Back

Prospective Evaluation of a 90-day Mortality Prediction Model: From Silent Pilots to Real Time Deployment in the EHR

Rossi, L. A.; Roberts, L. M.; Zachariah, F. J.

2023-01-28 health informatics
10.1101/2023.01.25.23284977 medRxiv
Show abstract

Prognostication in oncology is increasingly difficult due to the rapid evolution of therapies with significant improvement of survival. Accurate prognostication is essential to provide optimal, value-driven end of life care for cancer patients, and can promote goals of care (GOC) conversations with the potential to minimize chemotherapy or ICU utilization in the last weeks of life, and possibly increase hospice admission and length of stay.1 There are several recent publications on the application of machine learning for prognostication.2,3 We developed a 90-day mortality prediction model trained with data in the Electronic Health Records (EHR). After a non-interventional pilot stage, we deployed the model in February 2021 in the real-time Electronic Health Record Epic infrastructure of our cancer center. Here we present the model and evaluate its overall performance for the first 7.5 months since the go-live and outline our evaluation process for the next stages.

Matching journals

The top 5 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.