Author Spotlight
Making an Impact Sometimes Means Sacrificing Depth for Breadth
Adrienne Kline reflects on a multidisciplinary career that bridges medicine, machine learning, and engineering
In the Author Spotlight series, TDS Editors chat with members of our community about their career path in data science, their writing, and their sources of inspiration. Today, we’re thrilled to share our conversation with Adrienne Kline.
Adrienne Kline is a postdoctoral fellow in the Department of Preventative Medicine at Northwestern University. She completed her Ph.D. in biomedical engineering (with imaging specialization in machine learning), an M.D. in medicine, preceded by a B.Sc. in electrical engineering.
Her interests lie at the intersection of medicine and engineering, specifically in leveraging algorithmic decision-making support for translational applications to medicine. Her work has led to the development of novel methods for handling missing data, innovative metrics for evaluating the reliability of machine learning predictions, and information fusion of multimodal data streams. With an emphasis on structural data, computer vision, reinforcement learning and generative algorithms, she hopes to change the efficiency and reliability with which medicine is practiced.
What kinds of topics and projects are you most drawn to to these days?
My work has always consisted of a mix of novel applications and methodology. Applications, because I often want to leverage an existing tool for a task; but when the tool I want doesn’t exist in the field, I work to create it. Currently, I am working on novel imputation and machine learning capability tools, both of which are methods- and math-heavy.
Two other projects I am working on in the field of cardiology are in generative models and reinforcement learning. These are both application-based and serve to improve the reliability of decision-making in heart-failure patients. Additionally, I seek out novel projects combining hardware and software when I see problems that call for them. I’m always returning to theory, solving a meaningful problem.
As someone with a multidisciplinary background in engineering, machine learning, and medicine, how did you decide which career path to pursue?
My machine-learning career began in the latter portions of my undergraduate degree in engineering. My capstone project was on brain-computer interfaces (BCIs) — working in signal/image processing to control robotic hands using electroencephalography (EEG).
Following undergrad, I was accepted to PhD and MD programs simultaneously, and opted to defer medical school while I completed graduate work. During my PhD, I continued implementing machine learning into BCI robots I had designed using fMRI (functional magnetic resonance imaging) and EEG for spinal-cord injury patients.
Once I started medical school with an engineering lens on, I began to appreciate the nuance and difficulty of medical decision-making. Medicine is a demanding profession, but as I learned how to think like a clinician, I started noting all the many ways in which I would like to improve upon the current paradigms. Although avant-garde to my anxious classmates, I forewent a residency match as I saw myself as a bridge between worlds, rather than as a medical practitioner.
I enjoyed caring for patients, but my ideal role, playing to my chimeric strengths, became one where I could provide care to patients I’d never meet, sacrificing depth for breadth. I’ve always said I’m not after a job, I’m after a role—one in which I can work ‘at the bench,’ manage a team of engineers and scientists, collaborate with our clinician partners, and speak both languages with credibility.
Ultimately, I want to develop and translate technology (algorithms, devices, information) that improves the reliability, decreases the cost, and improves the efficiency of patient care.
Data practitioners today have many career choices. What kinds of questions would you recommend they ask themselves as they set out on their journey?
There is an abundance of opportunities in this booming and exciting field.
I think it’s easy to be sold on buzzwords with the overuse of the phrase “artificial intelligence.” Ultimately, career decisions are a value-based judgement contingent on the individual. If you’re motivated by salary, you are better off in industry. If you want to march to the beat of your own drum, maybe academia or a startup is for you.
That said, it used to be academia vs. industry, but many companies have robust, purist R&D departments that publish white papers, and many academics patent their research and birth companies from work in their lab. So when making any decision, like with any good algorithm, find the most important facets to you (salary, upward mobility, autonomy, geographic mobility, etc.) on which to base a decision. Then score these (and differentially weight them if you wish) against how well each of your prospective options fulfill those intrinsic desires; this is how I make my life and career decisions.
Adaptability and malleability are an absolute necessity for staying current regardless of area, and to grow skills, people need development room in their jobs. So if you’re considering industry more heavily, use caution and ensure some development room is built into the job description; it tends to be automatically built into academic learning centers.
Regardless of career stage, I encourage people never to forego the fundamentals of math, statistics, computer science, and engineering. These fundamentals are the crux on which novelty is built. New architectures and algorithms pop up daily, and if you focus on the leaves you lose sight of the tree. Always ask yourself “so what?” to ground your work and as a reminder to answer the big questions.
Why did you decide to write about data and machine learning for a wider audience?
My very first article on TDS was about a Python package I had created — psmpy. There is a formal white paper available now that I’ve since cited in the Medium article. The white paper submission occurred early in 2022 and wasn’t presented until July at an IEEE conference. But because the paper required a working package, I felt it presented a unique opportunity to make the data science and statistics community aware of its existence rather than via a ReadMe that would sit in an unseen repository.
Oddly enough, it became a popular article, so I began writing others inspired by questions our graduate students ask me in the lab. I have always enjoyed teaching and place a heavy emphasis on mentoring. By creating a repository of these articles, I have resources I can refer students to. These articles double as a resource for teaching and mitigate some repetition in the questions I am asked.
In recent months you’ve been publishing the excellent Statistics Bootcamp series. What inspired that idea, and what do you hope your readers take away from it?
I have had the privilege to teach students at both the undergraduate and graduate level. Teaching has always been a rewarding experience, and my students have always been my inspiration. The bootcamp is built on an introductory statistics graduate course I taught previously.
My hope for the bootcamp is that it can be an easily accessible and affordable repository of notes for those who have attended graduate school (and may want a refresh), and for those who want to acquire the knowledge from a graduate-level course in statistics but wouldn’t have had the opportunity to do so otherwise. By reading it, people become better generators and consumers of information.
Whether in your own specific field or in the broader ecosystem, what kinds of progress do you hope to see in the near to mid-term future?
Over the next few years, I hope to see more deployment of machine learning into translational applications. The caveat I’d add to that is I want to work and build in an environment where the reliability or uncertainty of the predictions are as paramount as the ability to get a prediction in the first place.
Knowing the limitations of algorithms, i.e. who it works well for and who it doesn’t, is incredibly important. It’s a tradeoff between the algorithm and expert decision makers, and it should occur on a case-by-case basis.
Lastly, in the health sector I want to see multimodal data drive more patient decision-making. In medicine, many decisions are termed clinical diagnoses; this means that it is not a single lab or imaging study that clinches the diagnosis, but rather a constellation of signs, symptoms, labs, imaging, etc. This is where ML has the ability to really shine: it can weight disparate data types into a unified diagnosis, since clinicians (and humans more broadly) don’t excel at performing matrix algebra when they examine a patient.
To learn more about Adrienne’s work and stay up-to-date with her latest articles, follow her here on Medium, on Twitter, and on Google Scholar. For a taste of Adrienne’s TDS articles, here are a few standouts:
- Explore all current posts in the Statistics Bootcamp series.
- Implementation and Limitations of Imputation Methods (May 2022, 9 minutes)
- Creating a Dashboard App with Dash/Heroku: A Streamlined Process (June 2022, 5 minutes)
- Parametric vs. Non-parametric Tests, and When to Use Them (April 2022, 4 minutes)
Feeling inspired to share some of your own writing with a wide audience? We’d love to hear from you.
This Q&A was lightly edited for length and clarity.
