Mortality prediction models may support medical decision-making in older individuals. However, these models often involve manual scoring, are hindered by methodological issues, and their performance in persons with dementia is unknown.This study aims to leverage machine learning techniques in a large routine care database to automate identification of those with high mortality risks.We utilized routine primary care data from a large nationally representative cohort of 355,958 community-dwelling older adults in XXX. We developed three manually defined models, an automated machine learning model and a model including time series to predict 1- and 5-year mortality based on the electronic health record. A split-sample design was used evaluate discrimination and calibration, while using method specific internal validation procedures. Model performance was specifically assessed among individuals with dementia.The best models for both 1- and 5-year mortality were the time-series models with AUC-ROCs of respectively 0.861 and 0.846. Performance in people with dementia was lower with AUC-ROCs of 0.727 and 0.748. Calibration was good but slightly overestimated the risks for individuals with dementia. Although the time-series models performed best, there was no significant performance difference compared to the manually defined models.We developed accurate, interpretable, and potentially easily implementable models for mortality prediction based on routine primary care data in older adults. The model could help primary care physicians automatically identifying individuals at risk and facilitating timely advanced care planning, but external validation remains needed. Caution is needed when applying the model to persons with dementia.This study demonstrates the feasibility of developing accurate, interpretable, and potentially implementable mortality prediction models using routinely collected primary care data. Our findings show that (electronic) primary health care records contain important predictive information, especially regarding diagnostic codes and contact information, and that they can contribute to automating individual risk assessments. However, the captured information varies among subgroups such as those with dementia. Following research should focus on the robustness of the models through external validation and the implications for practice.