PaperlayerView exact revision on arXiv

Source-linked AI summary

FedNI: Federated Graph Learning with Network Inpainting for Population-Based Disease Prediction

Liang Peng, Nan Wang, Nicha Dvornek, Xiaofeng Zhu, Xiaoxiao Li

arXiv:2112.10166v3cs.LGcs.SI

TL;DR

Medical institutions face privacy barriers to collaboration and incomplete local graphs that limit federated GCN disease prediction. FedNI uses federated network inpainting followed by GCN training on fused local graphs. It reports better performance than local and baseline federated methods across two neuroimaging datasets.

  • Problem

    Institutions face privacy and ownership barriers to sharing patient data, while their local graphs lack cross-institution neighbors needed for complete GCN inputs.

    Source: §I · p. 1, B. Network inpainting · p. 4

  • Method

    FedNI federatively trains an SN-GAN-based missing-node generator to complete local graphs, then federatively trains a GCN classifier on the fused graphs.

    Source: B. Network inpainting · p. 4, 2) Local network inpainting · p. 5

  • Results

    FedNI outperforms comparison methods across accuracy, AUC, precision, recall, and F1-score on both datasets, with significant improvements reported over comparison methods.

    Source: 2) Results and analysis · p. 8

  • Takeaways & Limitations

    The experiments support FedNI as a high-performing approach to federated population-based disease prediction on ABIDE and ADNI.

    Source: §I · p. 2, §V · p. 11

    The authors identify data heterogeneity as an issue for future work to address.

    Source: §V · p. 11

Abstract

from arXiv · show

Graph Convolutional Neural Networks (GCNs) are widely used for graph analysis. Specifically, in medical applications, GCNs can be used for disease prediction on a population graph, where graph nodes represent individuals and edges represent individual similarities. However, GCNs rely on a vast amount of data, which is challenging to collect for a single medical institution. In addition, a critical challenge that most medical institutions continue to face is addressing disease prediction in isolation with incomplete data information. To address these issues, Federated Learning (FL) allows isolated local institutions to collaboratively train a global model without data sharing. In this work, we propose a framework, FedNI, to leverage network inpainting and inter-institutional data via FL. Specifically, we first federatively train missing node and edge predictor using a graph generative adversarial network (GAN) to complete the missing information of local networks. Then we train a global GCN node classifier across institutions using a federated graph learning platform. The novel design enables us to build more accurate machine learning models by leveraging federated learning and also graph learning approaches. We demonstrate that our federated model outperforms local and baseline FL methods with significant margins on two public neuroimaging datasets.

I. INTRODUCTION

Population-based disease prediction can use GCNs to combine individual imaging features with relationships among subjects, but institutions face privacy constraints and local graphs lack cross-institution neighbors. FedNI combines federated learning with network inpainting to address these gaps.

I. INTRODUCTION

  • Population graphs represent individuals as nodes and their similarities as edges, allowing GCNs to use population interactions when predicting disease labels.The introduction notes that CNNs and RNNs process imaging information independently and may fail to capture population-level interactions.Source: §I · p. 1
  • Federated learning lets institutions collaboratively use private data without sharing it, addressing privacy and ownership barriers to centralized patient-data collaboration.The passages cite GDPR and HIPAA as examples of relevant constraints.Source: §I · p. 1
  • Cross-silo GCNs face missing edges between local graphs, which can corrupt ego-network information and undermine GCN effectiveness.Connecting local graphs would require sharing node information and may violate privacy regulations.Source: §I · pp. 1–2
  • FedNI first trains SN-GAN generators to predict missing nodes and edges, then uses fused local graphs to federatively train a GCN classifier.The classifier weights are iteratively shared with a trusted administration center and averaged.Source: §I · p. 2
  • FedNI reports 66.7% accuracy on ABIDE and 75.8% on ADNI, outperforming alternative federated-learning methods.The authors describe this as a contribution and report experiments on both datasets.Source: §I · p. 2

A. GCNs for disease prediction

The paper motivates population graphs for disease prediction, identifies limitations in graph construction and federated learning, and introduces a two-phase approach that constructs multimodal graphs and trains across institutions.

A. GCNs for disease prediction

  • Single-modality GCN graphs may miss complex subject relationships, while incomplete or low-quality graphs can impair classification.The paper proposes combining imaging and non-imaging information to represent the population graph.Source: A. GCNs for disease prediction · p. 2

B. Federated learning for disease prediction

  • Existing federated graph learning does not fully address small local disease graphs, auxiliary phenotype information, or healthcare privacy requirements, which FedNI targets.The authors identify Zhang et al. [19] as the closest related work and distinguish FedNI from it on these points.Source: B. Federated learning for disease… · p. 2

III. THE PROPOSED METHOD

  • FedNI first inpaints missing nodes and augments edges, then federatively trains a GCN classifier on the resulting fused local graphs.The framework's two modules are trained in a federated-learning fashion.Source: §III · p. 3, Figure 1 · p. 3

A. Population graph formulation

  • The population graph combines image-feature similarities with phenotype similarities, sparsifies the fused edges to each node's top k, and adds self-loops.The edge fusion is a Hadamard product of the similarity and phenotypic graph matrices.Source: A. Population graph formulation · pp. 3–4

B. Network inpainting

FedNI addresses incomplete local graphs by generating missing nodes and edges for federated GCN learning. Its node-masking procedure uses BFS tree leaves to create generator training examples while preserving graph structure for feature propagation.

