{"id":"d454d104-a766-4bf0-aa59-083906b7932b","arxiv_id":"2507.21853","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"In simulated four-layer head tissue, the two-layer DCS model recovers cerebral blood flow best overall, while the three-layer model uniquely extracts scalp and brain flow simultaneously.","lead":"This paper compares three analytical models for extracting cerebral blood flow from diffuse correlation spectroscopy (DCS) measurements, using computer-simulated head tissue as ground truth. It finds the two-layer model best balances accuracy and robustness, while the simpler semi-infinite model is only reliable for relative blood flow changes at large source-detector distances.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The model ranking is vulnerable to circularity: each model's 'optimal' fitting strategy is selected on the same simulated data used for the subsequent performance comparison, so the reported advantages may reflect strategy selection rather than model quality.","rationale":"The reader's stated weakest assumption was that the four-layer slab MC simulation is treated as ground truth for the human head, with fixed layer thicknesses and assumed optical/physiological parameters. That concern is real but is about external validity: the ranking may not generalize to real heads. The concern I identify is about internal validity: even within the simulation, the comparison is biased because the fitting strategies are chosen on the same data used for evaluation. This is a more load-bearing threat to the central claim because it can change the conclusions even if the simulation is perfectly realistic. The reader did mention this issue as point (2) in the rationale ('the optimal fitting strategy for each model is selected on the same data used for performance evaluation, which can bias the comparison'), but did not make it the weakest assumption. Therefore I mark agreement as 'partial.' The concrete test I propose is tractable: it requires re-running the existing pipeline with a different protocol or a split of the already-generated data, and it would directly reveal whether the ranking is stable. Given that the paper is a simulation study with acknowledged limitations, and the concern is addressable with additional analysis, the appropriate verdict remains CONDITIONAL (unchanged from the reader's verdict). I do not see a basis to reject the paper outright, but the current evidence does not fully support the strong practical recommendations without addressing this circularity.","tokens_in":25954,"tokens_out":4240,"duration_ms":50516,"concrete_test":"Re-run the full evaluation (rCBFi recovery, CBFi sensitivity, crosstalk, and timing) under two alternative protocols: (1) a train/test split, where the optimal strategy for each model is selected on a training subset (e.g., 10 of the 20 noise realizations, or a separate MC simulation) and then evaluated on the remaining held-out subset; (2) a fixed protocol, where all multi-layer models are evaluated with SD fitting and β fixed, and the semi-infinite model uses ETLR with β fixed, without per-model strategy selection. If the qualitative ranking changes under either protocol—e.g., two-layer SD no longer outperforms other strategies for rCBFi at ρ = 35 mm, or three-layer SD matches three-layer MD for simultaneous recovery—then the paper's central recommendations are artifacts of strategy selection rather than robust model differences.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Sec. 3.1 the authors compare fitting strategies and then state: 'Based on these results, we conclude that ETLR fitting with β fixed is optimal for the semi-infinite model, SD fitting with β fixed is preferable for the two-layer model, and MD fitting is best suited for the three-layer model. Therefore, we will adopt these respective fitting methods for further analyses.' All subsequent comparisons (rCBFi recovery in Figs. 5, CBFi sensitivity in Figs. 6, crosstalk in Figs. 8-9, and timing in Table 4) use these same selected strategies on the same simulated dataset, with no held-out data or cross-validation. This creates a selection bias: the strategies are chosen because they give the best CBFi estimates on the evaluation data, so the later model comparison is effectively a comparison of 'model + strategy tuned to the test set.' The central claim that the two-layer model best balances accuracy and robustness, and that the three-layer model is uniquely suited for simultaneous CBFi/SBFi recovery, could change if a different but equally defensible strategy were used. For example, two-layer MD fitting was rejected because it overestimates rCBFi (Fig. S3), yet three-layer MD fitting was adopted; had two-layer MD been selected, the two-layer model might have performed differently at large ρ. Similarly, the timing comparison is confounded because three-layer MD is intrinsically slower than SD, but this choice was motivated by accuracy on the same data. Thus the paper's practical recommendations are not a fair test of the models themselves.