{"id":"d7a73b40-5fd8-4d4b-a855-ca7af7335aa9","arxiv_id":"2412.07991","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"dsLassoCov is a federated LASSO with unpenalized covariates; its main contribution is an efficient DataSHIELD implementation rather than a new statistical principle.","lead":"This paper introduces dsLassoCov, a federated version of LASSO regression that adjusts for covariates without pooling patient data. It aims to let hospitals and research cohorts run high-dimensional biomarker screening together while controlling confounders.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1's proximal update is algebraically wrong: it solves a different objective than Formulation (1), so the covariate-independence guarantee and simulation results are not tied to the stated method unless the code differs from the pseudocode.","rationale":"The load-bearing condition for the central claim is that the federated procedure actually converges to the minimizer of Formulation (1). The paper's own Supplementary Eq. (8) gives the correct soft-threshold update; Algorithm 1 line 8 gives a different update. This is an internal inconsistency, not a disagreement with external consensus. I checked the algebra on a scalar example: for z = S - G/L, the stated update is sign(z)*max(z + λ/L, 0), whereas the true prox is sign(z)*max(|z| - λ/L, 0). The two coincide only when z = 0. If the implementation follows the pseudocode, all simulation and real-data results are for a penalized estimator with a misspecified proximal step, and the theoretical equivalence in the Covariate control mechanism section does not apply. If the implementation follows Eq. (8), then the paper's algorithm box is wrong in a way that prevents replication. Either way the manuscript as written cannot support the central claim until this is resolved. I did not choose the linearity assumption as the primary concern because the paper explicitly states it as an assumption and lists it as a limitation; it is a scope condition, not an internal inconsistency. The proximal update error is more fundamental because it undermines the regression core even when linearity holds. A quick code check can settle which variant is actually implemented.","tokens_in":21085,"tokens_out":8158,"duration_ms":81815,"concrete_test":"Take a small dense test case (e.g., n=50, p=8, one covariate, known sparse w). Compute the exact solution of Formulation (1) by coordinate descent (glmnet with penalty.factor=0 on the covariate column) and compare it to (a) the proximal gradient update exactly as written in Algorithm 1 line 8, and (b) the corrected operator from Supplementary Eq. (8). Also run the dsLassoCov code from the linked GitHub repository on the same data. If output (a) or the repository output differs from glmnet, the algorithm as described (or as implemented) is not solving the stated problem and the simulation results must be regenerated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Algorithm 1 line 8 does not implement the proximal update that Formulation (1) requires. For the scalar subproblem in Supplementary Eq. (7), the minimizer is the soft-threshold operator w_j^+ = sign(w̃_j) max(|w̃_j| - λ/L, 0), with w̃_j = S_j - G_w,j/L. Algorithm 1 instead writes max{S_i - (G_w - λ)/L, 0} × sign(S_i - G_w/L). Algebraically this is not equivalent: for a scalar with S=1, G=0, λ=0.1, L=1, the stated rule gives 1.1, whereas the correct proximal step gives 0.9; negative values are also mishandled. Consequently, if the shipped code follows the pseudocode, dsLassoCov is not minimizing the claimed LASSO-with-unpenalized-covariates objective, and the covariate-independence argument in Eq. (2) does not apply to the computed solution. If the code is correct, the pseudocode is internally inconsistent with Eq. (8) and must be corrected. Either way, the reported simulation accuracies and equivalences cannot be verified from the manuscript as written.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces dsLassoCov, a federated learning method that fits a LASSO-type linear model with an additional set of unpenalized covariates for confounder control, and applies it to high-dimensional biomedical data in the DataSHIELD environment. The proposed objective is the standard penalized regression with an unpenalized covariate block, and the optimization is a federated proximal-gradient method using exact aggregated gradients and a regularization path with warm starts. The authors provide a profiling argument that the solution is equivalent to minimizing the distance between outcome residuals and prediction residuals after removing linear covariate associations. The