- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
GeoNIRS studies pretraining across fNIRS datasets with different probe layouts and channel counts.
For each channel, the model receives the measured signal together with source-detector coordinates. These channel-wise representations are compressed into a fixed latent space. HbO and HbR keep separate signal streams. During pretraining, masked portions of one chromophore are reconstructed using the other stream and the visible context.
GeoNIRS-base uses eight pretraining datasets. The larger version uses 22. The eight downstream datasets are excluded from both pretraining sets.
On the main benchmark, full fine-tuning gives 63.54 average balanced accuracy. Linear probing gives 63.06. The strongest non-GeoNIRS entry in Table 1 is 61.73.
The paper includes a random-initialization comparison and a version trained without geometry. Reduced-label adaptation and missing-channel robustness are examined separately.
The cross-task benchmark is encouraging. The point requiring the closest scrutiny is whether coordinates collected under different probe conventions actually provide the common spatial reference assumed by GeoNIRS.
Critical strengths
The separation between pretraining and downstream datasets is a strong aspect of the evaluation. None of the eight downstream datasets is included in the pretraining corpus.
The random-initialization control is useful. Under linear probing, average balanced accuracy changes from 58.94 to 63.06 after pretraining. Under full fine-tuning, the corresponding comparison is 60.25 versus 63.54. The gain cannot be explained by architecture alone.
The frozen-backbone experiment is informative as well. GeoNIRS-LP reaches 63.06 without updating the pretrained encoder and remains above the reported non-GeoNIRS baselines on the task average.
The geometry experiments are more extensive than a single perturbation test. Retraining without geometry reduces the average result from 63.34 to 60.36 across the three adaptation settings. Corrupting geometry at inference causes a further performance drop.
The reduced-label result is useful. With 10% of the labels, GeoNIRS-FT reports 60.70 balanced accuracy, versus 55.98 from random initialization.
Missing-channel robustness is relevant for fNIRS. At 50% channel loss, GeoNIRS degrades less than several of the comparison methods.
Code and pretrained checkpoints are available for anonymous review, which is a useful contribution to reproducibility.
Critical weaknesses
The main methodological concern is the interpretation of the geometry input across datasets. GeoNIRS converts coordinates into common physical units, applies a fixed scale, and centers each probe. These steps remove differences in units and translation, but they do not align arbitrary coordinate frames. Two probes can cover comparable cortical regions while being stored under different rotations or reflections, in which case their midpoint coordinates and source-detector direction vectors differ even though their physical arrangement relative to the head is similar. This is especially important because the direction vector is explicitly part of the geometry representation. The manuscript does not describe an anatomical registration that maps all probes into a shared head coordinate system, and the current representation is not rotation invariant. The mixture of planar and three-dimensional layouts adds another layer of inconsistency because depth coordinates are structurally absent for some datasets but meaningful for others.
The ablation results establish that GeoNIRS uses geometry, but they do not yet establish what information within the geometry produces the gain. Removing geometry during pretraining lowers average performance by about three points, which is useful evidence. However, probe geometry is usually stable within a dataset, so the coordinate pattern can also act as a cue for probe configuration or acquisition source. A network could exploit this dataset-specific structure during multi-dataset training without learning anatomical correspondence in the stronger sense claimed by the paper. The zero-coordinate and shuffled-coordinate experiments do not separate these possibilities because both interventions disrupt an input the pretrained model expects. A stronger attribution test would replace physical coordinates with a learned probe-layout identifier while keeping the remaining architecture fixed. A rigid rotation of the complete probe coordinates would also reveal whether the model depends on arbitrary coordinate axes rather than invariant spatial relationships.
