Fig 7. Comparison between the ported models and the baselines.

Performance metric is the mean macro F1 scores with 95% CIs. Modularised fine-tuning or updating on additional local features (gray) consistently increases the model’s performance compared to statically using a source model that only uses shared features (teal). The modularised update scenario achieves this without changing the model’s behaviour on patients in the source dataset. The fine-tuning approaches perform almost as well as the global baseline (purple) that trains on the union of shared data. When the percentage of shared features is 80 or 100%, fine-tuning is significantly better than training only locally on the small ‘target’ dataset (green).