empirical part compares dsLassoCov with a per-feature residualization baseline ('ds.glm + ds.Lasso') and local meta-analysis methods in simulations, and reports a replication of a HELIX exposome analysis of childhood hypertension.","tokens_in":21296,"tokens_out":6205,"duration_ms":68358,"significance":"If the implementation were verified to solve the stated objective, the paper would be a useful practical contribution: it packages a known statistical construction (LASSO with unpenalized covariates, equivalent to residualization in the linear case) into a communication-efficient federated algorithm, reports run-time gains over per-feature residualization, and provides an open-source DataSHIELD implementation. The mathematical core is not deep, but the engineering and application value to exposome consortia is real. The paper also ships concrete code repositories and a tutorial, which is commendable. However, as written, the supporting evidence cannot be fully trusted because the central pseudocode appears to implement a different proximal update than the one derived in the supplement, and the simulation data-generation equations are dimensionally inconsistent.","major_comments":[{"comment":"Line 8 of Algorithm 1 does not implement the soft-threshold update derived in Supplementary Eq. (8). For z = S_i - G_w/L, the correct proximal update is sign(z) max(|z| - λ/L, 0), equivalently max(|z| - λ/L, 0) * sign(z). The pseudocode instead computes max{S_i - (G_w - λ)/L, 0} * sign(S_i - G_w/L). These differ: for the scalar example S_i=1, G_w=0, λ=0.1, L=1, the correct update gives 0.9 while the pseudocode gives 1.1, and negative arguments are also handled incorrectly. Consequently, if the shipped code follows the pseudocode, the algorithm is not minimizing the stated objective and the covariate-independence argument in Eq. (2) does not apply to the computed solution; if the code is correct, the pseudocode is inconsistent with Eq. (8). Either way, the reported simulation results cannot be verified from the manuscript as written. Please correct the pseudocode and provide a minimal numerical check against a standard solver.","section":"Supplementary Methods, Algorithm 1, line 8"},{"comment":"The data-generation equation is dimensionally inconsistent. After defining Xsig as n×p_t and Xc as n×p_c, the expression Y = [Xsig Xc a] [w; 0_c] does not conform: the scalar intensity a cannot be appended as a column to the design matrix, and [w; 0_c] has length p_t + p_c, not p_t + p_c + 1. The intended generative model presumably includes the confounding vector c ∈ R^{n×1} multiplied by the scalar a, but this is not what is written. Since every simulation in the paper depends on this generative scheme, please rewrite Eqs. (11)-(12) with explicit dimensions and state the exact R code used, so the simulations are reproducible.","section":"Supplementary Methods, 'Confounded data simulation', Eq. (12)"},{"comment":"The paper claims that dsLassoCov is 'mathematically equivalent to an iterative procedure that alternates between a covariate removal step and a parameter update step,' and that this 'guarantees the estimation of covariate-free parameters at each iteration.' The derivation in Eq. (2) is a static profiling identity at the optimum of the joint objective, not a property of the iterates of Algorithm 1. In Algorithm 1, the covariate coefficients w(c) are updated by a gradient step (line 9), not by residualization or by solving the closed-form conditional minimizer. If the per-iteration equivalence claim is intended, a proof should be supplied; otherwise the claim should be revised to describe the profiled objective and its solution.","section":"Methods, 'Covariate control mechanism' and Discussion, first paragraph"},{"comment":"The text and Figure 6 caption state that selected exposures are those with odds ratio below 0.95 or above 1.05, but Table S3 lists many exposures with odds ratios within [0.95, 1.05], for example Cobalt 0.969, PBDE 153 0.970, Lead 0.978, DMTP 0.985, and MEHP 0.994. The selection criterion is therefore not reproduced by the displayed results. Please define the exact selection rule (including how confounders are treated), reconcile Table S3 with Figure 6, and report the confidence intervals or stability of the selected set if a threshold is used.","section":"Results, 'Real data analysis'; Table S3; Figure 6 caption"}],"minor_comments":[{"comment":"There are recurring language and typographical issues: 'this allow' should be 'this allows', 'odd ratio' should be 'odds ratio', 'refered' should be 'referred', 'counfounding' should be 'confounding', and 'fedeated' should be 'federated'.","section":"Abstract and throughout"},{"comment":"The supplementary text contains duplicated paragraphs: the 'Read data analysis' section repeats the HELIX description, and the 'Data cohorts and preprocessing' section contains two nearly identical paragraphs describing the exposure categories. Please remove the duplicates.","section":"Supplementary Methods, 'Read data analysis' and 'Data cohorts and preprocessing'"},{"comment":"The line-search candidate sequence is given as {L_{i-1}, 2L_{i-1}, 4L_{i-1}, 16L_{i-1}, ...