# Sex Trouble: Common pitfalls in incorporating sex/gender in medical machine learning and how to avoid them

**Authors:** Kendra Albert, Maggie Delano  
**arXiv:** [2203.08227](https://arxiv.org/abs/2203.08227)  
**Date:** 2022-03-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2203.08227v1/

## Key points

- False assumptions about sex and gender in the medical system
- Three common mistakes made by researchers dealing with sex/gender data: "sex confusion," "sex obsession," and "sex/gender slippage"
- Pitfalls of misusing or misinterpreting sex/gender data in machine learning studies based on electronic health record (EHR) data
- Importance of working with data experts and collaborating with clinicians to handle such data properly
- Clear descriptions of how sex/gender data were collected and utilized should be provided in research papers
- Acknowledgment of limitations or changes in variables over time
- Careful incorporation of gender, considering multiple dimensions beyond just biological sex
- Need for machine learning researchers to engage thoughtfully with questions of sex/gender for inclusive research outcomes that serve all patients, including transgender individuals.

## Summary

False assumptions about sex and gender are deeply ingrained in the medical system, leading to binary, static, and concordant interpretations. Machine learning researchers must be aware of these assumptions to avoid perpetuating them. This perspective piece highlights three common mistakes made by researchers when dealing with sex/gender data: "sex confusion," which refers to the failure to understand what sex in a dataset signifies; "sex obsession," which assumes that sex assigned at birth is the relevant variable for most applications; and "sex/gender slippage," which conflates sex and gender even when only one is known. The authors then discuss how these pitfalls manifest in machine learning studies based on electronic health record (EHR) data. EHR data is widely used for various purposes, including retrospective analysis of patient outcomes and developing risk prediction algorithms. However, the misuse or misinterpretation of sex/gender data can lead to biased results and inadequate care for transgender individuals. To avoid methodological pitfalls, researchers are advised to work with data experts who have experience handling such data. Collaboration with clinicians can provide insights into how the data is used in clinical practice. Additionally, clear descriptions of how sex/gender data were collected and utilized should be provided in research papers. It is crucial to acknowledge any limitations or changes in variables over time. Incorporating gender carefully is another important consideration. Gender encompasses multiple dimensions such as identity, relations, roles, and institutionalized aspects. Researchers should not conflate sex and gender or assume that sex represents an absolute truth while disregarding individual gender identities. Overall, this perspective piece emphasizes the need for machine learning researchers to engage thoughtfully with questions of sex/gender in their research and algorithm development processes. By avoiding methodological pitfalls and considering the complexities of sex/gender, researchers can produce more inclusive research outcomes that better serve all patients, including transgender individuals.