{"id":"1a7a2ca4-043c-4565-8daf-cc3f1710a15d","arxiv_id":"2508.00930","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":1,"one_line_summary":"A local Hi-Fi framework assigns unique, redundant, and synergistic importance scores to individual data points and reports synergistic O3 and NO2 associations with Alzheimer mortality.","lead":"This paper extends Hi-Fi feature importance to individual data points, splitting each feature's contribution into unique, redundant, and synergistic scores, and applies it to air pollutants and Alzheimer's disease mortality in Italian provinces. A generalist reader might care because it offers a new way to see when combinations of pollutants, not just single pollutants, are associated with health outcomes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Per-pattern variance-based synergy scores are model-relative and locally underdetermined; without a specified local reference distribution or a synthetic no-synergy control, the O3/NO2 and green-area synergy claims may reflect the regression model rather than the data.","rationale":"The reader's verdict is UNVERDICTED, and I agree that the manuscript cannot currently be assessed because the full text is unreadable. My stress-test identifies a more specific version of the reader's weakest assumption: the method attributes variance-based synergy scores to individual patterns, yet per-instance variance is not well-defined without an explicit local conditional distribution over features. This is not merely a matter of empirical model misspecification; it is an identifiability issue in the proposed decomposition itself. The abstract and the corrupted text do not disclose how the local distribution is constructed, how the LOCO baseline is estimated, or how the regression model is specified and validated. Therefore the claimed synergistic associations between O3 and NO2, and between green areas and pollutants, are currently unsupported by publicly readable evidence. My concern does not move the verdict away from UNVERDICTED, because the appropriate status remains 'not enough information to judge'; acceptance should be withheld until the method is specified and the synthetic control is run. For the empirical claim, the decisive check is the no-synergy synthetic control: if the method reports synergy on additive data, the central application-level claim collapses into a model artifact. For the methodological claim, the decisive check is stability of per-pattern scores under two different local conditional distributions. I set agreement_with_reader to 'partial' because the reader emphasized model misspecification, while my concern is more sharply about the underdetermination of per-pattern variance without a local reference distribution, although the two are closely related and the reader did mention the LOCO baseline as a potential source of artifacts.","tokens_in":8149,"tokens_out":4310,"duration_ms":55148,"concrete_test":"Obtain the manuscript's code and data (or, failing that, run a synthetic replication) and perform a no-synergy control: generate y = a*x1 + b*x2 + epsilon with independent x1, x2, and also with correlated x1, x2, fit the same regression family used for AD mortality, and compute per-pattern Hi-Fi synergy scores. If mean synergy is significantly nonzero, or if it varies substantially across plausible model classes (linear, GAM, random forest) or across LOCO baselines, the reported O3/NO2 and green-area synergy cannot be distinguished from model artifacts. Additionally, require a stated local conditional distribution: repeat the per-pattern decomposition under two defensible choices (e.g., k-NN empirical conditional versus Gaussian copula); if the scores change materially, the decomposition is not a stable property of the data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that per-pattern Hi-Fi scores measure real cooperative effects in the data. The abstract defines these scores through a variance-based LOCO decomposition, but for a single pattern (one observation), 'variance' is not defined without a distribution over feature values conditioned on that pattern, and the LOCO baseline is not specified (e.g., which estimator, which resampling scheme). If the local distribution is chosen ad hoc, the unique/redundant/synergistic split is not identifiable, and synergy scores can be manufactured by the model class: a model with interaction terms can produce synergy even when the data-generating process is additive, and a misspecified additive model can suppress true synergy. The reader's weakest assumption is therefore load-bearing: the regression model and the local reference distribution must be specified, validated, and shown not to drive the reported O3-NO2 and green-area synergy. No model specification, validation, or uncertainty quantification is disclosed in the abstract, and the supplied full text is unreadable, so this condition is currently unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a per-pattern extension of the Hi-Fi (high-order feature importance) framework, in which each input feature is assigned unique, redundant, and synergistic importance scores for individual data points, and it compares these scores with Shapley-effect feature importance. The method is applied to an ecological One-Health dataset linking air pollutants (O3, NO2), urban green density, and Alzheimer's disease mortality, with the reported main finding being a synergistic association between O3 and NO2 with mortality, especially in the Italian provinces of Bergamo and Brescia, and a synergistic influence of urban green density with pollutants. The abstract frames local Hi-Fi as a promising, widely applicable XAI tool.","tokens_in":8350,"tokens_out":3419,"duration_ms":40567,"significance":"If the per-pattern variance-based decomposition is theoretically justified and empirically validated, the framework could be a useful addition to explainable AI, particularly for uncovering higher-order interactions at the level of individual predictions. The One-Health application is topical and could be of broad interest. However, the current manuscript does not make this case accessible: the full text after the abstract is unreadable due to encoding corruption, and the abstract alone supplies no model specification, no local reference distribution, no out-of-sample or synthetic validation, and no uncertainty quantification. The significance is therefore entirely conditional on details that the manuscript does not presently make verifiable.","major_comments":[{"comment":"The submitted full text is unreadable: after the abstract the characters are garbled, so the mathematical definitions of the per-pattern Hi-Fi scores, the LOCO estimator, and the empirical analysis cannot be checked. This must be corrected in a resubmission; without a readable manuscript no technical claim can be verified.","section":"Full text (after abstract)"},{"comment":"The central claim of synergistic O3/NO2 and green-area effects is model-relative: the scores are computed from a fitted regression model for AD mortality, but the abstract reports no model class, no hyperparameters, no validation, and no uncertainty quantification. Because a model with interaction terms can manufacture synergy even under an additive data-generating process, the authors must report the model specification and an out-of-sample or synthetic benchmark before interpreting the reported associations.","section":"Abstract"},{"comment":"For a single pattern (one observation), a variance-based decomposition requires a distribution over feature values conditioned on that pattern; the abstract never specifies this local reference distribution or the LOCO baseline (estimator, resampling scheme). Without a stated and justified local distribution, the unique/redundant/synergistic split is not identifiable. The authors should define the local distribution explicitly and show that the reported synergy is robust to its choice.","section":"Abstract"},{"comment":"The application is an ecological, observational analysis at province level, but the abstract gives no adjustment for potential confounders (e.g., socioeconomic status, healthcare access, smoking) and no measure of uncertainty for the reported province-specific results. The statement 'especially in the provinces of Bergamo and Brescia' requires at least interval estimates or a multiple-comparison-aware analysis.","section":"Abstract"}],"minor_comments":[{"comment":"The phrase 'Bergamo e Brescia' should be 'Bergamo and Brescia' in an English-language manuscript, and the disease name should be used consistently as 'Alzheimer's disease'.","section":"Abstract"},{"comment":"The terms 'Hi-Fi' and 'Shapley effect' are used without citations; the manuscript should give references for the baseline method and the comparison metric.","section":"Abstract"},{"comment":"The manuscript does not state data sources or availability; since the application is a key selling point, a reproducibility statement is needed.","section":"Abstract"},{"comment":"The abstract refers to a recently proposed adaptive version of LOCO, but the citation is not visible in the readable portion; ensure the reference is explicit in the resubmission.","section":"Full text"}],"recommendation":"major_revision","confidential_remarks":"The full text is unreadable, which may be an encoding error in the submission. If so, please ask the authors for a clean PDF or source file before review. The novelty relative to existing Hi-Fi and Shapley-effect literature is not assessable from the abstract alone, and the current manuscript would need substantial additions (model specification, local reference distribution, synthetic controls, uncertainty quantification) before a fair evaluation is possible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe one thing you should know: the full text in the copy I was given is corrupted beyond use—mostly replacement characters. So anything said about the method and experiments is built from the abstract, which is coherent but does not answer the question that matters.