Step 1: add a bunch of diverse experts to your team
Step 2: take more precise measurements
Step 3: take measurements over a wider range, and with more modalities. For example, if you are currently measuring infrared, add ultrasound, x-ray, and magnetic resonance spectroscopy
Step 4: have world experts establish ground-truth information and classify the samples
Step 5: analyze the data with a really thorough program that tries a lot of different hypothesis and techniques
Step 6: Repeat Step 1 through Step 5 a number of times until you finally find the rather unexpected combination of easy-to-take measurements that gives you 90%+ accuracy
Step 7: Publish a paper on all of this, to international acclaim, and the amazement of world-class experts who had only been rated as 84% to 87% accurate themselves.
Step 8: Spend the next few years figuring out why that combination works, as it involves previously unknown physical or biological processes
... Because that is what is takes to get 90%+ accuracy on real-world data.
Our experience was that there was a huge gap between being able to get 80%-ish accuracy and being able to get 90%-ish accuracy in real-world classification situations.