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper compares three analytical models for diffuse correlation spectroscopy (semi-infinite, two-layer, and three-layer) for estimating cerebral blood flow index (CBFi) and relative CBFi (rCBFi). The authors generate synthetic DCS data with a four-layer slab head model using the Monte Carlo eXtreme (MCX) toolkit, add realistic Gaussian noise validated against a milk-phantom experiment, and then evaluate multiple fitting strategies for each model: ETLR fitting with fixed or variable beta for the semi-infinite model, and single-distance (SD) versus multi-distance (MD) fitting for the two- and three-layer models. They assess CBFi sensitivity, absolute and relative CBFi recovery accuracy, crosstalk from scalp and skull blood flow variations, sensitivity to assumed parameter errors, and computational cost across source-detector separations of 20 to 35 mm. The optimal strategies are reported as ETLR with fixed beta for the semi-infinite model, SD with fixed beta for the two-layer model, and MD for the three-layer model. The central conclusions are that the two-layer model offers the best balance of accuracy and robustness for both CBFi and rCBFi, the three-layer model enables simultaneous recovery of CBFi, scalp BFi, and rCBFi, and the semi-infinite model is suitable only for rCBFi recovery at larger separations (>=30 mm).","tokens_in":26187,"tokens_out":3706,"duration_ms":44840,"significance":"The paper addresses a practical gap in DCS: there is no systematic, head-to-head comparison of the semi-infinite, two-layer, and three-layer analytical models, nor clear guidance on model selection and fitting strategies. The study is carefully designed: the Monte Carlo ground truth is generated independently of the analytical models, the noise model is validated against experimental phantom data, and the performance metrics (CBFi sensitivity, rCBFi recovery, crosstalk, computation time) are directly relevant to clinical and research users. The finding that multi-layer models nearly double CBFi sensitivity (up to ~100%, versus 36.8% for the semi-infinite model at 30 mm) is a useful quantitative benchmark, and the recommendation of a single-exponential alternative for real-time rCBFi monitoring is practical and falsifiable. However, the central model-ranking claim is weakened by the fact that each model's 'optimal' fitting strategy is selected on the same simulated dataset used for the subsequent performance comparison; this in-sample selection procedure could bias the relative ranking.","major_comments":[{"comment":"The choice of 'optimal' fitting strategy for each model is made by comparing strategies on the same simulated dataset that is subsequently used to rank the models. Sec. 3.1 concludes 'ETLR fitting with beta fixed is optimal for the semi-infinite model, SD fitting with beta fixed is preferable for the two-layer model, and MD fitting is best suited for the three-layer model,' and all later comparisons use these selected strategies. Because the strategies are chosen for best CBFi accuracy on the evaluation data, the reported model ranking is a comparison of 'model plus strategy tuned to this test set,' not a comparison of the models on an equal footing. This is load-bearing for the central claim that the two-layer model best balances accuracy and robustness. A concrete illustration is that two-layer MD fitting was rejected because it overestimates rCBFi (Fig. S3), yet three-layer MD fitting was adopted; had a different, equally plausible strategy been selected for the two-layer model, the comparison at large separations could change. I recommend validating the selected strategies on a held-out MC dataset (e.g., different noise realizations or perturbed tissue parameters) or performing a sensitivity analysis showing that the ranking is robust to this selection. Without such evidence, the relative merits of the models are not established independently of the fitting-strategy choice.","section":"Sec. 3.1 and Figs. 5-9, Table 4"},{"comment":"The manuscript claims, in the abstract and in Table 1, that 'sensitivity to assumed parameter errors' was assessed in this work, but Sec. 2.6 explicitly states that 'we provide only a brief review of this topic in the present work and do not perform a quantitative analysis.' Sec. 3.5 is likewise a literature review, not a new evaluation. Because this