Swarm Learning Models to Predict Cancer Biomarkers Outperform Other AI

A new study published in Nature explored swarm learning (SL) models as predictors of molecular alterations directly from standard histopathology images and hypothesized that these models could be a substitute for current artificial intelligence methods, positing that SL could improve prediction performance and generalizability.

Currently, AI is being used to extract potentially clinically useful information from conventional histology images of cancer, including predictive and prognostic molecular biomarkers. However, traditional AI methods require large datasets, and there are practical, ethical, and legal challenges to the data collection necessary for a robust dataset, the researchers pointed out. Training these AI models typically requires that patient-related data be shared with a central repository, but data sharing across institutions may require patients to forfeit their rights to data control.

Federated learning models have addressed this issue by training multiple AI models independently on separate computers. This ensures that the models do not share any input data but do share learned model weights. But the learning progress of the models is governed by a central coordinator, which the researchers argue can monopolize control and commercial exploitation.

These limitations have been addressed by new decentralized learning technologies such as SL. SL models are trained locally and combined centrally without requiring central coordination. The researchers note that SL in healthcare data analysis can lead to equality in training AI models and encourage collaboration among several parties. This could improve the quality, robustness, and reliability of these models in the future.

To evaluate whether SL models could be a substitute for current methods, the researchers developed three SL models designed for the molecular classification of solid tumors based on histopathology images. The models were evaluated against local AI models and a merged model. All models were trained using large datasets from three different patient cohorts.

Following the training of the models, the researchers evaluated their ability to predict BRAF mutational status, a clinical biomarker, in a fourth patient cohort. Overall, the SL models significantly outperformed each of the three local models and performed on par with the merged model. They achieved similar results when the models were evaluated based on their ability to predict microsatellite instability (MSI) and mismatch repair deficiency (dMMR), two other cancer biomarkers.

The researchers also investigated whether the SL models could compensate for the performance loss that occurs when a small dataset is used for training since AI prediction performance generally increases with the size of the dataset. When they restricted the training sets to 400, 300, 200, and 100 patients, the researchers saw significant reductions in the prediction performance for the local models.

While performance losses for the merged model were less pronounced, the SL models exhibited a comparable level of high performance. The SL models also achieved high performance when evaluated for pattern detection using 1,400 images scored by an expert based on relevant pattern or structure presence.

The researchers concluded that using SL models to predict cancer biomarkers offers a potential solution to the most common issues associated with conventional methods. Their SL models significantly outperformed locally trained AI models, which would improve prediction overall.

Additionally, the SL models performed similarly to the merged model but allowed for greater collaboration because they are not controlled by a central coordinator.

The authors argued that these advantages to the SL approach could make the development and training of AI prediction models more accessible to researchers with reduced hardware capabilities and smaller datasets.