26,28,21]', and Section 2.2 contains the typo 'derive the reader to [18]' instead of 'refer the reader to [18]'.","section":"Introduction and Section 2.2"},{"comment":"Figure 5 reports comparisons across four training-data ratios from repeated runs, but no error bars are visible or described in the caption, so the stability of the displayed differences cannot be assessed.","section":"Figure 5"},{"comment":"Figure 6 shows radar charts without numerical values or error bars, making the claim of 'considerable performance improvements' difficult to verify from the figure alone.","section":"Figure 6"},{"comment":"Table 1 prints 'JSR T' in the column heading, and CheXlocalize is listed as an additional attention evaluation set, but no attention-localization evaluation results are reported in the paper.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The main text relies heavily on the authors' own prior method (DWARF, ref [18]) for the alignment module and on KEVL (ref [30]) as the baseline; this is not a problem in itself, but the dependence on the authors' own tool is not flagged explicitly. The headline count error and the missing significance tests are the key substantive issues; I would not recommend acceptance until the numbers are corrected and the attribution to human insight is tested against a matched random control."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know up front. This is the first systematic study of whether explanation-guided / human-AI alignment training changes demographic fairness gaps in medical imaging, and that question is genuinely worth asking. The setup is substantial: three diseases, six public chest X-ray datasets, multiple alignment levels including a randomized-attention control, and both age and sex subgroups with a standard fairness-gap metric. If the effect holds, it's a useful training lever for groups working on fairness and robustness.\n\nThe problem is that the paper's own data are weaker than the abstract claims. The headline '27 out of 30 comparisons' doesn't reproduce from Table 2. Counting the four fairness-gap columns across the six dataset/subgroup rows gives 18 of 24 directional improvements; adding Hit Rate gives 24 of 30. So 'consistently reduces fairness gaps' is too strong, and the table-versus-abstract mismatch has to be fixed. No significance tests are reported anywhere, and with five repeats and overlapping error bars on several rows (JSRT nodule age accuracy, pleural effusion AUC), many of those 'improvements' could be noise.\n\nThe deeper issue is the randomized-attention control. The paper shows that random attention maps also reduce fairness gaps while degrading performance. That means the fairness benefit could be a generic decorrelation or regularization effect, not a consequence of encoding radiologist knowledge. The abstract says 'incorporating human insights consistently reduces fairness gaps,' but the authors' own control undermines the causal reading. A matched comparison with random guidance under equal loss weight and training budget would settle it.\n\nNone of this makes the paper worthless. The design is close to right, the authors ran the control that most papers skip, and the question is timely. But the claims need to be rebalanced, the counts need to be corrected, and statistical support is missing. I'd send this to peer review with a request for major revision, and I'd bring it to a reading group as a case study in how fairness interventions can be oversold. Not a desk reject, but nowhere near a clean accept.","headline":"First systematic study of human-AI alignment and fairness in chest X-ray, but the headline count doesn't reproduce and the random-attention control muddies the attribution; worth a careful revision, not a desk reject.","tokens_in":8892,"tokens_out":3767,"would_cite":false,"duration_ms":30699,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that steering a chest X-ray classifier's attention toward radiologist-annotated regions reduces fairness gaps across sex and age groups in most configurations, while also improving out-of-domain performance—provided the…","keywords":["Human-AI alignment","fairness","medical imaging","vision-language model","attention alignment","chest X-ray","out-of-domain generalization","explanation-guided learning"],"falsifier":"Train the same models with the alignment loss attached to random attention maps served at the same epochs, and compare fairness gains; if random maps produce the same fairness improvement as radiologist maps, then the content of the human guidance is not the active ingredient and the causal story fails. Alternatively, remove the attention loss and add an equivalent-magnitude regularizer such as dropout or weight decay; if fairness gains persist, the alignment mechanism itself is not responsible.","tokens_in":7977,"feed_emoji":"🩻","tokens_out":5567,"duration_ms":50539,"temperature":0.7,"pith_summary":"The paper asks whether teaching a chest X-ray classifier to look where radiologists look makes it fairer across demographic groups, not just more accurate. It reports that attention alignment reduces fairness gaps in 27 of 30 comparisons spanning three diseases and two demographic splits, and improves all four performance metrics on out-of-domain test sets. The benefit is strongest with partial alignment; full alignment can reverse gains. If true, this gives hospital AI developers a practical knob for fairness that does not require discarding accuracy.","feed_headline":"Radiologist-guided AI cuts sex and age bias in chest X-rays","feed_subtitle":"Adding radiologist attention maps improved fairness in 27 of 30 chest X-ray comparisons while boosting out-of-domain accuracy.","key_machinery":"The machinery is a pre-trained medical vision-language model whose visual and text embeddings are fused by cross-attention into per-class attention maps; an Attention Aligner head supervises these maps with a modified dice loss (equation 1) that only penalizes positive pixels and suppresses false positives, added to the standard cross-entropy classification loss. This alignment loss is the only difference between baseline and aligned models, so fairness changes are attributed to it.","core_discovery":"The