is one of the six performance dimensions listed as a contribution of the paper, the claimed scope exceeds what is actually delivered. This is a load-bearing inconsistency with respect to the stated aims, and it should be fixed either by adding a quantitative perturbation analysis (varying, e.g., mu_a, mu_s', or layer thicknesses for each model) or by removing the claim from the abstract, the introduction, and Table 1.","section":"Sec. 2.6, Sec. 3.5, Table 1, Abstract"},{"comment":"The entire model comparison is based on a single set of assumed tissue parameters in the four-layer slab simulation (scalp 5 mm, skull 7 mm, CSF 2 mm, and skull DB = 8e-8 mm2/s, among others). The paper itself acknowledges in the Discussion that 'the thicknesses of the extracerebral layers were kept constant' and that 'variations in layer thickness can influence brain sensitivity at different rho values.' However, the central recommendation that the two-layer model offers the best balance of accuracy and robustness across all tested separations is derived from this one parameter set. Given that real adult heads vary in scalp and skull thickness and in skull blood flow, the ranking could change under different but still physiologically plausible parameters. I recommend adding a small parameter-sweep study, at least for scalp/skull thickness and skull DB, to demonstrate that the reported ranking is stable. Without such a robustness check, the generalizability of the model-selection guidance is uncertain, although this is a limitation rather than an internal inconsistency.","section":"Sec. 2.2, Sec. 4 (Discussion, limitations)"}],"minor_comments":[{"comment":"The noise model in Eq. (5) depends on the coherence factor beta, which is fixed to 0.5 when generating the simulated g2 curves. It would be helpful to state explicitly that the same beta is used in the noise model and in the fixed-beta fits, to avoid ambiguity in the role of beta as a 'preknown' parameter (Table 3).","section":"Sec. 2.3"},{"comment":"The comparison between semi-infinite analytical fitting and single-exponential fitting is presented only for rho = 30 mm. Since the paper later recommends the semi-infinite model for rCBFi at rho >= 30 mm, including rho = 35 mm in the Bland-Altman analysis would make the recommendation more complete.","section":"Sec. 3.6 / Fig. 10"},{"comment":"The sentence defining a negative sensitivity appears twice in essentially identical form ('A negative sensitivity indicates that the recovered perturbed CBFi decreases...'). One of the two duplicate sentences should be removed.","section":"Sec. 2.5"},{"comment":"There is a typographical error in the main text: 'fminsearchbnd function was used to to minimize the penalty function' should read 'to minimize'.","section":"Sec. 2.4 / Eq. (9)"},{"comment":"For the three-layer model, the preknown parameters are listed as including L_scalp and L_skull, but the values (5 mm and 7 mm) are only given in Table 2. It would be clearer to repeat these values in Table 3 or to refer explicitly to Table 2.","section":"Table 3"},{"comment":"The CSF layer has mu_s = 0.002 mm^-1, which is extremely low; the authors cite prior work (Custo et al., Zhao and Buckley) to justify including it. It would strengthen the paper to add a brief note in Sec. 2.2 that this low-scattering layer may challenge the diffusion approximation, and to refer to the supplementary derivation where the analytical models assume diffusive layers.","section":"Sec. 2.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a biomedical-optics journal and addresses a timely, practical question. The main technical concern is the in-sample selection of fitting strategies before the model comparison; I would urge the editor to require either a held-out validation or a clear sensitivity analysis showing the ranking is robust. The mismatch between the claimed and actual treatment of parameter-error sensitivity (Sec. 2.6 vs. abstract/Table 1) should also be corrected. If the authors can address these points, the paper would be a useful reference for the DCS community."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a genuinely useful comparison for the DCS community, but the headline ranking comes with a caveat that is easy to miss. The authors select each model's 'optimal' fitting strategy on the same simulated dataset they then use to compare models. That makes the later claims about relative performance partly a comparison of the models plus their tuning, not just the models themselves.