Optimizing Health Equity through AI — Defying racial bias that plagues medical appointment scheduling.
In almost every realm - from who gets to work from home to how families are coping with distance learning - the COVID-19 pandemic has laid bare the deep inequalities in our society. In health, the hardest hit have been traditionally marginalized people - in particular Black, Brown and Indigenous communities, where health inequities were already present. On Wednesday, September 9, scholars from Santa Clara Leavey School of Business discussed the racial health equity problems caused by algorithms - and how to solve them - in the realm of appointment booking drawing from their research and paper under review: Overbooked and Overlooked: Racial Bias in Medical Appointment Scheduling. The talk was hosted by Professor of Law, Colleen Chien, co-curator of the Artificial Intelligence for Social Impact and Equity Series currently ongoing through the High Tech Law Institute at Santa Clara. My personal interest in this topic stems from my previous work experience as a biomedical engineer and regular clinic patient. I found the authors' work here to be inspiring, and I hope you do too.
MEDICAL APPOINTMENT SCHEDULING.
All access to healthcare (or at least most) starts with an appointment. Some people get 8 A.M. appointments, and some get 4 P.M. appointments. But how does that actually work for clinics that employ schedule optimization technologies?
At some clinics, many patients do not show up for scheduled appointments. In fact, the ‘show rate’ for patients at the clinic observed in the study was only about 74%. So why do we wait so long when we go in for checkups?
A LOOK AT THE SCHEDULING PROBLEM.
The problem stems from various solutions. Providers can charge non-showing patients, but this is often unavailable for some (e.g. non-profit clinics) and still does not fix the problem. Providers could also send reminder calls and texts, but that is how we arrived at the 74% figure. The most commonly used practice relies on a ‘double-book’ approach — that is, to overbook patients in one appointment slot. This allows the provider to have a full day, optimizing the revenue of the clinic, but of course, means that sometimes we have to wait a long time.
According to Dr. Samorani, AI appointment scheduling systems work by employing two steps, (1) machine learning, and (2) schedule optimization. The first step is to predict the “show” probability of each patient. Each machine learning show prediction can be based on a variety of factors including (a) patient demographics (which could include race), (b) previous no-show data, and (c) other information (e.g. how long ago the appointment was scheduled).
After predictions are made for each patient, a schedule optimization component slots the patients based on their overall predicted no show probabilities throughout the day in an attempt to achieve a “desired result.” This is known as an objective function.
The method, used widely in clinics and the industry today, is called the Traditional Objective Function (TOF). The ‘desired result’ of the TOF is to minimize patients’ wait times while equally reducing providers’ overtime, thereby reducing costs. The TOF accomplishes this task very well. However, the TOF results in significantly longer waiting time for the patients at the highest risk of no-show, which are usually minority groups.
RACIAL BIAS IN SCHEDULING.
As it turns out, among the groups investigated, Black patients end up waiting the longest because they are predicted to have the highest no-show rate. In computational experiments based on the data from the clinic under study, Black patients ended up waiting 33% longer (on average) than non-Black patients. This is a significant problem because longer wait times exacerbate existing issues of inferior access to healthcare, poor health care quality, and worsen healthcare outcomes for certain populations. Over the past few months, the authors have worked to tackle this problem.
The authors’ research considers the issue from multiple angles. As Dr. Samorani elaborated, there are likewise two intervention points at which one can attack the problem. The first is at the machine learning point (reduce disparity) to make the show probabilities correlate less with race. The second is at the schedule optimization point (change the TOF to have a desired result of fairness).
METHODS & RESULTS.
In the first method, the authors try to mitigate some of the factors that are correlated to race or otherwise remove them completely:
- But even after reducing all demographic factors correlating with race, we only get a middle-of-the-road solution — schedule quality decreases and racial disparity is still present. Dr. Samorani shares how these are the two fundamental problems with this approach.
In the second method, the authors change the TOF so that the desired result is to reduce the waiting time of the racial group expecting to wait the longest. The authors refer to this method as the “race-aware” approach:
- Using this method, the authors achieved a schedule quality equivalent to [not statistically different from] the state-of-the-art scheduling method AND nearly 0 racial disparity.
Considering this approach, the authors acknowledge how some might argue that their “race-aware” approach is inappropriate because it takes race explicitly into account. Even though they (and we) believe using race-aware policies is morally warranted, the authors have developed an alternative version of the TOF using the second method which they call “race-unaware.” In this TOF variant, the ‘desired result’ is to reduce the waiting time of the patient expecting to wait the longest:
- Using this alternative, the authors achieved a middle-of-the-road schedule quality with less racial disparity.
In summary, alternative power or weight can be given not only through machine learning or optimization to fix current problems in healthcare, but also in other areas.
MOVING FORWARD.
It is important now, more than ever, to combat these disparities and implement countermeasures to our existing systems and before our technologies are developed. The authors here highlight some concrete ways that we can approach problems in medical appointment scheduling, and we are excited to keep learning and looking for more solutions in other areas.
We are thankful for the opportunity to have spoken with some of the authors. I hope that our exploration, albeit a light one, on the topic has shed some light on a problem in healthcare not generally known, and is useful for others when thinking about the day-to-day and greater scheme of health equity.
If you would like to read more on the work here please see the under review paper here.
Other work that Santa Clara Law’s High Tech Law Institute has done can be found here. As a Santa Clara Law Student, you can find the class (The Business, Law, Technology, and Policy of Artificial Intelligence) where this blog's work was initiated here.
About the author of this post:
Dalton Chasser is a 2L focusing on Health and Intellectual Property Law at Santa Clara University School of Law.
