Continuous diagnosis and prognosis by controlling the update process of deep neural networks.

Continuous diagnosis and prognosis are essential for critical patients. They can provide more opportunities for timely treatment and rational allocation. Although deep-learning techniques have demonstrated superiority in many medical tasks, they frequently forget, overfit, and produce results too late when performing continuous diagnosis and prognosis. In this work, we summarize the four requirements; propose a concept, continuous classification of time series (CCTS); and design a training method for deep learning, restricted update strategy (RU). The RU outperforms all baselines and achieves average accuracies of 90%, 97%, and 85% on continuous sepsis prognosis, COVID-19 mortality prediction, and eight disease classifications, respectively. The RU can also endow deep learning with interpretability, exploring disease mechanisms through staging and biomarker discovery. We find four sepsis stages, three COVID-19 stages, and their respective biomarkers. Further, our approach is data and model agnostic. It can be applied to other diseases and even in other fields.

Adaptive model training strategy for continuous classification of time series.

The classification of time series is essential in many real-world applications like healthcare. The class of a time series is usually labeled at the final time, but more and more time-sensitive applications require classifying time series continuously. For example, the outcome of a critical patient is only determined at the end, but he should be diagnosed at all times for timely treatment. For this demand, we propose a new concept, Continuous Classification of Time Series (CCTS). Different from the existing single-shot classification, the key of CCTS is to model multiple distributions simultaneously due to the dynamic evolution of time series. But the deep learning model will encounter intertwined problems of catastrophic forgetting and over-fitting when learning multi-distribution. In this work, we found that the well-designed distribution division and replay strategies in the model training process can help to solve the problems. We propose a novel Adaptive model training strategy for CCTS (ACCTS). Its adaptability represents two aspects: (1) Adaptive multi-distribution extraction policy. Instead of the fixed rules and the prior knowledge, ACCTS extracts data distributions adaptive to the time series evolution and the model change; (2) Adaptive importance-based replay policy. Instead of reviewing all old distributions, ACCTS only replays important samples adaptive to their contribution to the model. Experiments on four real-world datasets show that our method outperforms all baselines.