The statistical presentation of the main benchmark could also be stronger. Results are averaged over five subject-level folds and two random seeds, and standard deviations are shown in several tables, but the principal GeoNIRS-versus-baseline comparisons do not include paired uncertainty. The two seeds use the same fold definitions, and the cross-validation training sets overlap across folds, so the ten fold-seed outcomes should not be interpreted as ten independent replications. This matters because several headline differences are around one or two percentage points: GeoNIRS-LP reaches 63.06 on average, Jung-MAE-LoRA reaches 61.73, and GeoNIRS-FT reaches 63.54. Reporting matched per-task or per-fold differences with confidence intervals would make the strength of these improvements much easier to assess.
The task-average gains also conceal substantial heterogeneity. Semantic decoding remains near chance, with GeoNIRS-FT at 51.03 balanced accuracy. In the same task, random initialization gives 51.52, so pretraining does not improve every downstream problem. The cross-chromophore ablation shows a similar pattern. Cross-reconstruction improves the overall average, but the largest gains occur on ICH and Tinnitus, while several tasks change little. On Action Mode, self-reconstruction is higher under all three reported adaptation settings. A leave-one-task-out analysis would help determine how strongly the aggregate advantage depends on the tasks with the largest effects. The physiological interpretation should also remain narrow because HbO and HbR can share systemic or acquisition-related variation in addition to task-related hemodynamic structure.
The comparison with prior fNIRS pretraining is another limitation. Jung-MAE is reconstructed from the published description because the original implementation and pretrained checkpoint are unavailable. The comparison is still useful, but it evaluates the authors' Jung-inspired implementation rather than the original system itself. Capacity also differs substantially: GeoNIRS-base has approximately 6.5M parameters, whereas the Jung-inspired model has approximately 2.0M. The internal w/o-Geometry control helps with attribution, but an external channel-independent pretraining baseline closer to GeoNIRS in model size would provide a cleaner test of the geometry-aware contribution.
The missing-channel experiment is relevant, but it does not yet isolate geometry as the reason for robustness. Channel removal is applied only at inference, after all downstream models have been trained with complete inputs. The experiment therefore demonstrates tolerance to missing channels rather than adaptation to naturally heterogeneous recording-level layouts. The cleanest attribution test would apply the same 10%, 25%, and 50% channel-loss protocol to the w/o-Geometry checkpoint. The downstream benchmark also uses a frozen common-channel tensor within each dataset, which keeps model comparison controlled but removes part of the recording-specific channel variation that motivates the framework. The current evidence therefore supports transfer across dataset-level layouts more directly than adaptation to arbitrary per-recording channel configurations.
The GeoNIRS-large experiment is difficult to interpret as scaling evidence. GeoNIRS-base uses 6.5M parameters and eight pretraining datasets, whereas the large configuration uses 30.1M parameters and 22 datasets. Average balanced accuracy moves only from 63.54 to 63.84, while ICH decreases from 80.86 to 78.07 and CI Status also drops slightly. Because capacity and corpus composition change together, the source of the 0.30-point average increase cannot be isolated. The result is informative as a description of the larger configuration, but not yet as evidence for a reliable scaling trend. The smaller one-, four-, and five-dataset study has the same limitation because dataset composition changes with corpus size and the analysis is restricted to one downstream task.
Finally, the scope of the method is narrower than the phrase “heterogeneous fNIRS datasets” may initially suggest. The authors screened many public datasets, but only 30 provided both usable signals and enough spatial metadata to recover geometry. GeoNIRS therefore currently applies to heterogeneous fNIRS datasets with sufficiently reliable source-detector coordinates. Datasets lacking those coordinates remain outside the demonstrated setting, which limits how broadly the present framework can be used without an additional geometry-missing mechanism.
Questions for authors
Are all probe coordinates expressed in one common anatomical coordinate frame before they are supplied to GeoNIRS?
If not, how are rotations or reflections between dataset coordinate systems handled?
What happens when the coordinates of an entire probe are rigidly rotated while its channel-to-signal correspondence remains unchanged?
Can physical coordinates be replaced by a learned probe-layout identifier while keeping the rest of the architecture fixed?
Can paired uncertainty be reported for the main GeoNIRS-versus-Jung-MAE comparison?
