Trial and Error No More: Building Organs with AI

One thing we need to face is at this point in time, there are far too many people facing a sad end to their lives because they cannot get the organs that they need.

Because, right now, there is no supply chain, system, or approach that is able to feasibly scale up to the need for organs and various tissue.

Introduction

Given that patients cannot get the right tissue at the right time, a well-established and promising solution is making more organs.

You can see a full article about this here, but to summarize the most promising way to meet the demand is to create supply in the form of artificial tissue that can ideally, be indistinguishable from a natural counterpart.

Of course, this technology is still in its infancy, with various challenges that need to be addressed and worked on. Of those various challenges comes that of quality…

Pretend you make a brand-new leap in the 3D Bioprinting field and come up with a new tissue, or a new scaffold, or a new method to bioprint. When it comes to a field as broad as this one, we need a way to compare all these metrics and see what product gives us the most quality. Because comparing the “quality” of two different bioprinting solutions can help us find important attributes to consider for the next iteration.

However, at the end of the day, comparing things which are qualitative can be hard. You need to find the right patterns which are difficult and time consuming, especially when there isn’t an objective measure to go off.

Our Good Friend AI

What is something that's really good at understanding relationships we can’t see and can do a good job of classifying things?

Artificial Intelligence! AI once trained properly are some of the strongest ways to classify something in a consistent fashion. This means it is possible to create an AI that is able to take attributes of bioprinted tissue to assess its viability and quality.

Building an AI Classification Model

When looking at ways to approach this task, a team of researchers decided there were two main ways to train the model. Those being, using random forest regression or random forest classification.

The idea is that given a set of labeled data points (training data), the model after picking a random number of data points from this set can make a decision tree with an output. This process can be done many times resulting in many decision trees, whose outputs are averaged out to finalize the model. When a new data point is introduced the output of the model is the mean of outputs of each of decision trees.

Decision trees are highly accurate because we can think of them as many little models working in tandem.

Dataset

For this project the dataset used, from the scientific paper “Machine Learning Guided 3D Printing of Tissue Engineering Scaffolds.” The model in this article is drawn from the code and discoveries they made with their dataset.

The dataset had multiple permutations of different bioprinting standards which also had a label for success. Characteristics of material proportions, spacing of fibres, speed of the printing, printing pressure and layers per scaffold were all combined in different amounts and tested for success. This also means that at the end, the model given these various characteristics can do the inverse, and project the success of an unknown permutation.

Training

Now that there is a raw data set, the next step is to give these categories labels as to how effective they are as a bioprinting option. The researchers who compiled the data set also determined that a label for the machine’s precisioncould be determined by looking at the spacing of the fibres bieng printed and the accuracy could be determined with the width of the fibre.

The code above shows one approach to doing this: Specifically it is:

• Taking information from the dataset
• Preprocessing it by removing the unnecessary or emptyparts and then opening it
• Creating a new data frame and copying over the essential metrics
• Determining the accuracy and precision
• And returning two sets; “dataX” with the metrics and “dataY” with the corresponding labels

This second part of the code is processing the data furthermore. Specifically, we are training the model on data that is missing some parts to it. This is to ensure the model isn’t tested on data it has seen during training. The values that will be missing are that of the fiber’s spacing and that of the numbers of scaffold layers.

Results

With this information the model learns to classify how effective a permutation of printing methods is so that when, new information is given, which has not been tested the model can predict. In fact the model gets >95% accuracy when tested with data that has not been used for training.

Implications

The fact is, 3D-Bioprinting faces a specific problem in the area of growth.

There is just too much trial and error going on for effective and iterative improvements to be possible.

The model replicated here showas just how much pssibility there is in using machine learning tools to validate what we see in the real world. This simple model takes just couple factors and provides a pass or fail. However, a far more sophisiticated model supercharged by much more data can likely do much more.

A model that has data points regarding the preprocessing, creation, post processing and deployment stages can be far more effective at testing the liklihood that a certain kind of 3D-Bioprininting approach can work.

This means that researchers can save huge amounts of time and money that they would otherwise need to dedicate to types of 3D printing that might not be as effective. Meanwhile models even at a lower depth unlock the ability to run through millions of combinations, streamlining the trial and error approach. Even disregarding all of this, sophisticated validators can show use patterns present between successful 3D-Bioprinting methods which then can be validated and applied.

Conclusion

In this field, what AI does that humans cannot, is find patterns in mountains of data. Especially in bioprinting these patterns (if they exist) can be gems of knowledge that can change approaches and help the field grow out of its infancy. This can be one of the tools that changes the feild that can change medicine, and that gives me great optimism. What was done here, is a simply model, a proof concept to show what is possible given more resources, time and strategy.

Hey! I am Abhinav a high schooler who is passionate about the environment, people and technology. If you enjoyed this article, give it a clap! You can find me on LinkedIn, I would love to connect! Stay creative!