Can a 5-to-90-day Mortality Predictor Perform Consistently Across Time and Equitably Across Populations?

Area of Science:

• Health Services Research
• Medical Informatics
• Gerontology

Background:

• Advance care planning (ACP) is crucial for end-of-life care, but many patients die without it.
• Accurate mortality prediction can prompt timely ACP discussions.
• Predictor performance can vary across populations and degrade over time (concept drift).

Purpose of the Study:

• To assess the performance equity and consistency of a novel 5-to-90-day mortality predictor.
• To evaluate the predictor across diverse demographics, geographies, and timeframes, including pre- and during the COVID-19 pandemic.
• To determine if the predictor can support equitable advance care planning.

Main Methods:

• A retrospective dataset of 76,812 adult inpatient encounters was used.
• Predictions were generated for the first day of admission.
• Performance metrics (AUC-PR, recall, precision) were analyzed across subgroups and time periods (2018 vs. 2021).

Main Results:

• Overall AUC-PR remained stable at 29% pre- and during COVID-19.
• Recall and precision showed minimal changes between pre-COVID and COVID-19 periods at tested certainty cutoffs.
• While some subgroups showed minor performance variations, precision at the 12.5% cutoff was largely equitable across most demographics, irrespective of the pandemic.

Conclusions:

• The novel mortality predictor demonstrates consistent and equitable performance across various sub-populations and timeframes.
• This tool can reliably support advance care planning (ACP) by identifying at-risk patients equitably.
• Mortality prediction can be a valuable, consistent, and fair method to encourage ACP conversations.