Bio-Artificial Intelligence: A lesson in Why AI Algorithms must “Know” the Field

Artificial Intelligence (AI) is all the rage. It seems that every time I surf the web, I see a new AI product, ranging from consumer electronics, to cars, to social media apps, to high-end analytical software. For example, check out these other articles on AI in Internet Marketing, the Legal Profession, and several other application areas. There are even AI products advertised in the airports. Forget what the pundits say about AI being a big part of our future. Just look around to see that the future is here now.

AI Approaches Must Adapt to the Problem at Hand

With the increasing prevalence of AI, we see successes and failures in the use of the technology. I have seen many questions on Quora along the lines of “What Artificial Intelligence algorithm should I use for mining the web?” Clearly, this is an unanswerable question; without knowing what the goal is, the correct approach cannot be defined. This is an extreme example, but to me, it is indicative of the same problems we have seen historically with other new technologies. There is a tendency to rush in with the new hammer in the toolkit and bang on everything like it is a nail.

The Need for Biomedical-specific AI

The reality is that AI methods must be adapted to the specific problem at hand. When we started investigating the use of AI in our own work to improve discovery of relevant biomedical literature, we quickly learned that unless we tackled the problem from the perspective of our target audience of biomedical professionals (which we are too) and adapted the AI to solve the critical problems in the field, we weren’t going to have a practical solution. You cannot blindly apply, for example, a canned back-propagation neural network algorithm to biomedical text and expect to achieve a solution to your goal.

Here are just a few of the problems we encountered applying standard AI algorithms to biomedical text:

1. Multiple meanings for the same term (polysemy) in the biomedical literature is extreme, with acronyms and abbreviations sometimes having over a hundred different meanings, and biomedical entities such as genes and aliases of many chemicals being the same as regular words. To avoid an unfathomable amount of training, additional methods had to be created.
2. Traditional triplets derived from natural language processing were insufficient to address the complexity of biomedical text, where pathways and system-level perspectives are implied throughout. This means rethinking how to discover the depth of the author’s intent.
3. AI-powered visual analytics are essential to move away from lists of results which cannot be assimilated, but visualizations that simply represent an arbitrary view aren’t necessarily helpful for the biomedical literature. (Here is another commentary on one aspect of that problem.) Visual analytics that addresses real questions are critical, and these depend on a deep understanding of the biomedical problems.

These are just examples of the many issues where we found the field-specific problems could not be adequately solved with off-the-shelf AI algorithms, and we had to develop our own unique solutions.

Summary

Many years ago in my Physics training I learned that many problems require you to think beyond the implements at hand and to build the tools you need. Whatever your application, as you look to AI to address critical problems, consider the goals carefully, think like your users, and adapt your AI approaches accordingly.