}; the jump from 4 to 16 is presumably a typo for 8, and the Armijo-Goldstein condition itself is not stated explicitly. Please clarify the line-search rule and the stopping criterion used in Algorithm 1 and Algorithm 2.","section":"Supplementary Methods, Algorithm 1, line 5"},{"comment":"The displayed equation 'λ_p×1 − 1/n | x^T y − x^T x_c (x_c^T x_c)^{-1} x_c^T y / σ | = 0' is not a well-formed mathematical statement. The final formula for λ_max is correct in substance, but the intermediate line should be rewritten componentwise so that the role of σ ∈ [-1,1] is clear.","section":"Supplementary Methods, 'The non-federated training protocol', lambda_max derivation"},{"comment":"The classification extension is asserted by saying that 'one can derive all equations ... by switching to the logistic loss', but no logistic objective or proximal update is given. Since classification simulations are reported, please add the explicit classification objective and confirm that the same exact-gradient aggregation and line search apply unchanged.","section":"Methods, 'Modeling, optimization and algorithm'"}],"recommendation":"major_revision","confidential_remarks":"My main concern is verifiability rather than the statistical idea. The correction of Algorithm 1 and the simulation equations is essential before the empirical claims can be evaluated. I would ask the authors to include a small unit test demonstrating that the federated dsLassoCov solution matches a centralized solver (e.g., glmnet with unpenalized covariates) on a fixed simulated dataset; this would settle whether the code or the pseudocode is the source of the discrepancy."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe useful part of this paper is the engineering. They built a DataSHIELD implementation of proximal-gradient LASSO that keeps unpenalized covariates in the objective, aggregates exact gradients, warm-starts along a lambda path, and demonstrates it on six real HELIX servers. That is a real contribution to applied federated learning for exposome consortia. The covariate-control mechanism itself is not new: as the authors note, it is LASSO with unpenalized covariates, and their Eq. (2) derivation is essentially Frisch-Waugh-Lovell. The “first” and “only” claims in the Discussion should be toned down.\n\nThe math exposition is mostly standard, but one load-bearing piece is wrong. Algorithm 1 line 8 writes the update as max{S_i - (G_w - λ)/L, 0} × sign(S_i - G_w/L), which is not the soft-threshold operator derived in Supplementary Eqs. (7)-(8). For a scalar with S=1, G=0, λ=0.1, L=1, the algorithm gives 1.1, while the correct proximal step gives 0.9. Negative values are also mishandled. If the shipped code follows the pseudocode, the method is not minimizing Formulation (1); if the code is correct, the pseudocode is inconsistent with the paper’s own derivation. Either way, the current manuscript cannot be verified as written.\n\nThe supporting evidence also has soft spots. Supplementary Eq. (12) is dimensionally inconsistent: [Xsig Xc a] cannot multiply [w; 0c] as written, and the missing term is likely the confounding-effect vector c. The simulation results appear to lack error bars or repeated-seed variability, so the efficiency and accuracy claims are hard to assess. In the real-data analysis, the federated model is compared to a local model but not to the “ds.glm + ds.Lasso” baseline that is the paper’s main efficiency comparator, and the 13 confounders chosen from a DAG are treated as sufficient without sensitivity analysis.