\n\nWhat is new: assigning per-instance unique/redundant/synergistic scores is a natural next step after the global Hi-Fi framework. If done right, it would let you say, for a single prediction, which features carry unique information, which are redundant, and which act synergistically. That is genuinely useful for explainable AI. Benchmarking against the Shapley effect is the right sanity check.\n\nWhere it gets shaky: for one observation, variance is not defined. You need a distribution over feature values conditioned on that pattern, or an explicitly specified local LOCO reference scheme, to compute the decomposition. The abstract does not specify it. If that choice is ad hoc, the unique/redundant/synergistic split is not identifiable. Worse, a model with interaction terms will produce synergy even when the true data-generating process is additive, and a misspecified additive model can suppress real synergy. So the O3-NO2 and green-area synergy claims are model-relative until shown otherwise. That is not fatal in principle—it is standard for feature attribution—but it demands a synthetic control with known no-synergy structure, and I cannot see one in the visible text.\n\nThe Alzheimer-mortality application is an ecological observational association, so it carries the usual confounding caveats. The abstract highlights specific provinces without presenting uncertainty intervals. That reads as a selected finding, but it is not damning on its own.\n\nWhat the paper does well: the authors are explicit that they build on prior work, and the application is a legitimate stress test for a local interaction score. The abstract does not contradict itself.\n\nBottom line: this is a paper for the explainable-AI community and for people working on high-order interactions in health data. It deserves a serious referee, because the methodological claim is substantive. But the referee's first job is to find the definition of the local reference distribution and to check for synthetic no-synergy or out-of-sample validation. Without those, the headline synergy is not supported. My verdict is unverdictable rather than reject, because the idea may be fine once the full text is readable.\n\nRecommendation: send it to peer review. If the next version is readable and answers the identifiability question, it will have a fair chance.","headline":"A plausible per-instance extension of Hi-Fi that I cannot assess because the full text is unreadable; the abstract alone does not answer the key identifiability question about the local reference distribution.","tokens_in":8872,"tokens_out":2251,"would_cite":false,"duration_ms":27509,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Per-case feature scores reveal O3-NO2 synergy in Alzheimer's deaths","keywords":["feature importance","explainable AI","synergy","redundancy","partial information decomposition","Leave One Covariate Out","air pollution","Alzheimer disease"],"falsifier":"Recompute the local Hi-Fi decomposition on the same data with a different well-validated regression model (for example, a regularized linear model or a tree ensemble), and also on a control data set where $O_3$ and $NO_2$ values are independently shuffled. If the strong synergistic scores in Bergamo and Brescia disappear or appear in the shuffled control, the claimed $O_3$-$NO_2$ synergy is a property of the estimator rather than of the underlying relationship.","tokens_in":7985,"feed_emoji":"🧠","tokens_out":5677,"duration_ms":61568,"temperature":0.7,"pith_summary":"The paper's aim is to turn feature importance for regression models from a single global score per feature into a per-pattern decomposition: for each individual data point, each feature gets a unique (two-body) contribution, plus redundant and synergistic higher-order contributions. The framework is built on an adaptive version of the Leave One Covariate Out (LOCO) measure, called Hi-Fi, and is compared with the Shapley effect. Applied to a One-Health question, the method predicts Alzheimer's disease mortality from air pollutants and urban green-area density; the central finding is a synergistic association between $O_3$ and $NO_2$ with mortality, strongest in the provinces of Bergamo and Brescia, and a synergistic influence of green-area density with pollutants. A sympathetic reader would care because standard importance tools report only isolated feature relevance, whereas this framework claims to expose cooperative effects that only appear when features act together, at the resolution of single predictions.","feed_headline":"Per-case feature scores reveal O3-NO2 synergy in Alzheimer's deaths","feed_subtitle":"Per-case scores separate unique, redundant, and synergistic pollutant effects on Alzheimer's mortality.","key_machinery":"The