Deep learning-based prediction of one-year mortality in the entire Finnish population is an accurate but unfair digital marker of aging

BackgroundAccurately predicting short-term mortality is important for optimizing healthcare resource allocation, developing risk-reducing interventions, and improving end-of-life care. Moreover, short-term mortality risk reflects individual frailty and can serve as digital aging marker. Previous studies have focused on specific, high-risk populations. Predicting all-cause mortality in an unselected population incorporating both health and socioeconomic factors has direct public health relevance but requires careful fairness considerations. MethodsWe developed a deep learning model to predict 1-year mortality using nationwide longitudinal data from the Finnish population (N = 5.4 million), including >8,000 features and spanning back up to 50 years. We used the area under the receiver operating characteristic curve (AUC) as a primary metric to assess model performance and fairness. FindingsThe model achieved an AUC of 0.944 with strong calibration, outperforming a baseline model that only included age and sex (AUC = 0.897). The model generalized well to different causes of death (AUC > 0.800 for 45 out of 50 causes), including COVID-19 which was not present in the training data. The model performed best among young females and worst in older males (AUC = 0.910 vs. AUC = 0.718). Extensive fairness analyses revealed that individuals belonging to multiple disadvantaged groups had the worst model performance, not explained by age and sex differences, reduced healthcare contact, or smaller training set sizes within these groups. ConclusionA deep learning model based on nationwide longitudinal multi-modal data accurately identified short-term mortality risk holding the potential for developing a population-wide in-silico aging marker. Unfairness in model predictions represents a major challenge to the equitable integration of these approaches in public health interventions.