Robustly federated learning model for identifying high-risk patients with postoperative gastric cancer recurrence

Abstract

The prediction of patient disease risk via computed tomography (CT) images and artificial intelligence techniques shows great potential. However, training a robust artificial intelligence model typically requires large-scale data support. In practice, the collection of medical data faces obstacles related to privacy protection. Therefore, the present study aims to establish a robust federated learning model to overcome the data island problem and identify high-risk patients with postoperative gastric cancer recurrence in a multicentre, cross-institution setting, thereby enabling robust treatment with significant value. In the present study, we collect data from four independent medical institutions for experimentation. The robust federated learning model algorithm yields area under the receiver operating characteristic curve (AUC) values of 0.710, 0.798, 0.809, and 0.869 across four data centres. Additionally, the effectiveness of the algorithm is evaluated, and both adaptive and common features are identified through analysis.

Fig. 1. DCA curves of seven models using data from four data centres.

The solid grey line assumes that all patients were involved in the R-AGC group while the black line assumes no patients were involved. The threshold probability was the point where the expected benefit of the treatment and treatment avoidance were equal. The result showed that the net benefit of the RFLM was greater than that of the clinical model (range, 0.00–1.00). RFLM robust federated learning model, FedAvg, Fedprox, Moon, HarmoFL, pFedFSL and pFedMe are the comparison test algorithms.

Fig. 2. Threefold cross-validation ROC curves for the four centres.

The blue curve represents the average AUC of the threefold curve. The grey areas represent the upper and lower limits of the ROC curve. The error band in the grey areas is the upper and lower boundary of the three-fold cross-verified ROC curve. Mean the AUC average for three-fold cross-validation, Std standard deviation.

Fig. 3. Correlation heatmap of common recurrence features and adaptive features.

a R-AGC common features correlation heatmap. b R-AGC adaptive features correlation heatmap. A_1 first feature at centre A, R-AGC recurrent advanced gastric cancer.

Fig. 4. RFLM Algorithm Result Analysis Diagram.

a The heatmap shows the information acquired by the RFLM for images in the recurrent and nonrecurrent classes. The red areas indicate a high level of model attention, while the blue areas indicate a low level of model attention. b The Euclidean distance plots depict the distance between the common and adaptive features of the four central data points. The left side represents common features, while the right side represents adaptive features. c The score charts illustrate the positive and negative images of the four data centres evaluated by the RFLM. Statistical test: Independent t-test (two-tailed). RFLM robust federated learning model, NR-AGC no recurrent advanced gastric cancer, R-AGC recurrent advanced gastric cancer, p significance value.

Fig. 5. A representative dataset was generated based on the WGAN.

NR-AGC nonrecurrent advanced gastric cancer, R-AGC recurrent advanced gastric cancer.

Fig. 6. RFLM algorithm diagram.

a Construction process of the robust local model in the RFLM. b Details of generating robust weight parameters in the robust local model. c Feature extraction and feature classification in the RFLM. CBAM convolutional block attention module, NR-AGC nonrecurrent advanced gastric cancer, R-AGC recurrent advanced gastric cancer, GCN graph convolutional neural network, WGAN Wasserstein generative adversarial network.