central object is the local (per-pattern) Hi-Fi decomposition, an adaptive extension of the Leave One Covariate Out (LOCO) measure. For each data point, it uses a variance-based partial information decomposition to split a feature's importance into a unique two-body term, a redundant term, and a synergistic higher-order term. This decomposition carries the argument because it is what turns 'this feature matters' into a statement about how features matter together, and it is the object whose values identify Bergamo and Brescia as sites of $O_3$-$NO_2$ synergy and green areas as synergistic contributors.","core_discovery":"The paper claims that cooperative effects in feature importance can be measured for individual patterns, not only averaged over a data set. In the authors' formulation, the predictive contribution of a feature in a regression model splits into three components, unique, redundant, and synergistic, and this split is computed locally, pattern by pattern. The empirical result in the Alzheimer's disease application is that ozone ($O_3$) and nitrogen dioxide ($NO_2$) act synergistically in predicting mortality, particularly in the provinces of Bergamo and Brescia, and that the density of urban green areas also participates synergistically with the pollutants. These claims are established by comparing the local Hi-Fi scores with the global Shapley effect and by inspecting which provinces carry the synergistic component.","pith_inferences":["If the local synergy scores survive model resampling and cross-validation, one testable extension is to use them as a screening map: provinces with consistently high per-pattern synergy would be candidates for targeted joint-pollutant monitoring.","The same per-pattern decomposition could be applied with a different importance baseline, checking whether the reported $O_3$-$NO_2$ synergy is robust to the choice of LOCO or is a property of the data-generating process.","A natural next experiment is to feed the local scores into a spatial model and ask whether high-synergy provinces cluster geographically or follow pollution transport patterns; the paper does not do this."],"forward_implications":["Each feature in a regression model can be reported as three per-instance scores (unique, redundant, synergistic) rather than one global importance, giving explanations at the level of individual predictions.","$O_3$ and $NO_2$ should be treated as a pair in future analyses of air pollution and Alzheimer's disease mortality, since their joint contribution exceeds what either feature alone explains.","Urban green-area density enters the prediction of Alzheimer's mortality by modulating pollutant effects, not just as an independent protective factor.","Local Hi-Fi can be transferred to other regression settings that need high-order cooperative effects, such as biomarker panels or environmental exposure mixtures."],"supporting_citations":[],"fun_headline_variants":["Local feature scores expose O3-NO2 synergy in Alzheimer's mortality","Per-case importance splits reveal O3-NO2 synergy in AD mortality","O3 and NO2 per-case synergy predicts Alzheimer's mortality","Pattern-level scores show O3-NO2 cooperative effect on AD mortality","Green areas boost O3-NO2 synergy in per-case Alzheimer's risk"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the regression model predicting Alzheimer's disease mortality is well specified, so that decomposing its per-pattern predictions tells us about real cooperative effects in the health and environmental data rather than about artifacts of the model or the LOCO baseline.","fun_headline_variants_meta":{"raw":{"variants":["Local feature scores expose O3-NO2 synergy in Alzheimer's mortality","Per-case importance splits reveal O3-NO2 synergy in AD mortality","O3 and NO2 per-case synergy predicts Alzheimer's mortality","Pattern-level scores show O3-NO2 cooperative effect on AD mortality","Green areas boost O3-NO2 synergy in per-case Alzheimer's risk"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000623,"raw_usage":{"total_tokens":2890,"prompt_tokens":954,"completion_tokens":1936,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":570,"completion_tokens_details":{"reasoning_tokens":1855}},"tokens_in":570,"tokens_out":1936,"duration_ms":14657,"temperature":1.0,"reasoning_tokens":1855,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T11:22:00.161855+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the local Hi-Fi decomposition on the same data with a different well-validated regression model (for example, a regularized linear model or a tree ensemble), and also on a control data set where $O_3$ and $NO_2$ values are independently shuffled. If the strong synergistic scores in Bergamo and Brescia disappear or appear in the shuffled control, the claimed $O_3$-$NO_2$ synergy is a property of the estimator rather than of the underlying relationship.","supporting_citations":[],"review_version":1}