Skip to content

rationale.md

Latest commit

 

History

History
36 lines (35 loc) · 2.95 KB

rationale.md

File metadata and controls

36 lines (35 loc) · 2.95 KB

Data and Algorithmic Bias Workshop

Use Case: The peril of inaccurate Sp02 measurements in the EHR

by Leo Anthony Celi on some airplane from somewhere to somewhere, originally to be applied in the Australian Datathon

1. Talk: Introduction

  • Pulse oximetry has been found to be a poor proxy of arterial oxygen saturation among individuals of color
  • A certain value means different depending on the skin tone
  • How would this issue impact the development of algorithms in healthcare where Sp02 is one of the features?
  • Specific aim: To determine whether the issue of poor accuracy of pulse oximeters also affects indigenous patients in Australia

2. Breakout session #1: Design the study to address the research question

  • Inclusion and exclusion criteria: pros and cons of focusing on a specific condition
  • Outcome: Sp02 – Sa02 gap
    • Definition of Sp02 – Sa02 gap
    • Representation: first pair? mean of pairs across the entire ICU LOS? mean of pairs during the first 24 hours? standardized based on number of pairs?
  • Features including confounders affecting Sp02 (e.g., hemoglobin, bilirubin, vasopressor use) and those affecting the relationship of So02 and outcome (e.g., age, illness severity, comorbidities)
  • Methodology: traditional regression, causal inference

3. Flow diagram and Table 1, distribution of outcome and features across indigenous vs. non-indigenous patients

4. Breakout sessions #2: Sources of bias in the study design

  • Sampling selection bias as regards who have access to ICU care (look at literature)
  • Sampling selection bias as regards the hospital’s ICU admission criteria (look at literature)
  • Sampling selection bias based on study inclusion and exclusion criteria
  • Measurement bias from irregular sampling
  • Data imbalance: indigenous vs non-indigenous, outcome imbalance
  • Unmeasured confounding: patient-provider sex and race concordance, other drivers of clinical decision-making that are not captured by EHR

5. Exploratory data analysis

  • Look at proportion of indigenous vs indigenous patients of every cohort that is excluded
  • Number of blood gases indigenous vs non-indigenous
    • during the first 24 hours standardized based on illness severity
    • during the entire ICU length of stay standardized based on ICU length-of-stay
  • If only the first pair is considered, time to the first arterial blood gas indigenous vs non-indigenous patients

6. Breakout session #3: How to address the different sources of data bias

  • Irregular sampling of arterial blood gases
  • Imbalance in the number of indigenous vs. indigenous patients

7. Talk: What is the effect of indigenous status being embedded on the data?

8. Breakout sessions #4: What are the potential consequences of the inaccuracy of the pulse oximetry on indigenous patients on the development of decision-support algorithms? Provide use cases.

9. Breakout sessions #5: How should we evaluate whether a decision-support algorithm will not harm indigenous patients further?