Personalized Risk-AdaptIve Surveillance (PRAISE) - Implications of Algorithmic Bias

ABSTRACT
Algorithmic bias is an emerging and highly relevant topic in health policy that draws attention to the idea that
seemingly well-performing predictive algorithms built using biased data can propagate systemic biases and
disparities existing in clinical practice. Our primary goal in the parent grant was to innovatively use machine
learning based risk prediction and value-of-information methodology to develop a personalized risk-adaptive
surveillance (PRAISE) framework and assess its impact on outcomes, including costs, time to recurrence and
survival, for colorectal cancer survivors who have been treated for their primary cancer and are under
surveillance for recurrence. Decision-making in the PRAISE framework is driven by dynamic risk predictions for
recurrence, which identify high-risk patients to target for more frequent surveillance testing. However, potential
sources of racial/ethnic disparities along the cancer care continuum from diagnosis to survivorship can lead to
distorted risk predictions during surveillance. Using distorted risk predictions can have implications for
decision-making, potentially propagating and exacerbating biases that exist in clinical practice and resulting in
poorer outcomes for certain subgroups. The overarching goal of the proposed research is to understand and
address algorithmic bias in the PRAISE framework. Specifically, we will first characterize heterogeneity in
current practice with respect to surveillance testing patterns, recurrence detection and survival across
racial/ethnic subgroups for patients diagnosed with and treated for colorectal cancer (Aim 1). This will help us
better understand the sources of bias in our data and will better inform our approach in Aim 2, where we will
use emerging and novel methods to mitigate racial/ethnic bias in our previously developed dynamic risk
prediction model for colorectal cancer recurrence. Finally, in Aim 3, we will develop an outcomes-based
framework to assess the implications of using a biased versus a bias-corrected risk prediction model to guide
surveillance testing among colorectal cancer survivors, specifically through their effect on decision-making and
subgroup-specific health and cost outcomes. This important work will motivate the use of new methods for
addressing bias in risk prediction models in cancer and other clinical areas.

Grant Number: 5R37CA218413-07
NIH Institute/Center: NIH
Principal Investigator: Aasthaa Bansal