B. Network inpainting

  • Local graphs omit cross-institution neighbors, leaving ego-networks incomplete and potentially degrading GCN performance; FedNI generates missing nodes and edges to augment message passing.The generated nodes and edges are intended to complete local ego-networks.Source: B. Network inpainting · p. 4

1) Node masking for missing node generation:

  • BFS-based masking removes leaf nodes to form incomplete-graph and hidden-node training pairs, avoiding the isolated or disconnected components random removal may create.The hidden nodes usually comprise 10%–15% of total nodes in the reported experiments.Source: 1) Node masking for missing node… · p. 4

2) Local network inpainting:

Local network inpainting masks nodes and predicts their counts, features, and phenotypes from incomplete graphs. FedNI combines reconstruction and adversarial training, then uses predicted phenotypes to construct edges for the fused graph.

2) Local network inpainting:

  • BFS masking removes nodes while preserving connected structure, creating the incomplete local graphs used to predict missing neighbors.BFS-based masking is identified as a way to avoid isolated components in the corrupted graph.Source: 2) Local network inpainting · pp. 4–5
  • A GCN encoder embeds each incomplete graph, and regression models use those embeddings to predict missing-neighbor counts and features.The count predictor maps node embeddings to an integer; the feature decoder reconstructs masked-neighbor features.Source: 2) Local network inpainting · pp. 4–5
  • SN-GAN training supplements feature reconstruction loss to make generated missing-node features better match the hidden-neighbor data distribution.The generator is an encoder–decoder, with the GCN encoder processing incomplete graph information and an MLP decoder reconstructing masked-neighbor features.Source: 2) Local network inpainting · p. 5
  • Predicted phenotypes, together with node features, are used to connect missing nodes to known nodes and construct edges.Class-labeled phenotypes use cross-entropy loss, while age prediction can use regression losses.Source: 2) Local network inpainting · p. 5
  • FedNI predicts phenotypes as well as node features, then uses phenotype information to rebuild edges and augment message passing.The paper distinguishes this from Zhang et al. [19], whose method uses random masking and auxiliary information from other local graphs.Source: 2) Local network inpainting · p. 5

C. Federated learning

FedNI uses two federated phases to address incomplete local graphs and the limited data available to individual institutions. The phases train missing-node predictors and then a global node classifier.

C. Federated learning

  • FedNI uses federated network inpainting to fill incomplete local-graph features, followed by federated GCN training for a global node classifier.The paper notes that GANs can perform poorly on small data and GCN node classification also requires many samples.Source: C. Federated learning · p. 5

1) Federated network inpainting:

FedNI federatively trains a missing-node generator while retaining local discriminators, then uses the generator to complete each institution’s graph. The resulting fused graphs support federated GCN classifier training.

1) Federated network inpainting:

  • The federated inpainting phase averages local generator parameters into a global generator while keeping each discriminator local.The authors report that discriminator averaging can hurt performance and that local discriminators can better fit graph heterogeneity.Source: 1) Federated network inpainting · pp. 5–6
  • At each institution, the trained generator predicts missing-neighbor counts, features, and phenotypes, and the predictions are used to build a fused local graph.The graph is completed by inference on each institution’s original local graph, followed by edge construction.Source: 1) Federated network inpainting · p. 6

2) Federated GCN node classification:

  • Federated GCN training alternates local classifier training on fused graphs with server-side averaging and broadcasting of classifier weights.The method applies FedAvg across the local GCN classifiers.Source: 2) Federated GCN node classification · p. 6

3) Privacy preservation techniques:

FedNI adds Gaussian noise to model weights shared with the administration center as a differential-privacy measure. Its algorithm organizes training into federated inpainting, graph completion, and federated GCN classification.

3) Privacy preservation techniques:

  • FedNI adds Gaussian noise with mean 0 and standard deviation 0.01 to shared local generator and GCN-classifier weights to satisfy differential privacy.The paper motivates this measure by noting that shared model parameters can reveal input data.Source: 3) Privacy preservation techniques · p. 6
  • The algorithm first trains missing-node generators under FL, completes local graphs, and then trains GCN classifiers under FL.Its listed procedures separate server-side aggregation from local network-inpainting and classifier operations.Source: 3) Privacy preservation techniques · p. 6

IV. EXPERIMENTS

The experiments evaluate FedNI on the ABIDE and ADNI neuroimaging datasets. They describe the datasets’ subjects and feature representations.