\n\nEven so, this paper deserves a serious referee. The application is timely, the implementation is shipped, and the real-data demonstration across six distributed servers is a useful existence proof. I would send it for peer review with a clear request for major revision: fix Algorithm 1, correct the data-generation equation, report variance across simulation runs, and rewrite the novelty claims. The readership is specialists in federated learning and exposome epidemiology, not a broad ML audience.","headline":"A useful federated Lasso implementation is undercut by an algebraically wrong proximal update in the pseudocode, plus soft simulation evidence, though the real-data deployment is a genuine practical contribution worth refereeing.","tokens_in":21904,"tokens_out":3208,"would_cite":true,"duration_ms":34337,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62J07"],"pacs":[],"model":"deepseek-v4-flash","headline":"dsLassoCov is claimed to be the first federated learning algorithm that controls for covariate effects in high-dimensional linear models while exchanging only aggregated gradients and losses.","keywords":["federated learning","LASSO","covariate adjustment","confounding","feature selection","exposome","privacy-preserving analysis","high-dimensional regression"],"falsifier":"Generate simulated data in which a confounder affects the outcome nonlinearly (for example, quadratically or through a threshold) while also shifting a subset of features, run dsLassoCov and a correctly adjusted comparator, and compare feature-selection accuracy; if dsLassoCov's accuracy drops materially, the linear-covariate assumption is the point of failure.","tokens_in":20880,"feed_emoji":"🩺","tokens_out":8705,"duration_ms":84260,"temperature":0.7,"pith_summary":"Federated learning trains models across distributed databases by exchanging model updates rather than raw data, and the open problem is controlling for confounders in that setting. dsLassoCov is a federated implementation of LASSO that adjusts for covariates while the model trains, so data owners never need to run per-feature residualizations or reveal individual-level data. The paper's central claim is that minimizing a penalized loss over feature coefficients and unpenalized covariate coefficients yields feature coefficients that are independent of linear associations with the covariates. Because the covariate block is profiled out algebraically, the optimizer can be federated by exchanging only gradients and loss values, keeping communication cost low. Simulations and a six-cohort exposome analysis, which reproduces previously reported hypertension associations, support the method's efficiency and feature-selection accuracy.","feed_headline":"Federated LASSO controls confounders without sharing data","feed_subtitle":"One algorithm adjusts for covariates during training and matches or beats per-feature residualization baselines.","key_machinery":"The central mechanism is the profiled covariate objective: because the unpenalized covariate block is linear, the optimal covariate coefficients are a closed-form least-squares projection of the residual $y - Xw$ onto the covariate space. Plugging that solution into formulation (1) yields formulation (2), in which the loss is the squared distance between the outcome and the prediction, each with linear covariate associations removed. This identity carries the covariate-control claim, since it makes the feature solution independent of covariate associations; it also dictates the federated protocol, where each server computes local gradients and losses of that same objective and the client aggregates them with weights proportional to local sample sizes. The same profiled objective extends to classification by replacing squared error with logistic loss.","core_discovery":"On the paper's own terms, the discovery is a mathematical reduction plus a distributed implementation. For the objective $\\min_{w,w^{(c)}} \\frac{1}{2n}\\sum_i (y_i - x_i w - x_i^{(c)} w^{(c)})^2 + \\lambda|w|$, the first-order condition for $w^{(c)}$ gives $w^{(c)} = (X^{(c)T}X^{(c)})^{-1}X^{(c)T}(y - Xw)$. Substituting this back turns the objective into a comparison between the covariate-residualized outcome and the covariate-residualized prediction, so the optimal $w$ is the same as if the covariates had been regressed out of both sides. The paper calls this a double-machine-learning-like debiased property and notes that a similar device appears in double-LASSO. The federated algorithm then needs only shared gradients, losses, and sample counts, not the data themselves. In simulation, the method matches the conventional federated residualization pipeline on regression feature selection and beats it in classification, where residualizing a binary outcome is not possible; in the real-world exposome application it selects exposures whose odds ratios align with earlier studies.","pith_inferences":["A natural extension the paper does not develop: the same profiling trick could be applied to other penalties such as ridge or elastic net, but the linear-in-covariates form of the inner adjustment would need to change before those extensions inherit the current guarantees.","If confounders act nonlinearly, the profiled linear term will