{"id":"082a545c-ad27-434d-ad12-c6599fb68385","arxiv_id":"2502.01382","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An electrode-placement algorithm that treats neural response as thresholded reduces simulated off-target stimulation by up to 60 percent compared with least-squares montage design.","lead":"HingePlace is a new algorithm for placing scalp electrodes in transcranial electrical stimulation, and it deliberately allows small electric fields outside the target region instead of driving them to zero. In two computer models of the brain, this approach reduced the volume or count of off-target neural activation by up to 60 percent compared with traditional placement algorithms, which could make non-invasive brain stimulation more precise.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported focality gains may be inflated by tuning HingePlace's tolerance hyperparameters on the evaluation metric itself; the 60-75% reductions lack held-out or error-bar evidence.","rationale":"The paper makes a genuine and interesting algorithmic contribution: the unification theorem (Theorem 1) and the idea of replacing a quadratic off-target penalty with a threshold-ignoring hinge loss are defensible, and the sea-of-neurons platform is a substantially more realistic test than the VTh metric alone. I read the paper in good faith and attempted to identify the weakest point in the central claim that HingePlace produces more focal neural responses than LCMV-E. The reader's weakest_assumption focuses on the biological irrelevance of sub-threshold fields; that is a legitimate concern for translation, but the sea-of-neurons experiments directly count spikes, so the paper does provide some evidence for the thresholding mechanism in a biophysical model. The more load-bearing vulnerability is that the quantitative magnitude of the reported gains is not robustly estimated. The tolerance vectors are tuned on the same evaluation metric, with 30 random draws and re-tuning when performance was poor, and no uncertainty quantification is reported. This makes the 60-75% figures untrustworthy as point estimates; they could be substantially inflated by selection bias. The proposed concrete test—held-out evaluation after tuning only on a separate subset, plus a fixed physiologically-derived tolerance—would settle whether the algorithm's advantage persists without per-run fitting to the test metric. Since the reader already reached CONDITIONAL based on related but not identical concerns, my read does not change the verdict; it sharpens the condition under which the central claim should be accepted. I therefore assign UNCHANGED rather than a new verdict class, and partial agreement with the reader's weakest_assumption because I identify the evaluation-protocol issue as the weakest load-bearing assumption.","tokens_in":48242,"tokens_out":5718,"duration_ms":55964,"concrete_test":"Re-run the Sec. 5.2.3 optimal-direction sea-of-neurons study with a pre-registered evaluation protocol: (i) randomly split the 800 neuron locations into a tuning set (400) and a held-out test set (400); (ii) tune [Ex_tol, Ey_tol, Ez_tol] using the paper's 30-point random search on the tuning set only; (iii) freeze the tolerances and evaluate Nact on the held-out test set; (iv) compute the relative decrease versus LCMV-E and a 95% bootstrap confidence interval over the 400 test locations. If the held-out median reduction is much smaller than the reported ~65% or the confidence interval includes 0, the headline gains are selection artifacts. A second check: fix the tolerances from the physiological threshold in Appendix D.7 (e.g., 0.6× the neuron-activation threshold) with no per-run tuning and compare again; this isolates the algorithm's benefit from the tuning procedure.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest quantitative claim—up to ~60% (MRI head model) and ~75% (sea-of-neurons) reduction in off-target activation—depends critically on the HingePlace tolerance hyperparameters E+_tol and E-_tol. In every simulation study, these tolerances are selected to minimize the exact metric used for evaluation: a grid search over VTh for the MRI studies (Sec. 5.1, Appendix D.3) and a 30-point random search, with case-by-case re-tuning, over Nact for the sea-of-neurons studies (Sec. 5.2, Appendix D.3). The sea-of-neurons results use the same 800 sampled neuron locations both to tune the tolerances and to score the final montage, and no cross-validation, no independent tuning set, and no bootstrap confidence intervals are reported. Because the tolerance search is performed on the test label, and because the allowed slack (tolerances up to 0.7×Edes) sits just below the activation thresholds used by the metrics, any sampling noise can be amplified into large apparent gains. This is load-bearing even if the underlying thresholding assumption is biologically correct: the paper's central claim is that HingePlace montages produce more focal neural responses than LCMV-E, and if the margin is an artifact of per-run hyperparameter selection, that claim is not yet established. The reader's concern about sub-threshold field relevance is real, but the sea-of-neurons model directly tests spiking; the evaluation-tuning flaw is the weak link in the empirical evidence for the headline numbers.