\n\nWhat is new and worth taking seriously: the paper gives a single shared protocol for the semi-infinite, two-layer, and three-layer models, with a well-described MCX four-layer slab head and a noise model validated on a milk phantom. The skull-layer BFi crosstalk quantification is new, and the SD-versus-MD comparison for the two- and three-layer models has not been done this directly. The time-to-result table is practical and honest. The noise model validation is a real strength; that kind of attention to measurement realism is often missing in simulation papers.\n\nThe soft spots are proportionate. The in-sample strategy selection is the main one. In Sec. 3.1 the authors pick the best-performing method from the same data, then use it for all subsequent analyses. That is circular in a way that a simple cross-validation or a held-out strategy-selection set would fix. It is not fatal, but it should temper the strength of the conclusions. The abstract also says the paper assesses 'sensitivity to assumed parameter errors', but Sec. 2.6 explicitly says only a brief literature review is provided. That is an overclaim, and minor but real. Finally, the whole comparison is on a simulated slab head; the authors acknowledge this, but readers should not mistake the guidance for validated in vivo performance.\n\nWho this is for: anyone using DCS to estimate CBFi or rCBFi and wondering which analytical model to adopt. The paper is a solid reference for model selection debates. It deserves a serious referee: the methods are clear, the data are reproducible in principle, and the community will want this comparison even if the recommendations need revision. I would accept it for review, with a request that the authors either add cross-validation or soften the conclusions, and fix the abstract's parameter-error claim.","headline":"Useful systematic DCS model comparison, but the ranking is biased by in-sample fitting-strategy selection and the abstract overclaims a parameter-error analysis.","tokens_in":26786,"tokens_out":2318,"would_cite":true,"duration_ms":27821,"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":"The two-layer DCS model best balances accuracy and robustness for CBFi and rCBFi recovery at 20–35 mm, while the three-layer model alone recovers CBFi, SBFi, and rCBFi together.","keywords":["diffuse correlation spectroscopy","cerebral blood flow","multi-layer tissue model","Monte Carlo simulation","blood flow index","source-detector separation","autocorrelation function","model selection"],"falsifier":"Re-run the comparison with subject-specific head geometries obtained from magnetic resonance imaging, or with a two-layer liquid phantom whose top- and bottom-layer flows are controlled independently, across source-detector separations of 20, 25, 30, and 35 mm; if the two-layer single-distance fit no longer recovers CBFi and rCBFi more accurately than the three-layer multi-distance fit, or if the semi-infinite model matches the two-layer model for rCBFi at 30 mm and beyond, the central ranking would be contradicted.","tokens_in":2133,"feed_emoji":"🧠","tokens_out":2062,"duration_ms":103767,"temperature":0.7,"pith_summary":"This paper tries to establish which analytical model for diffuse correlation spectroscopy (DCS) should be used to measure cerebral blood flow index (CBFi) from the head's surface. It compares the standard semi-infinite model with two-layer and three-layer models that account for contamination from the scalp and skull, using Monte Carlo simulations of a four-layer head phantom as ground truth. The paper claims that the two-layer model offers the best balance of accuracy and robustness for both absolute CBFi and relative CBFi (rCBFi) across all tested source-detector separations (20–35 mm), while the three-layer model is the only one that simultaneously recovers CBFi, scalp blood flow index (SBFi), and rCBFi. It also finds that the semi-infinite model is only reliable for rCBFi at separations of 30 mm or more. A reader should care because choosing the wrong model or fitting strategy can distort cerebral blood flow measurements, and this work gives concrete selection guidance.","feed_headline":"Two-layer model wins for optical brain-flow readings","feed_subtitle":"Simulation ranking shows when to use each DCS analysis model, and which one runs in real time.","key_machinery":"The paper's engine is the correlation diffusion equation for the electric-field autocorrelation function $G_1(\\mathbf{r},\\tau)$, whose normalized form $g_1$ is linked to the measured intensity autocorrelation function through the Siegert relation $g_2(\\tau)=1+\\beta|g_1(\\tau)|^2$. Three analytical Green's-function solutions are used as forward models: a semi-infinite homogeneous medium, a two-layer slab (scalp and skull over the brain), and a three-layer slab (scalp, skull, and a semi-infinite brain). The fitted parameters are blood flow indices of the form BFi $=\\alpha