Download as pdf or txt
Download as pdf or txt
You are on page 1of 3

1 Three Papers Worth Reading (# 3 will surprise you!

)
1.1 Operating Room Scheduling for Doctors
This paper is an enjoyable read because the authors gave an overview of the model they constructed without getting bogged
down in details. Moreover, the model remained faithful to solving the underlying practical problem; it did not rely on
unnecessary or unrealistic assumptions.
Constraints for the MIP were decided on by consulting key stakeholders. One criticism that was repeatedly brought up
during discussion is that the model was more concerned with the preferences of the doctors than with how the schedule process
affects patients. While there is some merit to this argument- an analysis of how the schedule affects patients would have
certainly been welcome- it understates the value of removing sources of friction in the work environment. In truth, it seems
that the problem solved by the model is most likely an important one, but the true value provided by the model is difficult to
quantify as the main benefit stems from fostering a work environment that is more friendly, as there is (hopefully) less reason
to quarrel with colleagues over resources.

1.2 Flu Supply Chain


Ekici et al. succeeds on two fronts. Interesting model and methodology; interesting problem and exploration of it. When one
thinks of pandemic preparation, food distribution is not the first thing that comes to mind, and while this particular does not
seem to have seen use in the ongoing pandemic, who is to say it will not see use the next time a bat-native virus decides it
wants to immigrate to a host of a more concerning (to us, at least) bipedal variety? Part of the charm of the paper in question
is its bravery to imagine and explore a novel way of addressing at what its time was a hypothetical. The value in this is it
plants the seeds for readers of the paper to consider pursuing similarly exploratory and creative “black swan” type questions
in research, and this alone makes the paper worth discussing.
Even while exploring such an intriguing concept, the authors do not slouch on making methodological contributions that
could potentially see applications outside of healthcare as well. There isn’t much more to say- the presentation of the facility
location model is both thorough and lucid, and the discussion of solution methods is similarly good.

1.3 Flu Vaccine


A reader with a sound understanding of stochastic programming but a weak prior knowledge of the vaccine design process
will find Özaltin et al. an engaging read on both counts. The exposition on the flu virus is short and sweet, and allows such
a reader to impress family members in medical school with their knowledge of antigenic shifts and drifts. The authors go
on to explain the factors that go into choosing strains for the flu vaccine, and the details of the production process that are
pertinent to the problem under consideration. Furthermore, these things are just interesting to know.
One of the shortcomings of the paper brought up in discussion is that the model seems somewhat simplistic as it does not
consider egg production at a later stage, which in the stochastic programming business is often referred to as recourse. While
it is not clear why the authors formulated their problem without recourse, the fruits that their labour brought to bare did not
go unappreciated. Their objective functions were prudently chosen with the risk averse in mind, and they provided the reader
with a couple of choices to boot. Their solution techniques were clever as well.

2 Papers for the Chopping Block


2.1 Inverse Markov Decision Processes
While the idea of using inverse Markov Decision Processes to solve a problem such as the inference of preferences is certainly
an interesting one, one could argue that this paper was never about solving the underlying medical problem, but rather
demonstrating how inverse MDPs could be used on a toy problem with a medical flavour. Although it doesn’t seem so at
first, the authors’ true intent is betrayed by their choice assumptions that preserve the technical aesthetic of the model while
sacrificing its practicality. Indeed, the authors assumed that patients are risk-neutral, when in fact they are human (and
therefore risk-averse), and assumed that decisions are based solely on patient physiology.
One could conceivably argue that papers in this course should sincerely aim to solve problems in healthcare as opposed to
using them primarily for illustrative purposes. And while this paper could be retained with the idea of showing how inverse
optimisation should not be used in healthcare, it might be worth considering a paper that uses an “inverse” approach to solve
a healthcare problem whose characteristics make it more agreeable to be modelled as such.

2.2 Deep Learning Review


The goal of papers in this course is to provide improvement to some decision-making process in a healthcare or medical
context by use of a quantitative model. By that metric, this paper is inappropriate for this course as it merely alludes to other
papers that have used tools to solve problems in a medical context. In the paper’s defense, it does provide a reasonably decent
introduction to various tools and techniques of deep learning, this time would perhaps be better spent on a paper that delves
into the thought process of solving a specific medical / healthcare problem using a model from operations research.
A more subtle danger with this paper is that it does not explicitly emphasise the role “black magic” plays in machine
learning models. What can often happen is when one is following the methodology of a paper, a slight misspecification of a
hyperparameter or a careless choice of random seed can lead to much less impressive results than presented in a paper. What
this means is that readers of this paper may not fully appreciate the “hackery” needed in utilising such models, or the dearth
of interpratibility that accompanies them.

