4 ms·

I find it odd that the article seems to downplay the potential benefits machine learning provides by emphasizing a lack of explanatory power from classification

by cs2818 10y ago

I find it odd that the article seems to downplay the potential benefits machine learning provides by emphasizing a lack of explanatory power from classification algorithms. At the very least it would seem that a machine-based classifier provides human physicians and researchers with more examples to base their inquiries on (possibly even illuminating some features they may have previously missed as important portions of theoretical models).

randcraw 10y ago
But more opinions (from AIs) that differ from the advice of the 'responsible health professional' will surely add confusion to the doctor's life, and in the US medical space, legal liability. In general, doctors are understandably reluctant to invite more oversight, especially if it doesn't clearly add value and if it's not independently and certifiably trustworthy. Past AI apps, like 1980s expert-systems, generally relied on brittle binary criteria that were hard to match with certainty. Too often they produced results that were either obvious or implausible, but at least they could explain themselves. They were also poor at matching against fuzzy clues from patients (and doctors) who are notoriously inconsistent and nonquantitative at describing symptoms. No doubt a greater emphasis on quantitation lies at the heart of today's AI systems. But if the classifications and recommendations of tomorrow's AIs lack explicability, there's no way in hell they'll be trusted or given authority by risk-averse practitioners. A middle ground is needed, where the 'advice' from the AI is grounded in clear statistically significant bases and adds value to the process, rather than competing with humans. In some spaces like suggesting cancer therapies, that are more likely to succeed using quanta, I think AI will be adopted and appreciated first. Primary care medicine will probably see it last, though it probably is already employed invisibly behind the scenes by insurers for validation and quality control (like prescription drug contraindication).

by cs2818 10y ago

randcraw 10y ago

But more opinions (from AIs) that differ from the advice of the 'responsible health professional' will surely add confusion to the doctor's life, and in the US medical space, legal liability. In general, doctors are understandably reluctant to invite more oversight, especially if it doesn't clearly add value and if it's not independently and certifiably trustworthy. Past AI apps, like 1980s expert-systems, generally relied on brittle binary criteria that were hard to match with certainty. Too often they produced results that were either obvious or implausible, but at least they could explain themselves. They were also poor at matching against fuzzy clues from patients (and doctors) who are notoriously inconsistent and nonquantitative at describing symptoms. No doubt a greater emphasis on quantitation lies at the heart of today's AI systems. But if the classifications and recommendations of tomorrow's AIs lack explicability, there's no way in hell they'll be trusted or given authority by risk-averse practitioners. A middle ground is needed, where the 'advice' from the AI is grounded in clear statistically significant bases and adds value to the process, rather than competing with humans. In some spaces like suggesting cancer therapies, that are more likely to succeed using quanta, I think AI will be adopted and appreciated first. Primary care medicine will probably see it last, though it probably is already employed invisibly behind the scenes by insurers for validation and quality control (like prescription drug contraindication).