D_B$, and the paper selects an optimal fitting strategy for each model: early-time-lag fitting with fixed $\\beta$ for the semi-infinite model, single-distance fitting with fixed $\\beta$ for the two-layer model, and multi-distance fitting for the three-layer model. The ground-truth data come from a Monte Carlo simulation of a four-layer slab head, with realistic Gaussian noise added through a published DCS noise model; this ground truth allows the paper to quantify sensitivity, contamination, and accuracy across source-detector separations of 20 to 35 mm.","core_discovery":"The central claim is that multi-layer DCS analytical models nearly double CBFi sensitivity compared with the semi-infinite model, reaching approximately 100% versus 36.8% at a 30 mm source-detector separation, but that the best model depends on what quantity is being measured. The semi-infinite model, when processed with early-time-lag fitting and a fixed coherence factor, underestimates absolute CBFi and should be used only for relative CBFi recovery at larger separations (≥30 mm). The two-layer model, using single-distance fitting with a fixed coherence factor, recovers both CBFi and rCBFi accurately across all tested separations of 20–35 mm and is the most resistant to blood flow changes in the scalp and skull. The three-layer model, using multi-distance fitting, is the only tested model that simultaneously recovers CBFi, SBFi, and rCBFi, but it is far more computationally expensive (about 35,000 seconds per 500 samples) and remains sensitive to extracerebral blood flow changes. The semi-infinite model and a simplified single-exponential fit are fast enough for real-time monitoring, processing 500 samples in under half a second.","pith_inferences":["Because the ranking rests on one simulated head geometry (5 mm scalp, 7 mm skull, 2 mm CSF) and a small assumed skull blood flow, the two-layer advantage may not carry over to infants, skull defects, or conditions that alter scalp and skull thickness; repeating the comparison with varied thicknesses is a natural next test.","The result that a simpler single-distance, fixed-beta fit outperforms multi-distance fitting for the two-layer model suggests that adding fitting dimensions can trade accuracy for bias, a trade-off that may generalize to other layered optical measurements.","The near-equivalence of single-exponential and semi-infinite early-time-lag fitting at 30 mm implies that rCBFi monitors could drop model parameters entirely and rely on decorrelation speed, a simplification that could be tested directly on existing clinical hypercapnia datasets.","The fixed noise setting (10 kcps at 30 mm, 60 s integration) was chosen as one realistic point; lower photon counts or shorter integration times would likely worsen the large-separation instability seen in two-layer and three-layer fits, potentially shrinking the separation range over which the paper's ranking holds."],"forward_implications":["For absolute CBFi measurements at source-detector separations of 20–35 mm, the two-layer model with single-distance fitting and a fixed coherence factor is the recommended choice; the semi-infinite model underestimates CBFi and should be avoided.","For relative CBFi only, the semi-infinite model with early-time-lag fitting, or even a single-exponential fit, is adequate at separations of 30 mm or more and is the only tested option fast enough for real-time monitoring.","The three-layer model with multi-distance fitting is the only one that simultaneously recovers CBFi, SBFi, and rCBFi, but it costs about 35,000 seconds per 500 samples and remains sensitive to scalp and skull blood flow changes.","Errors in assumed optical properties and layer thicknesses barely affect rCBFi in any of the models, so relative measurements need less prior anatomical information than absolute ones.","Multi-layer models increase CBFi sensitivity from about 37% to near 100% at 30 mm separation, implying that the standard semi-infinite analysis can miss a large portion of cerebral blood flow changes."],"supporting_citations":[{"why":"Supplies the two-layer analytical model and the two-layer slab Monte Carlo head simulation approach that this work extends.","marker":"[4]"},{"why":"Supplies the three-layer analytical model and multi-distance fitting strategy, plus the parameter-error findings for the three-layer model.","marker":"[42]"},{"why":"Provides the multilayer Monte Carlo modeling framework and the early-time-lag fitting constraint used for the semi-infinite model.","marker":"[22]"},{"why":"Supports the use of early-time-lag fitting for the semi-infinite model and multi-distance fitting for the three-layer model, with validation on phantoms and animal