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HingePlace, a convex optimization algorithm for multi-electrode transcranial electrical stimulation montages that penalizes off-target electric field components only above user-specified tolerance levels, thereby exploiting the thresholding behavior of neural response. It extends the unification result of Fernández-Corazza et al. by proving (Theorem 1, Appendix A) that the CDM and LCMV-E formulations are essentially equivalent, and shows (Appendix C) that the L1L1-norm optimization of Prieto et al. is a special case of HingePlace with p=1 and direction-independent tolerances. The algorithm is evaluated in two simulation platforms: an MRI head model (ROAST) using the VTh metric, and a sea-of-neurons model combining spherical head models with biophysically realistic neurons using the Nact metric. The authors report consistent focality improvements over LCMV-E in both platforms, with reductions in off-target activation as high as 60-75% in some conditions.","tokens_in":48618,"tokens_out":8438,"duration_ms":74874,"significance":"If the empirical claims hold, the paper makes two contributions: a clean theoretical unification of major electrode placement algorithms, and a novel, convex design objective that permits sub-threshold off-target fields to improve focality. The formal equivalence proof in Appendix A is detailed and accompanied by numerical illustrations, and the sea-of-neurons evaluation is a strong independent check that goes beyond common field-based surrogates. The code is publicly available, which supports reproducibility. However, the headline performance numbers currently rest on simulation studies whose tolerance hyperparameters are selected to minimize the very metric used for evaluation, and the sub-threshold irrelevance assumption is embedded in both the loss and the MRI-head-model metric. The significance is therefore conditional on a more rigorous validation protocol.","major_comments":[{"comment":"The tolerance hyperparameters are selected on the evaluation metric. For the MRI studies, Etol is chosen by grid search and the paper states 'We always present the result with the least VTh'; for the sea-of-neurons studies, [Extol,Eytol,Eztol] is chosen by a 30-point random search minimizing Nact, and if HingePlace underperforms LCMV-E at some (Isafe, Imultot) the tolerances are re-tuned at that operating point. Because the same 800 sampled neuron locations are used both for tuning and scoring, and because no cross-validation, independent tuning set, or bootstrap confidence intervals are reported, the reported 60-75% reductions in off-target activation are not protected against selection bias and are likely optimistic. Please add an independent tuning/validation split (or fix tolerances a priori from biophysical thresholds), and report means and confidence intervals over repeated neuron samples.","section":"Appendix D.3 / Sec. 5"},{"comment":"The comparison against 'L1L1-norm optimization' may not implement Prieto et al.'s method correctly. Appendix C shows that the l1 regularizer in (C.1) is equivalent to reducing the total-current constraint to Itot=∥I∗∥1/2 on the HingePlace side, and that the target-fit term becomes a hard constraint. The text in Sec. 5.2.1 says only that L1L1 was analyzed 'by implementing the HingePlace algorithm with p=1 and the same tolerance across all directions,' with a footnote replacing the one-sided l∞ constraint. If the reduced Itot and/or the l1 regularizer are not included, the baseline is not the [28] algorithm, and the conclusion that direction-specific tolerances are essential may rest on an unfair comparison. Please state the exact optimization solved and verify the equivalence numerically for the reported operating points.","section":"Sec. 5.2.1 / Appendix C"},{"comment":"The hinge-loss objective and the MRI-head-model metric VTh both presuppose that sub-threshold off-target fields are harmless: the loss does so through E+tol and E−tol, and VTh through its 80%-of-Edes activation threshold. Consequently, the MRI results cannot independently validate the thresholding hypothesis, and the sea-of-neurons results test only direct spiking, not possible subthreshold modulation of plasticity or network activity (a limitation acknowledged in Sec. 6 but not examined). Since the paper's central claim is that HingePlace produces more focal neural responses in real tissue, please either provide a sensitivity analysis over tolerance values (for example, showing the focality gain as a function of the assumed threshold) or explicitly bound the conditions under which subthreshold fields can be ignored.","section":"Sec. 1, Eq. (9), Appendix D.1"}],"minor_comments":[{"comment":"The maximum reported focality improvement is stated inconsistently: the abstract says 60%, Sec. 1 says '~30% (MRI) and ~60% (sea of neurons)', Sec. 5.2.1 reports up to 75%, and Sec. 5.2.3 reports up to 65%. Please harmonize the headline numbers with the exact values in the results.","section":"Abstract / Sec. 1 / Sec. 5.2"},{"comment":"The sentence 'TP refers to target point in l and k' is unclear because k and l are subplot labels; please specify which panels are meant and what TP denotes.","section":"Fig. 7 caption"},{"comment":"The text says 'We run the same study as Sec. 5.1.4' but should refer to a preceding study (e.g., Sec. 5.1.3 or Sec. 5.1.1); please correct the cross-reference.","section":"Sec. 5.1.4"},{"comment":"The grid for p=2,3 is listed as {0.01Vm−1, 0.65Vm−1, 0.55Vm−1, 0.35Vm−1}; please explain the ordering and the inclusion of 0.01, and state whether the same grid was used for both p=2 and p=3.","section":"Appendix D.