Transfer Learning Across Diseases Opens a Novel Route Towards Pandemic Preparedness

Abstract Pandemic preparedness has gained high relevance throughout the SARS-CoV-2 pandemic, necessitating optimized therapeutic strategies for early phase disease management. Predicting patient-specific outcomes, particularly for novel infectious diseases with complex multimorbidity patterns, presents significant challenges due to limited availability of disease-specific patient datasets and an incomplete understanding of disease mechanisms early in the disease wave. We propose a concept for rapid predictive modeling of disease outcomes based on the hypothesis of common disease-promoting mechanisms and leveraging transfer learning across different disease types. Introducing the Medical Reference Space (MRS) as a unifying framework to integrate heterogeneous clinical data structures, allows us to monitor therapy response dynamics and predict individual outcomes across various diseases. We performed a retrospective analysis on mechanically ventilated patients from intensive care units, comprising 495 COVID and 6212 non-COVID patients. The results demonstrate that MRS-based models, trained on non-COVID patients, effectively predict individual survival in COVID-19 patients, comparable to models trained exclusively on COVID data. Moreover, the MRS approach provides insights into the impact of rapidly evolving events, such as sepsis, on the predictivity of machine learning-based prognostic systems indicating conceptual limitations for patient-specific decision support systems. Our findings highlight the potential of using retrospective patient data repositories for transfer learning as a rapid strategy for predicting disease outcomes of new evolving diseases, opening a route towards future pandemic preparedness.