1) Datasets:

  • ABIDE includes 1,029 subjects: 485 with ASD and 544 healthy controls.The data come from ABIDE-I and ABIDE-II.Source: 1) Datasets · p. 6
  • ABIDE subjects are represented by 7,503-dimensional feature vectors derived from correlations among 122 brain regions.Each subject’s functional-connectivity network uses Pearson correlations between ROI time series.Source: 1) Datasets · p. 7
  • ADNI includes 911 subjects: 375 with Alzheimer’s disease and 536 with mild cognitive impairment, represented by 345-dimensional feature vectors.The features are anatomical statistics extracted from structural MRI.Source: 1) Datasets · p. 7

2) Model training Setting-up:

The experiments compare centralized, local, and federated methods under controlled data splits and repeated cross-validation. They evaluate classification with five metrics and report mean performance across runs.

2) Model training Setting-up:

  • For each dataset, subjects were equally and randomly assigned to five institutions; experiments used an 80% labeled rate and five-fold cross-validation.The authors repeated each method five times with random seeds and report average performance and standard deviation.Source: 2) Model training Setting-up · p. 7, 3) Performance evaluation · p. 7

3) Performance evaluation:

  • Performance was assessed using Accuracy, AUC, Precision, Recall, and F1-score, with two-sample t-tests for significance testing.Higher values indicate better performance.Source: 3) Performance evaluation · p. 7

1) Alternative methods:

  • Baselines included centralized and local GCNs, FedMLP, FedGCN, and FedSage+, which uses a local missing-node regression generator with a GraphSage classifier.FedMLP does not use population-graph information, while FedGCN applies FedAvg to local GCNs.Source: 1) Alternative methods · pp. 7–8

2) Model training Setting-up:

  • FedNI used 30 communication rounds for federated inpainting followed by 10 rounds for federated GCN classification, with 10 local epochs per round.The authors chose two-phase training to stabilize GAN optimization and report better, more stable performance than end-to-end training.Source: 1) Alternative methods · p. 8

2) Results and analysis:

FedNI outperformed the comparison methods across two neurological disease datasets. Ablations report contributions from BFS masking, edge prediction, and the discriminator.

2) Results and analysis:

  • Across ABIDE and ADNI, FedNI outperformed all methods on accuracy, AUC, precision, recall, and F1; versus FedSage+, average gains were 1.2%, 1.4%, 0.6%, 2.6%, and 2.1%, respectively.The FedSage+ precision comparison had p = 0.073; the accuracy, AUC, recall, and F1 comparisons had p = 0.009, 0.013, 0.003, and 0.005, respectively.Source: 2) Results and analysis · p. 8

1) Ablation on BFS, edge prediction, and discriminator:

2) Different network inpainting module training strategies:

The experiments compare inpainting-module training strategies, random inpainting, client-count scalability, and computational cost. FedNI’s reported performance improves with more clients, while its additional inpainting cost is described as acceptable.

2) Different network inpainting module training strategies:

  • Federating the generator alone (FL-G) significantly outperformed NoFL-D-G and FL-D across Accuracy, AUC, Precision, Recall, and F1-score (p < 0.05).The authors report no significant improvement from federating the discriminator and suggest local discriminators fit local-graph heterogeneity better.Source: 2) Different network inpainting module… · p. 10

3) Comparison with random inpainting:

  • FedNI’s network inpainting consistently outperformed random inpainting on ABIDE and ADNI.The random baseline generated the same number of nodes using normally distributed features and randomly generated edges.Source: 3) Comparison with random inpainting · p. 10

4) The scalability of FedNI:

  • With local client size fixed at 10% of the dataset, increasing the number of clients from 2 to 10 improved performance while maintaining consistency.The authors report that more clients provide more training samples for global model learning.Source: 4) The scalability of FedNI · p. 10

5) The computational cost:

D. Hyperparameter Discussion

FedNI’s accuracy is robust to the tested loss-weight settings, while neighbor count and local-update epochs expose performance trade-offs. The conclusion reports state-of-the-art performance on public datasets and identifies data heterogeneity as future work.

1) Parameter sensitivity analysis:

  • Across α and β values from 10−2 to 5, FedNI is reported insensitive to the coefficients and consistently better than FedSage+’s best performance.The comparison uses mean accuracy over five independent runs; the default is α = 1 and β = 1 on both datasets.Source: 1) Parameter sensitivity analysis · p. 11
  • Accuracy increases as k rises from 3 to 10, but can be slightly hurt from k = 15 to k = 40; the reported optimum is around k = 10 on both datasets.The authors attribute the pattern to underused neighbors at small k and noise neighbors and oversmoothing at large k.Source: 1) Parameter sensitivity analysis · p. 11, Figure 8 · p. 11

2) Local updating epochs:

  • With local updating epochs E ∈ {1, 5, 10, 15, 20}, larger E requires fewer communication rounds to converge but lowers model performance, especially on ADNI.Loss decreases as communication rounds increase.Source: 2) Local updating epochs · p. 11, Figure 9 · p. 11

V. CONCLUSION

  • FedNI obtains state-of-the-art performance in experiments on public datasets, while addressing incomplete information in local population networks.The conclusion identifies data heterogeneity as future work to further improve performance.Source: §V · p. 11

Related papers

Loading 2112.10166v3…