Just got back from a conference, and it’s always so interesting to experience the difference between reading a paper vs talking to its author & getting a more contextualized picture of how thoughtful or thorough the research is. As I thought about that more, I decided that it might be useful (or at least fun) if I wrote about some of my research & explained some of the context and findings. I assume it’d be more fun to read this than the papers themselves.

For my first research post, I wanted to start with the work I’m most proud of: my Master’s Thesis last year, which looked at the possible links between end-of-life treatment decisions and trust in the doctor-patient relationship.

Link to slides from my Research Qualifying Exam presentation.

In 2016, one of our clinical collaborators (Leo) pointed Marzyeh and me to a set of papers [1,2,3] which studied racial disparities in end-of-life (EOL) care in North America. They looked at “aggressive” care (i.e. high-risk interventions that are unpleasant, like a tube in the throat to try to prolong life) vs comfort-based case (i.e. hospice). The main finding was that white patients received smaller amounts of aggressive care than nonwhite (in particular African American and Hispanic) patients.

Renowned author and surgeon Atul Gawande has extensively studied end-of-life decisions and how patients who are empowered to make informed EOL decisions overwhelmingly choose to live with dignity instead of overly medicalized procedures (and that they do so with much higher levels of well-being and sometimes even longer lives). But these studies found that white patients seemed to be transitioning to hospice care earlier and at higher rates. Why?

Of course the first thing I thought of was implicit bias from the caregiver leading them to unknowingly make different care decisions. And while I don’t want to minimize that because it relates to the whole doctor-patient relationship, some of the researchers speculated an even sharper hypothesis for what could be causing this: mistrustful patients are ignoring their doctor’s recommendation for hospice.

For anyone unfamiliar with end-of-life care and decisions, it’s hard to appreciate the gravity of the situation without context. There are serious dignity questions at stake about how you or a loved one would like to go, and it’s especially hard to make that decision for someone else or when you’re really cared & confused. West Virginia Public Broadcasting produced a one-hour documentary about the struggles and process of end-of-life looks like. All I can do is write words on a page and tell you what other people say, but this video lets you actually hear from patients and family members experiencing this in their own words. If you want to actually appreciate what people go through, you need to hear from them directly.

Imagine your father is in the hospital. His doctor approaches you and says that after a week of critical care, she believes they’ve done everything they can for your dad. She suggests that it might be time to consider a transition to withdrawing treatments and making him comfortable for what little time he has left.

If you trust your doctor — if you believe that she’s looking out for your family’s best interests — you might strongly consider her advice. But if you don’t trust her — if you think the healthcare system doesn’t want to waste the resources on you or that they want your father’s bed for someone else — then you might say “Keep fighting and do what you can to save my dad.” Currently, this scenario has not been thoroughly studied to understand whether that’s definitively what’s happening to create the existing racial disparity, but that is how we believe the causal mechanism would be causing this. We wanted to study that further.

Previous work has suggested that racial disparities in health outcomes may reflect higher levels of mistrust for the healthcare system among black patients. Family members of African American patients are more likely to cite absent or problematic communication with physicians about EOL care. When the doctor-patient relationship lacks trust, patients may believe that limiting any intensive treatment is unjustly motivated, and demand higher levels of aggressive care. To better understand what could be causing this lowered trust and diminished communication, I read Harriet Washington’s 2007 book Medical Apartheid, and it was the most informative and contextualizing work that I read throughout this entire multi-year project.

American Medicine’s Dark History of Exploiting Black Bodies

In her 2007 book, Washington suggests that the medical exploitation of African Americans by white institutions throughout American history has created “Black Iatrophobia.” The most infamous example of medical exploitation on black Americans was the Tuskegee Syphilis Experiment from 1932-1972, where black men were tricked into not receiving any medicine for syphilis (despite the invention of Penicillin) so that scientists could study the progression of the disease. But while the might still be the most notorious example, it is far from the only example.

Medical abuse in America has plagued the African American community from the beginning of US history all the way through modern times. Going back to 1801, Thomas Jefferson injected 80 of his own slaves with smallpox to prototype vaccines. In the late 1840’s, Dr. James Marion Sims (considered by some to be “the Father of Gynecology”) surgically experimented on and mutilated his female slaves — who were unable to refuse his operations — without anesthesia. Until the early twentieth century, medical schools were using bodies graverobbed from black cemeteries as cadavers for their dissection trainings.

As recently as 1987-1991, US scientists administered as much as five hundred times the approved dosage of the experimental Edmonton-Zagreb vaccine against measles to African American and Hispanic babies in Los Angeles without communicating to the parents on informed consent forms that the vaccine was experimental or unlicensed

Harriet Washington’s thesis of Black Iatrophia has manifested in the published literature. Socialized mistrust of the medical community in minority groups has been established as a factor in care differences. With this broader understanding from this book, we wanted to further study this notion of “mistrust” and how it related to EOL care. The first thing we needed to do what try to establish what mistrust was and how to quantify it for the patients.

Modeling Mistrust Algorithmically

The first thing we did was replicate that the racial disparity existed in our two ICU datasets, which we found largely to be the case (especially for mechanical ventilation, though less extreme for vasopressors). But even after controlling for severity of illness, the differences in treatment across race still existed.

