Background Accurate timing of ACP is considered challenging. Available tools are not widely implemented, can be unfeasible to screen large populations, and require prior awareness to be used proactively in individual patients. We are developing an automated signaling tool for early identification of patients who might benefit from ACP. In this talk, we will focus primarily on the model we created to predict life expectancy, which we consider to be a first step towards identifying these patients. Additionally, we will discuss further model development and deployment in a real-world setting.

Methods We used machine learning (ML) and natural language processing (NLP) techniques to train a recurrent neural network on 1234 medical records of deceased patients. We trained several models, and compared the best-performing model to doctor’s performance on a similar task as described in the literature.

Results While doctors were correct in 20% of the cases (allowing an error margin of 33% around the actual moment of death), our best-performing model attained a prognostic accuracy of 29%. Being overly optimistic about life expectancy harms anticipation to end-of-life care. Our model was less likely to overestimate life expectancy (in 31% of the incorrect prognoses) than doctors (63%) were.

Conclusions Our research shows that ML and NLP offer a feasible approach to predicting life expectancy. The results are promising, given that our model is trained on a relatively small data set. Current work focuses on further model development with an increased dataset, and implementation of the tool in primary care facilities.