How does the average cross-reconstruction advantage change when ICH is excluded? What is the result after excluding both ICH and Tinnitus?
Can the channel-loss experiment be repeated using the w/o-Geometry checkpoint?
Does GeoNIRS still show an advantage when downstream recordings retain their native channel availability instead of using a common frozen channel tensor?
Can a channel-independent pretraining baseline with capacity closer to GeoNIRS-base be evaluated?
Can the 30.1M-parameter architecture be pretrained using only the original eight-dataset corpus?
Can the 22-dataset corpus be used with the 6.5M-parameter model?
Would a geometry representation based on invariant distances or anatomically registered positions change the results?
3 comments
Coordinate Harmonization Across Datasets and Updated Downstream Results
We appreciate the reviewer’s concern about geometric consistency across datasets. Following the preliminary tests described in our previous response, we further investigated this important issue. After checking the coordinate sources for each dataset, we developed a procedure to harmonize their spatial representations. Specifically, we define the origin as the midpoint between the left and right preauricular landmarks and adopt RAS-like axes (+X right, +Y anterior, +Z superior), with all coordinates expressed in millimeters. For recordings providing LPA, RPA, and nasion landmarks, we construct the coordinate axes from these three points and apply rotation and translation. For template-based coordinates or datasets providing only planar layouts, we use publicly available template fiducials, supplementary three-dimensional coordinates, or explicitly identified approximate reconstructions. Dataset-specific procedures are summarized in Table 1.
Table 1. Coordinate sources and harmonization procedures.
| Stage | Dataset | Geometric information and coordinate harmonization |
|---|---|---|
| Pretraining | DS21, DS29, DS31, DS32, DS33, DS34 | Align each recording using its three-dimensional optode coordinates and LPA/RPA/Nz landmarks through rotation and translation. |
| Pretraining | DS28 | Match optode coordinates to the MNE standard_1005 template and align using the corresponding template fiducials. |
| Pretraining | DS35 | Align MNI152NLin2009bAsym template coordinates using the landmarks provided in the recording. |
| Downstream | DS7 (ICH) | Use the three-dimensional MNI optode coordinates published in the paper, provisionally aligned using ICBM152/Cutini template fiducials. |
| Downstream | DS47 (CI Status) | Align individually digitized three-dimensional optode coordinates using recording-specific LPA/RPA/NAS landmarks and convert meters to millimeters. |
| Downstream | DS36 (Tinnitus) | Align the three-dimensional optode coordinates stored in SNIRF files using their embedded LPA/RPA/Nz landmarks. |
| Downstream | DS19 (Pain Intensity) | Align the published MNI channel coordinates using Colin27 fiducials and approximately reconstruct optode endpoints from the planar probe layout and scalp projection. |
| Downstream | DS8 (Motor Recovery) | Align the three-dimensional optode coordinates stored in SNIRF files using their embedded LPA/RPA/Nz landmarks. |
| Downstream | DS42 (Action Mode) | Transform the mean three-dimensional optode coordinates provided in the supplementary material according to the inferred AtlasViewer/Colin27 frame, then align using template fiducials. |
All coordinates are expressed in millimeters with RAS-like axes (+X right, +Y anterior, +Z superior) and the origin at the midpoint between the preauricular landmarks, (LPA + RPA) / 2.
This procedure harmonizes coordinate origins, axis orientations, and units. After coordinate harmonization, we pretrained GeoNIRS again and evaluated it using the original downstream splits and evaluation protocols. Balanced accuracy, averaged equally across the six tasks, increased from 65.89%, 66.16%, and 66.36% to 66.70%, 67.70%, and 67.54% under LP, LoRA, and FT, respectively, corresponding to improvements of 0.81, 1.54, and 1.17 percentage points. LoRA improved on all six tasks. Taking the best-performing GeoNIRS adaptation setting for each task, the updated model ranked first on five tasks and second on Pain Intensity. These results support the practical value of further harmonizing coordinate representations across datasets; the complete changes are reported in Table 2. Coordinate verification and conversion for the remaining two downstream tasks are still being finalized. We therefore report six tasks in this update and compare against the original results on exactly the same task set.