experiments.","marker":"[25]"},{"why":"Justifies multi-distance fitting for multilayer models and shows how source-detector separation affects model accuracy.","marker":"[55]"},{"why":"Supplies the Gaussian noise model used to synthesize realistic, noisy DCS data.","marker":"[69]"},{"why":"Provides the GPU-accelerated Monte Carlo simulation toolkit used to generate the ground-truth four-layer head data.","marker":"[54]"},{"why":"Defines the CBFi sensitivity metric and supports the claim that early-time-lag fitting improves brain sensitivity.","marker":"[24]"}],"fun_headline_variants":["Two-layer DCS model best for brain blood flow","Multi-layer DCS nearly triples brain flow sensitivity","Real-time brain flow with semi-infinite DCS model","Three-layer DCS model yields brain and scalp flow","Simulation shows best DCS model depends on goal"],"cache_read_input_tokens":28800,"weakest_assumption_plain":"The ranking depends on treating the four-layer slab Monte Carlo simulation, with fixed layer thicknesses (scalp 5 mm, skull 7 mm, CSF 2 mm) and a low assumed skull blood flow (8e-8 mm2/s), as a faithful stand-in for the real adult head; if those tissue values are wrong, the model ranking could change.","fun_headline_variants_meta":{"raw":{"variants":["Two-layer DCS model best for brain blood flow","Multi-layer DCS nearly triples brain flow sensitivity","Real-time brain flow with semi-infinite DCS model","Three-layer DCS model yields brain and scalp flow","Simulation shows best DCS model depends on goal"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001489,"raw_usage":{"total_tokens":6077,"prompt_tokens":1143,"completion_tokens":4934,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":759,"completion_tokens_details":{"reasoning_tokens":4867}},"tokens_in":759,"tokens_out":4934,"duration_ms":43683,"temperature":1.0,"reasoning_tokens":4867,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:16:54.180115+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the comparison with subject-specific head geometries obtained from magnetic resonance imaging, or with a two-layer liquid phantom whose top- and bottom-layer flows are controlled independently, across source-detector separations of 20, 25, 30, and 35 mm; if the two-layer single-distance fit no longer recovers CBFi and rCBFi more accurately than the three-layer multi-distance fit, or if the semi-infinite model matches the two-layer model for rCBFi at 30 mm and beyond, the central ranking would be contradicted.","supporting_citations":[{"cited_title":"Inves�ga�on of diffuse correla�on spectroscopy in mul� -layered media including the human head,","cited_arxiv_id":null,"evidence_quote":"Supplies the two-layer analytical model and the two-layer slab Monte Carlo head simulation approach that this work extends."},{"cited_title":"Accuracy of diffuse correla�on spectroscopy measurements of cerebral blood flow when using a three -layer analy�cal model,","cited_arxiv_id":null,"evidence_quote":"Supplies the three-layer analytical model and multi-distance fitting strategy, plus the parameter-error findings for the three-layer model."},{"cited_title":"Improved accuracy of cerebral blood flow quan�fica�on in the presence of systemic physiology cross -talk using mul� -layer Monte Carlo modeling,","cited_arxiv_id":null,"evidence_quote":"Provides the multilayer Monte Carlo modeling framework and the early-time-lag fitting constraint used for the semi-infinite model."},{"cited_title":"A s s e s s m e n t o f a m u l �-layered diffuse correla�on spectroscopy method for monitoring cerebral blood flow in adults,","cited_arxiv_id":null,"evidence_quote":"Supports the use of early-time-lag fitting for the semi-infinite model and multi-distance fitting for the three-layer model, with validation on phantoms and animal experiments."},{"cited_title":"Influence of source –detector separa�on on diffuse correla�on spectroscopy measurements of cerebral blood flow with a mul�layered analy�cal model,","cited_arxiv_id":null,"evidence_quote":"Justifies multi-distance fitting for multilayer models and shows how source-detector separation affects model accuracy."},{"cited_title":"Diffuse op�cal correla�on tomography of cerebral blood flow during cor�cal spreading depression in rat brain,","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian noise model used to synthesize realistic, noisy DCS data."},{"cited_title":"Sensi�vity of near -infrared spectroscopy and diffuse correla�on spectroscopy to brain hemodynamics: simula�ons and experimental findings during hypercapnia,","cited_arxiv_id":null,"evidence_quote":"Defines the CBFi sensitivity metric and supports the claim that early-time-lag fitting improves brain sensitivity."}],"review_version":1}