Thoughts on Machine Learning and Bulk Data Processing

Hi everyone!

It’s been a while since I’ve published on here, but I’ve had some thoughts I wanted to share on learning speed, machine learning, and microfluidic processing. [1] To give some background information, when I learn, I try to connect new information I’ve learned to old information that I already understand in an attempt to be able to recover the data when I recall it as some later point. However, because of the vast amount of information available to us, it’s difficult to cede the point that computers can do a better job than us at bulk data processing, which naturally leads to fears of job security for future doctors.

Figure 1: The issue of Nature I read (Credits)

Interestingly, in an article that I read in the October 2017 issue of Nature, a highlight was released on machine learning and its relation to the scientific field. They stated that machine learning is not growing so quickly that it warrants the fear that skilled physicians will be out of jobs in the future. In particular, I was most interested by the example where they looked at machine learning applied to making tumor cell diagnoses within the medical field. Finally, the paper I read on microfluidic processing discussed the inefficiencies of bulk RNA -> cDNA conversion and the solution to the problem in single cell expression profiling, a different version of bulk data processing. [2]

The aim of this article is to:

1. Warm-up the portion of my brain responsible for writing
2. Talk about learning and the inevitable incorporation of machine learning
3. Approach data processing in a couple different ways

I think that my one greatest fear is that in comparison to the rest of the brilliant, Berkeley population, that I am learning too slowly and am incapable of contributing worthwhile information to the scientific field. It’s a prototypical case of imposter syndrome, but knowing the name of the problem doesn’t really make me feel any better.

In my mind, learning progresses linearly. That is, for every paper I read, I add a little bit more information to the collection of knowledge that I’ve already amassed. In my worldview, this idea could be applied to other people as well. For some people, the rates at which they learn are faster than mine; they connect the dots easily and can access disparate pieces of information without too much trouble. For others, the linear rates at which they learn are slower and it’s difficult for them to put the puzzle pieces together. This degree of “natural talent” applies to everything and may be a cause of alarm. The idea has been taught as a “fixed mindset” within education circles. However, although someone potentially can learn faster than you do, you can still make up that difference in knowledge given enough time. While it isn’t exactly a fun prospect, that you have to put in the work while they can choose to relax, it gives you a point to work towards, a plan of action that is better than not trying at all.

Figure 2: A linear learning curve (Credits)

Interestingly, Atul Gawande gets at this point in one of his pieces from 2004. [3] Labeled “The Bell Curve”, he discusses the implications that in any skilled activity: playing baseball, reading books, throwing footballs, and diagnosing patients, there is this natural bell curve that exists. It is a fact that not all doctors can be the best and it is even scarier to think that physicians may try to hide this fact. However, reassuringly, I also saw a recurring theme: efforts towards improvement can close the gap between those at the cutting-edge of the bell-curve and those at the opposite end. His article discusses the measures taken by those clinics who had low median survival rates with their patients who had cystic fibrosis and follows their story of improvement.

Figure 3: A Kaplan Meier survival growth curve (Credits)

Initially, I searched for Atul Gawande’s early writings because I was curious about the man behind the persona. Right now, he is one of the most prominent writers medicine presented to the public eye, but more importantly, an enigma in my mind. Moreover, because I knew he wrote in the New Yorker, a magazine that publishes almost every week, I could access some of his older archived articles for a hint to his murky past. To do so, you google: “New Yorker Atul Gawande” and scroll to the 6th page of his article page, where you’ll find some articles without pictures. These articles are some of his first, written all the way back to the 1990s and what you’ll find are less polished pieces of writing – excellent to be sure, but they lack the voice and character that the public figure now carries in his novel.

Figure 4: The first article to pop-up with a picture. Jeff also read this article at UCLA! (Credits)

Nevertheless, they give insight into the man before his ascent to fame. He was a learner just like you and me, polishing his writing, attempting to improve his ability to convey information. In many ways, I appreciate the flaws and cracks in his prose, it conveys his humanity in a way that just speaking about it never could. But, Gawande was also worried about the same problem that I had. In his article: “No Mistake”, he highlighted the idea that even since 1998, physicians had recognized that bulk data processing and specialization is critical to improving patients’ outcomes. [4] In one of the examples in the article, he compares a blind algorithmic approach versus human judgment in Pap smear diagnosis for cervical cancer. Surprisingly, what he found was that the blind algorithmic approach trumped human judgment. Consequently, it’s a very real fear that we’ll one day be replaced with computers to make the sensitive life-death decisions in the clinic.

Figure 5: The first article I could find archived on the New Yorker by Gawande (Credits)

Soon, a writer on my new blog, Sonic Hedgehogs, will be looking at such an application to seizure predictions. It is common knowledge that if computers can help us improve patients’ quality of life, that it’ll be implemented provided that the algorithm stands up to scrutiny. Furthermore, there is no good reason to fear the advance of machine learning from a patient’s perspective if it can only improve diagnoses. In the Nature article that I referenced in the introduction, machine learning could still only be implemented in specific procedures in the medical field. However, I feel that even if machine learning’s integration into the medical field is slow, the recurring cases of success for machine learning replacing physician intuition is telling. But, I think it is really awesome that we have the ability to improve patients’ outcomes – even if it’s replacing a human with a machine!

Figure 6: An example of a microfluidics processor (Credits)

Lastly, the paper I’ve just read is about microfluidic processing of single, embryonic human stem cells. [1] For me, it strongly evokes the idea that we still need human imagination to drive scientific progress, independent of the growth of machines. The paper highlights the problematic assumption that we can guess at individual cell mRNA expression levels given the expression profile of a population of cells. The authors emphasized that instead of running reverse transcription reactions in multiple tubes, the way that you and I would do it in a lab setting, that running it on a chip (which is normally like a glass microscope slide) would improve efficiency of the reaction, resulting in more DNA from less mRNA. While this idea isn’t exactly new, I could never imagine a machine replacing the design of these complex experiments. I for one, welcome the integration of medicine and machine learning and look forward to the revolution to come.

In conclusion, I’ve talked quite a bit about learning, Atul Gawande, and microfluidic processing, but what I really wanted to drive home is that a passion for learning will take you farther than any natural talent. All the brains behind these projects are no different from you and me. One day, I look forward to meeting you at the cutting-edge of the bell curve.

Warm regards,
Joseph