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript makes code available and the Appendix A proof is careful; the sea-of-neurons evaluation is a genuine strength. The main risk is the validation protocol: the tolerance hyperparameters are tuned on the test metric, which can inflate the headline gains. If the authors can provide a proper hold-out analysis or justify tolerances from independent biophysical data, the paper would be suitable for publication. The relationship to the authors' EMBC 2021 paper is transparently cited, and the patent disclosure is explicit; the present work's added value is the theoretical unification and the biophysical validation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the unification result is real: Theorem 1, showing CDM and LCMV-E are essentially equivalent, is proved carefully in Appendix A with a numerical check, and combined with Fernandez-Corazza et al. it does unify the four major electrode-placement families. That alone is a useful contribution. Second, the hinge-loss idea is genuinely new: replacing the quadratic off-target penalty with a symmetrized hinge that ignores fields below direction-dependent tolerances is not a routine tweak, and the paper correctly identifies that all traditional algorithms over-penalize sub-threshold fields. The sea-of-neurons evaluation with biophysically realistic neurons is a strong independent check, and the code is public.\n\nThe soft spots are real, and they are the ones the stress-test flags. The headline 60-75% reductions are simulation-only, and the tolerance hyperparameters are selected to minimize exactly the VTh or Nact metric used for comparison. The MRI studies use a grid search over Etol with the least VTh chosen; the sea-of-neurons studies use a 30-point random search over [Ex_tol, Ey_tol, Ez_tol] selecting the least Nact, with case-by-case re-tuning when HingePlace underperforms. Worse, the same 800 sampled neuron locations are used both to tune tolerances and to score the final montage. No cross-validation, no held-out tuning set, no error bars. That means the margin could be partly an artifact of per-run cherry-picking. The reader's concern about sub-threshold fields being behaviorally irrelevant is biologically fair, but the sea-of-neurons model directly tests spiking, so the tuning-on-the-test-label issue is the weaker link in the empirical evidence. The abstract's \"as much as 60%\" also understates the text's 75% sea-of-neurons claim, which does not build confidence.\n\nThe theory side holds up. The thresholding insight is sound, the convexity argument is clear, and the L1L1 equivalence in Appendix C is careful. The MRI metric VTh is a simplified radial-in threshold surrogate, but the authors admit that and use the biophysical model as a check. The citation pattern looks fair, including the authors' own earlier work and the independent Prieto et al. result.\n\nWho is this for? Anyone working on tES montage optimization or, more broadly, on replacing field-focality objectives with response-focality objectives. A serious referee should engage with it, mainly to demand proper held-out evaluation and error bars before the quantitative claims are taken at face value. I would accept it for peer review, but with the expectation that the empirical section needs substantial revision.","headline":"Solid unification theorem and a genuinely new loss idea, but the headline gains are inflated by tuning tolerances on the evaluation metric; still worth refereeing.","tokens_in":49115,"tokens_out":1167,"would_cite":true,"duration_ms":13658,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Replacing zero-forcing penalties with a symmetrized hinge loss that ignores sub-threshold electric fields yields more focal neural responses in transcranial stimulation simulations, with off-target activation reduced by up to about 60…","keywords":["transcranial electrical stimulation","electrode placement algorithms","hinge loss","neural thresholding","convex optimization","focality","LCMV-E","sea of neurons model"],"falsifier":"In an animal or ex-vivo preparation, compare HingePlace and LCMV-E montages delivering matched target activation while recording off-target spiking or calcium activity across the stimulated region: if LCMV-E produces equal or fewer off-target responses, the thresholding-based focality gains claimed here do not transfer to real tissue.","tokens_in":48071,"feed_emoji":"🧠","tokens_out":10516,"duration_ms":84790,"temperature":0.7,"pith_summary":"All major electrode-placement algorithms for transcranial electrical stimulation share a common restriction, the paper argues: they drive the off-target electric field toward zero, even though neural response is thresholded and non-zero fields below the threshold produce little or no direct stimulation. Extending a known unification result, the paper proves that the four main