The next thing we wanted to do was to split the group into trustful and mistrustful to see whether that gap was even larger, but we did not have scores for each patient indicating the quality of trust in the doctor-patient relationship. We did, however, have some clues about trust for some patients.

An example of a patient’s clinical note documenting their explicit mistrust of their doctors and how that manifested in frustration and noncompliance.

Clinical notes written by doctors and nurses provide a very vivid and comprehensive view into the interaction between the patient and their caregivers. In the above example, we can see a patient who was very frustrated and mistrusting. The relationship with their providers was clearly poor, but not every patient’s notes is as easy to discern as this one’s.

An illustration of how we used Machine Learning to create mistrust scores for all patients by learning patterns about the few hundred “labeled” examples and extrapolating to all patients.

In order to derive a “mistrust score” for every single patient, we used the “clear” mistrusting notes as anchors and tried to determine how similar other patients were to those cases using a simple supervised Machine Learning algorithm (technically speaking, we derived three different mistrust scores through a process like this in order to avoid over-committing to any one definition of what “trust” is or how it definitively manifests).

Figure showing the mimiciii.chartevents table from the MIMIC database. This table includes many events, including all of these indicators above, documented by the caregiver staff.

The “inputs” for this model were a comprehensive set of documented interactions in the chart events table from the EHR. It captures a very large number of interpersonal interactions in the doctor-patient relationship, such as whether the patient’s pain is being managed well, whether the patient is restrained and treated as a threat, whether the patient got their hair washed, how agitated the patient appears, how frequently the care team is communicating with the family, and much more.

We looked at the most predictive features for the mistrust scores, and saw that high levels of mistrust were associated with agitation, pain, and patients being restrained. On the other hand, the most trustful patients tended to enjoy little-to-no pain, low agitation, higher levels of healthcare literacy, and more consistent family communication. 

Correlation matrix between severity scores and mistrust scores. A 1.0 means perfect agreement (hence 1s on the diagonal comparing a given score to itself).

After fitting these models, we wanted to sanity check them to ensure our algorithmic scores that we are calling “mistrust” seem to align with our intuitive notions of what that should mean. We wanted to ensure that the learned scores were not simply capturing some existing trend such as severity of illness: we saw that the scores had moderate correlation with one another (r=0.26), the severity of illness scores had strong correlation with one another (r=0.68), yet the mistrust scores had virtually no correlation with severity of illness (r<=0.05). While this experiment doesn’t prove that the scores are capturing “trust,” it does show that whatever it’s capturing is decidedly not  just how sick the patients are.

These plots are cumulative distributions of the percentage of patients given some duration of treatment, with the dotted vertical lines indicated the median of the population

To understand these plots, let’s look at an example. In the top left quadrant, we see that the median black patient received 3,286 minutes of mechanical ventilation whereas the median white patient only received 2,454 minutes. The difference is about 14 hours, on average. But when we split our population into trustful and mistrustful (instead of white and black), we see that treatment disparities are much larger across trust-based cohorts than race-based cohorts. This trend was true for 3/3 metrics for mechanical ventilation and 2/3 metrics for vasopressors.

Of course anyone who has experience in EOL care would have already been able to tell you how essential trust, respect, and communication are. The main concern for these racial disparities is that levels of mistrust are higher in one population than in another.

CDF of the 3 mistrust scores for white patients (blue) and black patients (orange). The scores were scaled to be zero mean and unit variance at the full cohort level.

We find that as anticipated, the majority of metrics indicated that the black population of patients had a statistically significantly higher level of mistrust. This (in conjunction with the above findings that mistrustful patients receive longer durations of care) reinforces our belief that mistrust could be the mechanism which is causing racial disparities in EOL. Although  these metrics are likely not perfect, they show as a starting point that there is a strong signal which needs to be better understood and potentially addressed. There are many ways to improve upon this work, and I think that further study is warranted.

Future Work

There are lots of potential follow-up questions one could look at, some of which are technical, some of which are social/ethical/etc, and some of which are both. I’m happy to chat about ideas/reactions that people have!

  • Patient surveys. Don’t speculate how the patients feel. Ask them!
    • Everything in this analysis was through the eyes of what the caregivers recorded in the notes / chart events.
    • Work with an interdisciplinary team (sociologists, caregivers, patients, etc) for better definitions of trust.
  • Doctor Biases. Do some doctors have larger racial disparities than others?
    • Do some doctors have higher levels of mistrust than others?Do some doctors have high mistrust among black patients but not for white patients?
  • Causal Inference. Was mistrust the cause of different treatment patterns?
  • Cultural: This analysis only focused on white and black patients.
    • Different cultures handle death & dying differently.We have an autopsy-based metric, but some people wouldn’t want the body cut open at all.
  • Policy: What should we do if we believe the status quo hurts nonwhite populations? What should change?
    • One could mandate standardized care. But people choose aggressive care. Is it too paternalistic to tell them otherwise?
    • Is the problem that the choices are not truly informed & that people don’t necessarily know what they’re signing up for? If so, how do we better help people make whatever choice they want in a truly informed way?

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s