2.3 Schizophrenia
As with the the other paper in the data mining module (which was also discussed in the previous section), this paper feels
inappropriate for the course as it does not seem to be concerned with decision making processes. Now perhaps one could make
the argument the scope of the course is broader than just decision-making processes, but it difficult to see how this paper by
Schnack et al. successfully accomplishes the goal it sets itself.
The paper purports to investigate a causal relationship between aging in the brain and schizophrenia, but like a bad
geologist studying sedimentary rocks, forgets about stratification! In particular, the authors acknowledge the existence of
several potential confounding factors, including but not limited to, psychosis, cannabis use, and the use of medication, but
fail to control for these factors when conducting their analysis. The authors acknowledge the difficulty in controlling for
medication use as all the patients in the study were on some form of antipsychotic medication, but do not account for other
potential similarities in the patients that may result in important caveats concerning the trends the authors identify. For
instance, no mention is made of the sex of the patients involved in the study. This could prove to be important because the
same condition can present itself differently in men and women, as is the case with heart attacks, so it is important understand
the exact demographic the study concerns.

3 Papers That Would be Interesting to Add


3.1 Yet Another Operating Theatre Paper
During discussions, a criticism that was applied to many papers was that of the presentation of the mathematical model.
Quite often it was the case that descriptions of parameter and variables were not consolidated in one spot, making it difficult
to hunt down the meaning of a variable xj or parameter ci . This in turn led to challenges in understanding what quantities a
constraint equation was relating together. A related criticism is the lack of explanation behind the purpose of a constraint.
The proposed paper by Lodi and Tubertini is instructive in that it provides a good example of how a mathematical model
should be presented. Indeed, each decision variable is given is described on its own line, all in one section, making them
very easy to look up. Similarly, all model parameters and their descriptions are consolidated into one table. Although the
objective function of the MIP look intimidating at first, as do some of the constraints, the authors are quick to explain the
ideas behind the mathematical formulations. Moreover, the paper provides insight into considerations that need to be made
when constructing such a model, as the authors worked closely with medical professionals.

3.2 A New Discussion: Queueing Theory


Two modelling tools that have their own week dedicated to them are stochastic programming and Markov decision processes.
To say the least, this is understandable; both are popular modelling choices in healthcare due to their natural and explicit
accounting of uncertainty. Oftentimes the uncertainty in healthcare has a particular structure to it- that of servers, clients,
arrival and service distributions. Perhaps that makes queueing theory worth its own week of discussion.

3.2.1 Nursing Levels


Doctors seem to receive the lion’s share of the glory. Ask a person on the street (from six feet away at the time of writing)
to name a TV show about doctors and they would likely be able to name several. Ask them to name TV shows primarily
concerned with nurses and they would invariably struggle. In spite of this nurses are an absolutely integral part of the
healthcare system, and constitute an important resource to pay attention to, as this paper does.
In addition to acting as a reminder of the important role nurses play, Green and Yankovic’s paper provides several important
lessons in the field of healthcare engineering. Firstly, it demonstrates how to clearly present a mathematical model that does
not consist of an objective function and constraints. The paper also conducts a detailed analysis that illustrates the kind of
insights that can be gleaned from queueing models. Finally, a problem brought up time and time again in discussion is that of
the model being wrong due to its lack of some seemingly crucial assumption. The authors of this paper endeavour to validate
their model against a simulation model that does not suffer from the same lack, which demonstrates how to figure out if your
wrong model is one of the useful ones.

3.2.2 Discharge Decisions


Queueing models are fundamentally probabilistic, and when one wants to apply probabilistic models to real life, one often
needs to endure the dreaded process of estimation. But what happens when quantities are too variable to get a good estimate
of? Chan et al. use an idea to circumvent this problem by using an alternative approach that uses quantities that are easier to
estimate. The authors also do not forsake the methodological for the empirical or vice versa, which is something the paper
discussed in Section 2.1 failed to do. They provide both theoretical assurances as well as empirical validation to support the
soundness of the model.
An interesting aspect of the model is what it doesn’t rely on: diagnostic information about the patients at the time of
discharge, demonstrating the utility of a simple model, and how one does not necessarily need to incorporate all the available
information in order to solve a useful and interesting problem well.

You might also like