Dear Human,
Do you know that the fitness tracker you are wearing, or the smartphone app monitoring your daily physical activity, or the electronic health record (EHR) from your clinical visits are a few of the major sources of my training? What is this training for? Who is my employer? The answer to the latter question is the healthcare industry. And, I am training to fulfill my job responsibilities, which are plenty: correctly diagnose diseases, predict novel disease patterns, identify the invisible population who are at risk, predict the success of clinical drug trials, so on and so forth. But, aren’t clinicians and researchers worldwide already doing this? Yes! But, the manual labor, time and the costs are too high as compared to the results, which often have errors or missing values. Hence, they have put me to task. To meet these demands, my training needs to be robust. And boy, it is (or at least the opportunity is)! I am currently floating in a vast ocean of information from biomedical research, patient records, and wearable devices. The more information I get to train over, the better will be my efficiency. All this sounds hunky-dory till I get to the glaring stumbling blocks in my learning/training.
First, this vast ocean of information is extremely noisy and miscellaneous. Often, I don’t know what to do with a piece of information that I have picked up because it doesn’t have a label or notes attached to it. In other cases, I find conflicting labels on the same piece of information. This, typically, happens when two different parties have labeled the concerned piece of data after referring to two different medical ontologies. So, the problem is more with inconsistencies/conflicts amongst the ontologies and not so much with the labelling itself. But, I am at the receiving end of this, and indirectly is the healthcare industry. Second, despite its size, this ocean is not representative of the entire world population as we have a huge chunk of people without access to primary health care. So, in a word, the ocean that I am floating in is ‘biased’. Third, disease progression over time is often unsynchronized or discontinuous among patients. A major limitation in me is that I lack the ability to handle time-course analysis. Fourth, the causes and patterns of progression for most of the diseases are not completely known. Adding to this is the limited number of patients recorded for each disease type. Thus, to classify these medical conditions into domains for my improved training appears difficult right now. Finally, although I have been created by man, for man, interpreting me is like reading a black box – a complex system whose internal mechanisms are not clearly understood. Interpreting my workings is crucial and has far-reaching consequences as clinicians these days are increasingly relying on data-driven solutions for decision-making and patient monitoring.
With the stumbling blocks in place, let me now give you a quick overview of what you can do to make me work better for you. First, since patient numbers for capturing information are limited, collect as many features as possible for each available candidate. They could be electronic health records, wearable devices, information from social media, environments, surveys, online communities, genome profiles, and beyond. Try working on an approach that could let you throw all these data together and integrate them to help me generate actionable insights. Protecting me, and sensitive patient data would be another priority that you’d need to work on. Add more expert knowledge that is curated from medical journals, research papers, and professionals into the existing information ocean. This would help me train in the right direction. Moreover, considering that time is an important factor in every health care-related problem, another dimension to work on is to make me time-sensitive. And finally, of course, make me more interpretable for the clinicians. The more they will understand me, and my results, the finer I can work for them.
And while I end this letter, I realize that I forgot to introduce myself. Anyway, the subject is way deeper than my name, so in a way, it matters less (?). Some of you might already know me. For others who don’t, I am Deep Learning (DL), a subset of machine learning in AI.
My way of functioning and decision-making mimics your brain. I am currently used in a lot of other application domains, but today I wanted to have a word with you on my role in health care. Because if you survive, I survive. And, vice-versa. Get the drift?
Thanks for your time.
Yours faithfully,
DL
Reference: Deep learning for healthcare: review, opportunities and challenges
Photo courtesy: Nvidia blog
Cover image courtesy: http://startupheretoronto.com/partners/cvca/vc-investing-convergence-ai-healthcare/
Author: Saikata Sengupta
Saikata Sengupta is currently pursuing her Ph.D. from Department of Neurology at Friedrich Schiller University, Germany. You can follow her on Linkedin or Twitter
Editor: Arunima Singh, PhD
Arunima obtained her PhD in Computational chemistry from the University of Georgia, USA, and is currently a postdoctoral researcher at New York University. She enjoys traveling, reading, and the process of mastering a new cuisine. Her motivation to move to New York was to be a part of this rich scientific, cultural, and social hub.
Second editor: Manoja Eswara, PhD
Manoja Eswara did her Ph. D. from University of Guelph, Canada and is currently doing her postdoctoral fellowship in Cancer Epigenetics at Lunenfeld Tanenbaum Research Institute, Toronto, Canada.
The contents of Club SciWri are the copyright of PhD Career Support Group for STEM PhDs (A US Non-Profit 501(c)3, PhDCSG is an initiative of the alumni of the Indian Institute of Science, Bangalore. The primary aim of this group is to build a NETWORK among scientists, engineers and entrepreneurs).
This work by Club SciWri is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.