algorithm families—weighted least squares, reciprocity-based, constrained directional maximization, and LCMV-E—are essentially equivalent, all minimizing off-target field magnitude. On that basis it introduces HingePlace, a convex algorithm whose symmetrized hinge loss penalizes only off-target field components above user-set tolerances. Across an MRI head model and a biophysically realistic 'sea of neurons' model, HingePlace montages consistently reduced off-target activation relative to LCMV-E, by as much as about 60 percent and up to 75 percent in one case. If the thresholding premise holds in real tissue, the result reframes montage design from creating focal electric fields to creating focal neural responses.","feed_headline":"Hinge loss boosts brain-stimulation focality up to 75 percent","feed_subtitle":"Designing electrode montages to ignore below-threshold fields cuts off-target activation.","key_machinery":"The load-bearing object is the symmetrized hinge loss, L_hinge(I) = ‖Γ_C L_HP(I)‖_p, where L_HP(I) = max(0, T_c I − E⁺_tol) + max(0, −T_c I − E⁻_tol) for the off-target forward matrix T_c, voxel-volume weights Γ_C, and user-set upper and lower tolerance vectors. It is what lets the algorithm leave sub-threshold off-target fields unpenalized while still punishing supra-threshold leakage. The argument also rests on Theorem 1, which proves that the CDM and LCMV-E optimization problems are essentially equivalent (for α < α_max they are dual with identical unique solutions; for α ≥ α_max, LCMV-E picks the CDM solution with least off-target energy), so the four major traditional algorithm families are all special cases of HingePlace at zero tolerances.","core_discovery":"The paper's central discovery is that the traditional objective in electrode placement—minimizing the off-target electric field magnitude—is stricter than the neural response requires, because neurons respond nonlinearly and only fire or change state when the local field crosses a threshold. Theorem 1 establishes that the CDM and LCMV-E formulations are essentially equivalent (they are each other's dual for α < α_max, and for larger α, LCMV-E selects the most relevant CDM solution), thereby completing the unification of the four major algorithm families and showing that all of them force off-target fields to be close to zero. HingePlace replaces the quadratic off-target term with a symmetrized hinge loss, L_hinge(I) = ‖Γ_C L_HP(I)‖_p, whose per-voxel components max(0, T_c I − E⁺_tol) + max(0, −T_c I − E⁻_tol) are zero whenever the field stays within user-set tolerances; the problem stays convex and reduces to LCMV-E when p=2 and the tolerances are zero. In two simulation platforms, HingePlace montages generated more focal neural responses than LCMV-E montages, with off-target activation volume or spike counts reduced by up to about 60 percent (MRI head model) and up to 75 percent (sea-of-neurons model), and the paper shows that the L1L1-norm method is a special case of HingePlace with equal tolerances in all directions.","pith_inferences":["If sub-threshold off-target fields in real tissue still modulate plasticity, network synchrony, or subthreshold membrane dynamics, the simulated focality gains could shrink or vanish; a direct plasticity assay comparing HingePlace and LCMV-E montages at matched target activation would settle this.","The tolerance vectors could be made patient-specific by deriving them from individual input-output curves or neuron models, which the paper tunes only by grid or random search in simulation; this may increase the gains beyond the reported values.","Because the paper proves equivalence of the classical families, attaching the same hinge loss to reciprocity-based or weighted-least-squares formulations should yield comparable focality gains without re-deriving the algorithm.","A testable prediction of the thresholding hypothesis: the relative focality gain should be largest when the desired target field is just above the neural threshold, and should shrink when off-target fields are far supra-threshold."],"forward_implications":["All four major traditional algorithm families (WLS, reciprocity-based, CDM, LCMV-E) are subsumed by HingePlace at zero tolerances, so the measured focality gains apply to the whole classical family, not just the LCMV-E baseline used in the simulations.","Direction-specific tolerances matter: encoding the neuronal preference for certain field orientations (instead of one scalar tolerance) substantially improves focality, which means physiological orientation-sensitivity should be built into the loss.","Since the hinge-loss problem is convex, the thresholding idea can be added to existing electrode-placement pipelines with modest extra computation, the only added hyperparameters being the three tolerance vectors.","The same loss applies to other multi-electrode stimulation modalities—ECoG, deep brain stimulation, TTFields, temporal interference—wherever below-threshold off-target fields are known to be inert.","The correct design target is the focal neural response, not the focal electric field, opening the door to incorporating further neural response properties (e.g., orientation selectivity, adaptation) into montage optimization."],"supporting_citations":[{"why":"Supplies the prior