Figure 5.

Evaluating fine-tuning to generate individual models on the discovery cohort. (A) For a specific individual under investigation, we first trained 5 global models using all other individuals by resampling the training and validation set for the deep learning training process. Next, we split records of the individual of interest by days into training, validation, and test sets and fine-tuned the global model using the training and validation set. We evaluated the performance by the area under the ROC curve for both the global and the individual fine-tuned models for (B) 5 separated models and (C) ensemble models. ROC: receiver operator characteristic.