Examining MIMIC III Electronic Health Records (EHR) for Bias
Introduction
The question of equality has been splashed across the headlines for the last few years, including Me Too, Black Lives Matter, transgender rights, and abortion rights. It is not surprising then that people have raised the question of achieving equality in medical care. Even today, Black Americans experience higher mortality rates.
The American Medical Association stated that minorities receive a lower quality of healthcare, and the CDC reported mistreatment of women in maternity care. This doesn't even capture medical care concerns for those that are uninsured or unable to communicate their health issues due to language barriers. A recent Washington Post article quoted Kortney James, a nurse practitioner and founder of Black Mamas:
“People do not want to go to the hospital. They do not. They feel like ‘if I go there, something bad is going to happen, people aren’t going to hear me, they aren’t going to see me, they aren’t going to care for me.”'
This begs the question, what is happening in our medical systems for such widespread disparity to occur? What underlying mechanisms are at play? To delve deeper into the topic, I looked at the MIMIC III dataset, which contains both demographic and treatment data.
The Data - MIMIC III
MIMIC III is a clinical database including critical care unit patient stays from 2001-2012 at the Beth Israel Deaconess Medical Center in Boston. The detailed dataset has records for more than 40,000 patients that are de-identified, and access is restricted to protect privacy. The dataset is fully described and hosted on PhysioNet via MIT, and it was published in an article in Nature. A demo of the MIMIC III dataset is also available on PhysioNet for those without access. To access the full dataset, I had to obtain a CITI Human Research certificate, a conflict of interest course certificate, and I signed the PhysioNet Credentialed Health Data Use Agreement [see Sources below].
Preprocessing and Data Insights
The files of interest for this project contained information for patients, admissions, ICU stays, prescriptions, lab events, and microbiology events. Missing data was imputed, and groups of individuals were combined to condense the fine distinctions for language, ethnicity, and religion that allowed me to analyze the different groups. Specifically, different types of unknowns were combined for marital status, language, ethnicity, and religion. Also, ethnicity was grouped by those that identified as White, Other (unknown), Black, American (Hispanic, Native American, Caribbean, etc), Asian, Multi-race, and Middle Eastern. I began by looking at the diversity of the patients in the data set. As the plots below show, the majority were white and married, and more people were on Medicare, male, Christian, English-speakers than not.
MIMIC III Demographics
Data Analysis
We can see the average mortality for each of the groups below; the average for all patients was 0.12. The mortality rate for some groups, such as those Medicare and widowed individuals, are logically high as they are very likely correlated with an older age group. Interestingly, those in the category of self-pay, Medicaid, women, Jewish people, and "other" language groups also have a higher mortality rate.
Understanding the factors behind why these groups have higher mortality rates is worth further investigation and finding answers to these questions:
- Do the self-pay group wait to get medical help because they have to pay out-of-pocket, causing health problems not to be caught in time?
- Does the Medicaid group have poverty related issues such as diet and stress that factor into health?
- Does the "other" language group have communication problems such that they are not understood?
MIMIC III Mortality Rate
I further examined six features: the mean length of stay (los) in the ICU and in the emergency department, number of admissions per patient, number of prescriptions, and number of labs and microbiology labs. These features best represented quality of care for the patients. The correlation heatmap between these features revealed that the number of prescriptions and the number of lab/microbiology tests were highly correlated (0.82/0.74), as well as the number of labs and microbiology labs (0.73). The more tests that were done, the more prescriptions were prescribed, which seems logical.
Somewhat correlated were the number of admissions to the number of prescriptions (0.64), labs (0.57), and microbiology labs (0.45). The following plot helps us visualize the non-correlated features of length of stay in the ICU and number of lab tests, where we clearly see that the group of self-pay patients had fewer labs and fewer prescriptions. Also, we observe that the number of prescriptions (circle size) increases as the number of labs does. However, the same relationship is not shown with respect to the length of hospital stay, since the former is correlated and the latter is not.
Average length of stay in the ICU versus the number of labs test
ANOVA and Pairwise t-test
Rather than looking at the barcharts for each of our six outcomes (los ICU, los ER, number of admissions, labs, microbiology labs, and prescriptions), it is better to determine if they were statistically significant or not. To that end, I performed an ANOVA test with a pairwise t-test between each feature for six different features (ethnicity, gender, language, insurance type, marital status, and religion) for each outcome. To summarize the large number of results, the following were among the key statistically significant differences:
- Language
- The "other" languages group (excluding Spanish and English) had longer ICU stays than Spanish and English speakers, and they had fewer prescriptions than the English speakers.
- The English speakers had fewer labs than the "other" languages and Spanish speakers.
- Insurance Type
- The insurance type was very often significantly different for each outcome. So much so that it defined the type of treatment received, especially so with the "self-pay" group. The "self-pay" group had shorter stays in the ICU and ER, fewer ER admissions, prescriptions, lab tests, and microbiology tests than any other group.
- Ethnicity
- The highest number of ER admissions was for the Black ethnicity group, compared to nearly every other group.
- Between the Black and White ethnic groups, the length of stay in the ER was longer for the Black group.
- The length of stay in the ICU was shorter for Asians than for every other ethnicity across the board.
- Marital Status
- The widowed group was most often different, which may be correlated with age. They had fewer prescriptions and lab tests, and they had longer stays in the ER than single or married people.
- Religion
- Jewish people had more admissions to the ER than Christian people, fewer prescriptions, and longer stays in the ER.
- Gender
- The length of stay in the ER was longer for women, and women received fewer prescriptions.
- Small Groups
- The groups of people that are Middle Eastern, Multi Race, and Life Partners are very small, which is why no significant relationships came out of the analysis. While these groups may have different treatment, this dataset doesn't capture that difference.
Conclusions and Future Work
Within this dataset I identified health disparities that were not simply a Black and White issue, but one that extends to different ethnicities, religions, genders, and language groups. A third party or internal review could help to resolve this problem, but at present equal treatment in healthcare remains complex. Future work for analysis would involve breaking the groups down further into age brackets, inquiring about unknowns in the data, and looking at long term mortality amongst the groups.
Sources:
Johnson, A., Pollard, T., & Mark, R. (2016). MIMIC-III Clinical Database (version 1.4). PhysioNet. https://doi.org/10.13026/C2XW26.
Goldberger, A., Amaral, L., Glass, L., Hausdorff, J., Ivanov, P. C., Mark, R., ... & Stanley, H. E. (2000). PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals. Circulation [Online]. 101 (23), pp. e215–e220.
About Author