unification result (WLS and reciprocity are special cases of CDM) that this paper extends, and defines the CDM formulation used in Theorem 1.","marker":"[13]"},{"why":"Provides the quasi-static linear forward model relating electrode currents to the induced field, and the LCMV-E-style optimization that serves as the classical baseline.","marker":"[8]"},{"why":"Provides the LCMV-E variant whose linear target constraint is adopted as the canonical form in the paper's LCMV-E formulation.","marker":"[26]"},{"why":"The standard MRI head-model simulation pipeline used as the first comparison platform.","marker":"[32]"},{"why":"Supplies the biophysically realistic pyramidal neuron models used in the sea-of-neurons simulation platform.","marker":"[37]"},{"why":"Supplies the sea-of-neurons modeling approach (spherical head model plus realistic neuron models) used as the second comparison platform.","marker":"[33]"},{"why":"Defines the L1L1-norm optimization that the paper proves is a special case of HingePlace with equal tolerances across directions.","marker":"[28]"},{"why":"Provides the experimental observation behind the VTh activation metric (no neural response below about 80 percent of the desired field).","marker":"[43]"},{"why":"Earlier special case of HingePlace and the origin of the VTh metric used in the MRI head model comparison.","marker":"[35]"}],"fun_headline_variants":["Hinge loss lets brain stimulators ignore weak fields","Threshold-aware electrode placement boosts focality 75%","HingePlace: exploit neuron thresholds for sharper tES","Ignoring sub-threshold fields sharpens brain stimulation","Hinge loss trims off-target activation by 75%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that sub-threshold off-target electric fields are behaviorally irrelevant—fields below a neuron's activation threshold have no meaningful off-target effect—so if they still modulate plasticity, network activity, or subthreshold membrane dynamics, the simulated focality gains could shrink or vanish in real tissue.","fun_headline_variants_meta":{"raw":{"variants":["Hinge loss lets brain stimulators ignore weak fields","Threshold-aware electrode placement boosts focality 75%","HingePlace: exploit neuron thresholds for sharper tES","Ignoring sub-threshold fields sharpens brain stimulation","Hinge loss trims off-target activation by 75%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000214,"raw_usage":{"total_tokens":1488,"prompt_tokens":1068,"completion_tokens":420,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":684,"completion_tokens_details":{"reasoning_tokens":341}},"tokens_in":684,"tokens_out":420,"duration_ms":4648,"temperature":1.0,"reasoning_tokens":341,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T15:29:18.597386+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In an animal or ex-vivo preparation, compare HingePlace and LCMV-E montages delivering matched target activation while recording off-target spiking or calcium activity across the stimulated region: if LCMV-E produces equal or fewer off-target responses, the thresholding-based focality gains claimed here do not transfer to real tissue.","supporting_citations":[{"cited_title":"Unification of optimal targeting methods in transcranial elec- trical stimulation","cited_arxiv_id":null,"evidence_quote":"Supplies the prior unification result (WLS and reciprocity are special cases of CDM) that this paper extends, and defines the CDM formulation used in Theorem 1."},{"cited_title":"Accessibility of cortical regions to fo- cal tes: Dependence on spatial position, safety, and practical constraints","cited_arxiv_id":null,"evidence_quote":"Provides the LCMV-E variant whose linear target constraint is adopted as the canonical form in the paper's LCMV-E formulation."},{"cited_title":"Re- alistic volumetric-approach to simulate transcranial electric stimula- tion—roast—a fully automated open-source pipeline","cited_arxiv_id":null,"evidence_quote":"The standard MRI head-model simulation pipeline used as the first comparison platform."},{"cited_title":"Biophysically realistic neuron models for simulation of cortical stimulation","cited_arxiv_id":null,"evidence_quote":"Supplies the biophysically realistic pyramidal neuron models used in the sea-of-neurons simulation platform."},{"cited_title":"Barth, and Pulkit Grover","cited_arxiv_id":null,"evidence_quote":"Supplies the sea-of-neurons modeling approach (spherical head model plus realistic neuron models) used as the second comparison platform."},{"cited_title":"L1-norm vs","cited_arxiv_id":null,"evidence_quote":"Defines the L1L1-norm optimization that the paper proves is a special case of HingePlace with equal tolerances across directions."},{"cited_title":"Effect of focality of transcranial 30 currents on neural responses","cited_arxiv_id":null,"evidence_quote":"Provides the experimental observation behind the VTh activation metric (no neural response below about 80 percent of the desired field)."},{"cited_title":"Hingeplace: Focused transcra- nial electrical current stimulation that allows subthreshold fields outside the stimulation target","cited_arxiv_id":null,"evidence_quote":"Earlier special case of HingePlace and the origin of the VTh metric used in the MRI head model comparison."}],"review_version":1}