central discovery is that human-AI alignment—placing a loss on the model's cross-attention maps so they overlap with radiologist-annotated regions—acts as a fairness intervention in chest X-ray classification. Across edema, pleural effusion, and nodule/mass detection, aligned models show smaller accuracy, AUC, sensitivity, and F1 gaps between sex and age subgroups than the same model trained without alignment, while simultaneously improving out-of-domain performance. The paper also identifies a boundary condition: alignment is beneficial up to a point, and 100% alignment can degrade gains; randomized alignment destroys performance but still narrows fairness gaps, indicating that part of the fairness effect is a decorrelation effect rather than purely semantic guidance.","pith_inferences":["Editorial inference: the same alignment mechanism might transfer beyond chest X-rays to other modalities with expert gaze maps, such as dermatology or pathology, where demographic skew also exists; the paper does not test this.","Editorial inference: because randomized alignment preserves fairness gains but kills performance, a practical extension would apply the alignment loss to a different layer or representation to isolate the generic regularizing component that drives the fairness benefit.","Editorial inference: a stricter multiple-comparison or permutation analysis across the 30 reported comparisons could reveal how many individual gains are statistically robust; the paper reports point estimates with error bars but no adjusted significance tests."],"forward_implications":["Fairness gaps over sex and age shrink under attention alignment in the large majority of tested configurations (27 of 30).","Out-of-domain performance rises on all four metrics for all three diseases, so fairness gains do not have to come at the cost of generalization.","Full 100% alignment can erode the gains, so the dose of human guidance should be tuned rather than maximized.","In low-data settings, alignment reduces fairness gaps more strongly, making it a candidate intervention when training data is scarce.","Randomized alignment lowers performance while also lowering fairness gaps, which suggests that part of the fairness effect is a decorrelation effect independent of the semantic content of the human guidance."],"supporting_citations":[{"why":"Supplies the pre-trained medical vision-language model and the attention-alignment loss formulation that the experiments vary.","marker":"[18]"},{"why":"Supplies the fairness gap metric used to define and measure demographic disparity.","marker":"[15]"},{"why":"Supplies the hit-rate metric used to measure how closely model attention matches human attention, and the CheXlocalize evaluation set.","marker":"[22]"},{"why":"Supplies the baseline model and the knowledge-enhanced text backbone used in both aligned and non-aligned conditions.","marker":"[30]"},{"why":"Supplies the NIH ChestX-Ray14 training data for nodule/mass, pleural effusion, and edema.","marker":"[27]"},{"why":"Supplies the MIMIC-CXR training data for pleural effusion and edema.","marker":"[13]"},{"why":"Supplies the VinDr-CXR training data with radiologist attention annotations.","marker":"[19]"},{"why":"Supplies the PadChest training data for nodule and mass detection.","marker":"[2]"},{"why":"Supplies the CheXpert data used for out-of-domain testing of edema and pleural effusion.","marker":"[12]"}],"fun_headline_variants":["Aligned AI reduces sex and age bias in chest X-rays","Radiologist-mapped AI: fairer and sturdier X-ray models","Expert alignment helps X-ray AI, but only up to a point","Even random alignment narrows X-ray bias gaps","Over-alignment with radiologists hurts X-ray AI performance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the cross-attention maps the model produces honestly reflect the features driving its classification, so penalizing those maps to match radiologist annotations changes the learned representation rather than merely redrawing a visualization; if that link fails, the fairness gains could come from generic regularization rather than human guidance.","fun_headline_variants_meta":{"raw":{"variants":["Aligned AI reduces sex and age bias in chest X-rays","Radiologist-mapped AI: fairer and sturdier X-ray models","Expert alignment helps X-ray AI, but only up to a point","Even random alignment narrows X-ray bias gaps","Over-alignment with radiologists hurts X-ray AI performance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000428,"raw_usage":{"total_tokens":2113,"prompt_tokens":789,"completion_tokens":1324,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":405,"completion_tokens_details":{"reasoning_tokens":1237}},"tokens_in":405,"tokens_out":1324,"duration_ms":12339,"temperature":1.0,"reasoning_tokens":1237,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:12:47.438252+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same models with the alignment loss attached to random attention maps served at the same epochs, and compare fairness gains; if random maps produce the same fairness improvement as radiologist maps, then the content of the human guidance is not the active ingredient and the causal story fails. Alternatively, remove the attention loss and add an equivalent-magnitude regularizer such as dropout or weight decay; if fairness gains persist, the alignment mechanism itself is not responsible.","supporting_citations":[{"cited_title":"Scientific Data9(1), 429 (2022)","cited_arxiv_id":null,"evidence_quote":"Supplies the VinDr-CXR training data with radiologist attention annotations."},{"cited_title":"In: International Conference on Medical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"Supplies the pre-trained medical vision-language model and the attention-alignment loss formulation that the experiments vary."},{"cited_title":"Nature Medicine30(4), 1166–1173 (2024)","cited_arxiv_id":null,"evidence_quote":"Supplies the fairness gap metric used to define and measure demographic disparity."},{"cited_title":"In: Proceedings of the IEEE CVPR","cited_arxiv_id":null,"evidence_quote":"Supplies the NIH ChestX-Ray14 training data for nodule/mass, pleural effusion, and edema."}],"review_version":1}