Table 2. Balanced accuracy (%) before and after coordinate harmonization.
| Downstream task | LP original → updated | Δ (pp) | LoRA original → updated | Δ (pp) | FT original → updated | Δ (pp) | Updated GeoNIRS best rank |
|---|---|---|---|---|---|---|---|
| ICH | 77.82→79.02 | +1.20 | 80.10→81.26 | +1.16 | 80.86→79.26 | −1.60 | 1 |
| CI Status | 68.06→69.84 | +1.78 | 69.53→70.76 | +1.23 | 69.19→71.09 | +1.90 | 1 |
| Tinnitus | 58.70→63.41 | +4.71 | 58.72→62.84 | +4.12 | 57.70→61.41 | +3.71 | 1 |
| Pain Intensity | 56.19→52.96 | −3.23 | 54.36→55.50 | +1.14 | 54.70→55.09 | +0.39 | 2 |
| Motor Recovery | 64.27→64.56 | +0.29 | 63.92→64.71 | +0.79 | 64.75→65.85 | +1.10 | 1 |
| Action Mode | 70.29→70.41 | +0.12 | 70.34→71.16 | +0.82 | 70.99→72.51 | +1.52 | 1 |
| Six-task average | 65.89→66.70 | +0.81 | 66.16→67.70 | +1.54 | 66.36→67.54 | +1.17 | 1 |
Rigid-rotation sensitivity analysis
Rigid-rotation sensitivity analysis
We thank the reviewer for suggesting a rigid-rotation test to examine whether GeoNIRS depends on coordinate-frame orientation. We had initially deferred this evaluation because of computational constraints, but have now added it. We used the same checkpoint pretrained in the unified coordinate frame and applied a shared random 3D rotation of 120° to each channel’s complete 6D geometry representation, rotating both the channel midpoint (x, y, z) and the source–detector direction vector (Δx, Δy, Δz). The same rotation was used across all six downstream tasks, folds, seeds, and adaptation protocols. We then followed the original downstream adaptation and evaluation protocols and paired the results with those obtained using the same checkpoint and data splits without rotation. Table 1 reporRigid-rotation sensitivity analysists changes in BAcc, defined as rotated minus unrotated performance.
Table 1. Change in BAcc after rigid rotation relative to the unrotated results (percentage points).
| Downstream task | LP ΔBAcc (pp) | LoRA ΔBAcc (pp) | FT ΔBAcc (pp) |
|---|---|---|---|
| DS7 ICH | −10.17 | −9.04 | −4.76 |
| DS47 CI Status | −0.90 | +1.55 | +0.07 |
| DS36 Tinnitus | +0.91 | −0.66 | −3.19 |
| DS19 Pain | −0.23 | −2.33 | −2.22 |
| DS8 Motor Recovery | −3.02 | −3.38 | −1.56 |
| DS42 Action Mode | −2.33 | −1.38 | −1.11 |
| Six-task average | −2.62 | −2.54 | −2.13 |
Across the six tasks, equally weighted mean BAcc decreased by 2.62, 2.54, and 2.13 pp for LP, LoRA, and FT, respectively. Macro-F1 decreased by 2.88, 2.73, and 2.07 pp, and AUROC by 3.45, 2.38, and 2.28 pp, respectively. BAcc decreased on five of the six tasks under each adaptation protocol. These results indicate that downstream performance is somewhat sensitive to coordinate-frame orientation, although the effect varies across tasks and adaptation protocols. Together with the geometry-free control that preserves channel-index identity, this analysis tests whether channel identity alone accounts for the observed gains and whether the model is sensitive to coordinate orientation.
