42 Cannot Be The Answer
- AI in health care has enormous potential to improve the quality of living and reduce overall health costs
- AI its strength is its weakness: very small and sometimes undetectable ‘noise’ to the human eye or composition is decisive in classifying an image
- It is quite normal in human interaction to explain why and/or why not.
- If AI is to be integrated into health care, the AI must not just tell what it ‘thinks’ is the answer, the AI must also tell how it came to its ‘conclusion; the “AI-lineage” of the algorithm. Clearly pointing out the area that is different. Preferably backed up with known cases that have similar deviations, and cases with the absence of the deviations.
“The Answer to the Great Question... Of Life, the Universe and Everything... Is... Forty-two,' said Deep Thought, with infinite majesty and calm.” ― Douglas Adams, The Hitchhiker's Guide to the Galaxy
A study using Google’s Deep Thought Mind AI could identify breast cancers with a similar degree of accuracy to expert radiologists while reducing the number of both false positives and negatives; not by much but still. Great developments that will improve the quality of living while reducing the overall health costs.
But every positive has its own drawbacks and challenges. Within the field of imaging classifying studies, the term ‘adversarial attacks / patches / images’ has emerged over the years.
With the right kind of noise to an image and with 99% confidence, a panda is identified as a gibbon, or just about any image as an ostrich. Or, with an adversarial patch by put ‘smartly’ a physical sticker on an object and a stop sign becomes a 45mph sign, and a toaster a banana. But even untouched natural images can be wrongly classified and thus a spider has been labelled with 99% confidence as a manhole, and bird as a jeep.
Of course, progress is being made to reduce these wrong classifications. While the breast cancer study looks as if we are about to beat the best human experts, other studies show that natural images sometimes are classified with 99% confidence compared to humans with an accuracy of less 5%. Enough Black Swans to seriously hinder AI in health care.
As it stands now, the strength of AI is its weakness: very small and sometimes undetectable ‘noise’ to the human eye or composition is decisive in classifying an image.
Last year my oldest needed an X-Ray and it was for me as an untrained easy to spot the fracture in his thumb (thumb metacarpal) because I knew where it hurt and I know how a fracture looks like on an X-Ray. My oldest knew where to look, but couldn’t see the fracture because he didn’t know how a fracture would look like. After some explanation he could recognize the fracture and that all his other bones were okay.In a recent X-Ray of my jawbone, though I knew where it hurt, I could only spot the inflammation after being pointed out and double-checking that this ‘colouration’ indeed deviated from the rest of in the X-Ray visible jawbone.It is quite normal in human interaction to explain why and/or why not. The answer is almost secondary. Among experts, be it in health or IT or any other field, the answer can even shift or be completely different as a result of the interaction.
If AI is to be integrated into health care, the AI must not just tell what it ‘thinks’ is the answer, the AI must also tell how it came to its ‘conclusion; the “AI-lineage” of the algorithm. Clearly pointing out the area that is different. Preferably backed up with known cases that have similar deviations, and cases with the absence of the deviations.
