A secret sauce of innovation through analytics is to develop the capability to extrapolate analytics methods that were leveraged in one domain to a completely different domain. Sometimes, that may be the answer to a persistent problem that you could not solve with analytics. We can use this article to highlight this approach: https://lnkd.in/gGih5y8p

If you zoom out, the key elements of the solution here are:

1. The scientists have data points on impacting variables (the conditions during spaceflight)

2. They have data on the current state of the dependent feature (which is the eye, the condition of which will be dependent on spaceflight conditions, so they have data on the current condition of the astronaut’s eyes)

3. They have data on how the impacting variables(the conditions during spaceflight) generally impact the feature (eyes)

If you think about it this way, removing all the jazz, it becomes a simple prediction problem. With the three data points available, you now want to build a model to predict for specific dependent feature. Obviously, you need specific skills for model building and training, but I am talking about problem formulation here.

So what is the most challenging part here? It is the three data points. Having access to reliable data. That is why I always say, if you have data, have designed your data architecture prudently, you are going to kill it in the AI age.

Returning to our article and being more specific, what did the scientists do?https://lnkd.in/gGih5y8p

Researchers at UC San Diego have developed an AI model that predicts the risk of Spaceflight Associated Neuro-ocular Syndrome (SANS) in astronauts before they leave Earth.  SANS refers to vision and eye-structure changes observed in astronauts during and after spaceflight (optic disk swelling, shape distortions, shifts in vision).

The AI is trained on optical coherence tomography (OCT) scans, high-resolution cross-section images of the eye and optic nerve collected before and during space missions.

Because astronaut data are limited, the team augmented the dataset by slicing the eye scans into many subimages, using techniques like data augmentation and transfer learning to generalize from small samples.  They also incorporated Earth-based analog data from head-down tilt bedrest studies, which simulate some of the fluid-shift effects of microgravity.