Additional analyses on coordinate frames, geometry attribution, robustness, and evaluation scope
1. Cross-dataset coordinate frames
Thank you for raising this important issue. Coordinates were already converted to mm, centered per recording, and normalized using a fixed scale. Metadata/fiducial audit indicates RAS-style axes in all eight pretraining datasets, but differing origins (relative to (LPA+RPA)/2). These steps do not anatomically register probes. We will clarify that GeoNIRS uses within-recording relative source–detector/probe-layout geometry. Future work will map probes to a common frame using shared scalp fiducials and MRI/MNI registration where available, then reevaluate pretraining.
2. Geometry attribution
We appreciate the concern that the original ablation does not isolate which geometric information accounts for the gain. We added a learned probe-layout ID control: one learnable 6D vector per layout, shared across its channels, replaces physical geometry; other settings match GeoNIRS-base. Table 1 reports eight-task mean BA (%); Δ=Learned ID−GeoNIRS-base.
GeoNIRS-base performs better under all three protocols, suggesting layout identity alone does not explain the full gain.
3. Paired uncertainty
We added matched comparisons with MOMENT and Jung-MAE. Since both seeds use the same subject-level folds, we averaged them within each fold before computing paired differences. Estimates are in our response to Reviewer 2km3, Comment 1.
4. Task heterogeneity
We added leave-one-task-out (LOTO) analysis, recomputing equal-weight means over the remaining seven tasks after each exclusion. Pretraining averages LP/FT; cross-reconstruction averages FT/LoRA/LP. Values are BA differences (pp).
Both gains remain positive after any single-task exclusion, though contributions vary; excluding ICH reduces the cross-reconstruction gain. We will distinguish task averages from task-level outcomes and avoid implying universal improvements. Cross-chromophore reconstruction tests predictive structure between HbO and HbR, which may also reflect shared systemic or acquisition-related variation.
5. Jung-inspired MAE
We appreciate the opportunity to clarify. As original code and checkpoint were unavailable, we evaluated a paper-guided Jung-inspired implementation. The initial model matched GeoNIRS in encoder/decoder depth and width, but not parameter count. We retrained a channel-independent, capacity-matched version (~6.5M parameters). Results are in our response to Reviewer 2km3, Comment 2. GeoNIRS-base retains higher average performance, so parameter count alone does not explain the gap.
6. Missing-channel robustness
We agree the original experiment did not isolate geometry as the source of robustness. It tests inference-time channel-loss tolerance after downstream training with complete channels. We applied the same 10%, 25%, and 50% loss protocol to the w/o-Geometry checkpoint; GeoNIRS showed no clear robustness advantage over this control. We will remove the geometry-based explanation and retain only that GeoNIRS performs favorably relative to non-GeoNIRS baselines under inference-time channel loss. This does not establish adaptation to naturally varying recording-level layouts.
7. Scaling analysis
We agree Base–Large changes both model capacity and corpus composition; its +0.30-point mean gain cannot be attributed to either alone. We will report it in the Supplementary Material as the larger configuration's result. Separately, we fixed GeoNIRS-base (6.5M parameters) and varied the corpus from 5 to 8 datasets. The four configurations represent ~12%, 28%, 30%, and 100% of pretraining hours; FT mean BA across eight tasks was 61.06%, 61.51%, 61.63%, and 63.54%. We will present this as corpus-size sensitivity: performance rises across tested nested configurations, without attributing the trend to dataset count alone.
8. Scope and geometry metadata
We appreciate the point about scope. GeoNIRS requires spatial metadata sufficient to recover channel geometry. Of ~100 public datasets screened, 30 met our criteria. In our check of 10 public fNIRS dataset papers in Scientific Data (2024–2026), 9 provided explicit spatial information; we therefore do not consider this requirement to severely restrict practical use. Missing coordinates also limit localization and cross-study comparison. fNIRS best practices recommend reporting source/detector positions relative to head landmarks; NIRS-BIDS supports measured or standard-space coordinates, helping data reuse and cross-dataset modeling.