Review Claim by Claim

 Quick Explanation Copied

 Long Explanation

What the paper reports

The study transforms ~5,000 T1-weighted OASIS-1-based MRI images (176×208 px, four dementia stages) into brain graphs via a Brightness Distance Matrix and thresholding, computes six distance-based topological indices (Wiener, Szeged, Graovac–Ghorbani, PI, Mostar, NGG), normalizes them with a Watts–Strogatz small-world model, and feeds them to four classifiers Graph invariants for AD staging paper.

The optimized 4-hidden-layer neural network achieved fold-wise accuracies of 88.4–90.3% (mean 89.45%, AUC 0.9766–0.9821, mean 0.9788) Fold-wise NN performance metrics. Benchmarking places this below raw-image deep-learning baselines on the same data family (ResNet-50 96.7%, GoogleNet 97.54%) Transfer learning CNN accuracies on OASIS but above traditional SIFT+SURF SVM pipelines (~86%) SIFT+SURF ML baseline on OASIS — the trade-off being interpretability and privacy compatibility rather than raw accuracy.

Critical appraisal

Strengths. The framework is transparent (six scalar indices), open-source (GitHub code released), and the WS-model normalization with a systematic ablation over rewiring probability p and degree δ is a genuine methodological contribution. Privacy claims are logically grounded since only numeric indices leave the source site.

Concerns. (1) The dataset is a single Kaggle mirror of OASIS-1; no independent external validation. (2) Class imbalance was handled with SMOTE on training data only — reasonable, but synthetic samples can inflate minority-class performance. (3) The claimed federated-learning compatibility was not experimentally demonstrated — it is forward-looking, as the authors themselves acknowledge in limitations Paper-stated limitations. (4) Table 1 differences across AD stages are extremely small (clustering coefficient C: non-demented 0.8035 vs moderate 0.7997) with no inferential statistics reported, so the biological separability of stages via these indices alone is not established. (5) Pixel-level graphs are not anatomical connectomes; the 'brain network' framing risks over-interpretation, since BDM adjacency reflects brightness similarity, not neural connectivity.

Context. Deep-learning surveys repeatedly identify dataset heterogeneity, interpretability, and cross-center generalization as the field's core challenges DL neuroimaging survey; this paper addresses interpretability and privacy explicitly, but generalization remains untested.

What would change the verdict: replication on independent multi-site cohorts; ablations showing WS normalization (vs raw indices) actually improves classification; statistical testing of stage-wise index differences; and a demonstrated federated-learning experiment.

Author Review: G. NallappabhavithranAuthor Review: R. SelvakumarDeeper Paper Review on BGPT

 BGPT Paper Review

Study Novelty

First combined systematic ablation of six distance-based topological indices across WS-model parameters applied to MRI-derived brain graphs for AD staging; individual ingredients (topological indices, WS model, graph-based AD classification) are established, but the synthesis is new.

Scientific Quality

Open code and careful five-fold CV are positives, but missing inferential statistics for Table 1, untested federated-learning claims, single Kaggle dataset, threshold-sensitivity of the BDM adjacency, and no comparison of WS-normalized vs raw indices limit quality.

Study Generality

The pipeline is modality-portable (authors propose EEG/ECG extension), but validation is confined to one OASIS-derived dataset and one disease, so demonstrated generality is narrow.

Study Usefulness

Offers an interpretable, low-dimensional, privacy-compatible alternative to black-box CNNs; useful as a biomarker scaffold, though its 89.45% accuracy trails image-level deep learning on comparable data.

Study Reproducibility

Public Kaggle/OASIS-1 data and a GitHub repository support replication, but unspecified threshold τ values, hyperparameters, and WS normalization details reduce exact reproducibility.

Explanatory Depth

Theoretical work (Theorem 1, index behavior across p and δ) is solid; mechanistic explanation of why indices differ between AD stages biologically is thin.