only approximate the true confounding surface; sensitivity analyses varying the covariate set or adding nonlinear terms would be needed to know how much residual confounding remains in real data.","The communication advantage is largest when the number of covariates is small relative to the number of features; when covariates are numerous, the closed-form projection may need to be computed or approximated carefully in the federated protocol.","A testable consequence: on a cohort with known technical batch structure, replacing per-feature residualization with dsLassoCov should produce nearly identical marginal effect estimates while cutting wall-clock time, a prediction that can be checked directly on existing consortium data."],"forward_implications":["In high-dimensional federated biomarker screens, dsLassoCov removes the need for one generalized-linear-model fit per feature to residualize covariates, which is the main cost of the conventional pipeline.","The method supplies a principled covariate adjustment for federated classification, a setting where residualizing a binary outcome is not well-defined.","Simulation results support that, as the number of features grows, dsLassoCov keeps runtime lower than the conventional federated approach while retaining at least equal feature-selection accuracy in regression and higher accuracy in classification.","The real-data replication suggests that federated selection of early-life exposures for hypertension is stable enough to reproduce previously reported odds ratios across six distinct databases.","Adoption could lower the communication ceiling for consortium-wide exposome and omics analyses, since per-iteration exchanged messages do not scale with the number of residualization regressions."],"supporting_citations":[{"why":"establishes the low-dimensional equivalence of regression adjustment and residualization that the covariate-control reduction relies on.","marker":"[17]"},{"why":"defines the LASSO penalty and its shrinkage and selection behavior that dsLassoCov extends to the federated setting.","marker":"[22]"},{"why":"supplies the regularization-path and tuning strategy used to select the lambda sequence.","marker":"[23]"},{"why":"provides the double/debiased machine-learning perspective the paper invokes for the residualized objective's debiased property.","marker":"[25]"},{"why":"supplies the double-LASSO covariate-control formulation that the paper adapts.","marker":"[26]"},{"why":"provides the privacy-preserving federated architecture in which the server-side functions are executed.","marker":"[8]"},{"why":"supplies the six-cohort birth-cohort exposome data used in the real-data analysis.","marker":"[18]"},{"why":"supplies the multi-task federated learning framework from which the optimizer and regularization-path protocol are drawn.","marker":"[19]"},{"why":"reports the earlier urban-environment blood-pressure associations the real-data results are said to replicate.","marker":"[27]"}],"fun_headline_variants":["Federated LASSO with built-in covariate control","Covariate control in federated learning without sharing data","Privacy-preserving LASSO that adjusts for confounders","dsLassoCov: efficient federated covariate adjustment","Federated learning that handles confounders analytically"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the included covariates capture all confounding and act linearly on the outcome, because the derivation that the feature solution is independent of covariate associations holds only for that linear model.","fun_headline_variants_meta":{"raw":{"variants":["Federated LASSO with built-in covariate control","Covariate control in federated learning without sharing data","Privacy-preserving LASSO that adjusts for confounders","dsLassoCov: efficient federated covariate adjustment","Federated learning that handles confounders analytically"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1408,"prompt_tokens":991,"completion_tokens":417,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":336}},"tokens_in":607,"tokens_out":417,"duration_ms":4312,"temperature":1.0,"reasoning_tokens":336,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:20:17.262119+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate simulated data in which a confounder affects the outcome nonlinearly (for example, quadratically or through a threshold) while also shifting a subset of features, run dsLassoCov and a correctly adjusted comparator, and compare feature-selection accuracy; if dsLassoCov's accuracy drops materially, the linear-covariate assumption is the point of failure.","supporting_citations":[],"review_version":1}