Standard machine-interpretable evidence specification is NEEDED for reliable prescriptive analytics

[Brian Alper]

I want to make this assertion boldly and clearly available for:

1. Communication to HHS in the next few weeks while RFI is open
2. A peer-reviewed publication to scale the dissemination of this message
3. Providing a basis for AI development for prescriptive analytics
4. Providing a basis for EBMonFHIR as a key foundation

 

Let me know if you want to join a co-authoring group, especially with #2 in mind.   Perhaps we start with the HEvKA StatisticsOnFHIR Working Group meeting (Mondays at 2 pm Eastern) as the focused meeting time for developing this message.  Being a co-author for this one means an expectation of active participation in the next 2 weeks to shape a substantial manuscript, perhaps drafting text or finding strong references to back it up. (As this email may be seen by hundreds of people, I want to be clear I am looking for active co-authors, not a long list of ‘supporters’)

 

A standard machine-interpretable way to define the evidence used for prescriptive analytics should improve transparency, accuracy, precision, explainability, reduced bias, validity, trustability, efficiency, and reproducibility.

 

An outline for making the argument could be:

 

1. AI is now expected to be used in nearly every cognitive function.
2. AI has established acceptable performance for descriptive analytics, diagnostic analytics, and predictive analytics in many contexts, but the use of AI for prescriptive analytics in healthcare triggers alarms for ethical, legal, and social concerns regarding trustability for both internal and external validity.
3. The evidentiary basis for diagnostic analytics and predictive analytics is pattern matching, and this is why AI has scaled to perform better than human cognition for these areas.
4. The evidentiary basis for prescriptive analytics is critical appraisal of findings from controlled observations with deep understanding of the methods for obtaining these findings.
5. Using pattern matching to approximate the evidentiary basis for prescriptive analytics is exponentailly problematic because:
1. It introduces confounding bias by converting large numbers of chance associations (findable with big data analytics) into misleading findings from ‘uncontrolled’ observations.
2. It exacerbates the extensive problem of reporting bias in conventional scientific publications, by using the pattern of language in study reports instead of the data of study findings.
6. For AI to provide prescriptive analytics with the smallest possible degree of bias, the AI needs a data source with a precise, unambiguous expression of the findings from controlled observations.
7. For AI to provide prescriptive analytics with recognition of the potential for bias, the AI needs a data source with a precise, unambiguous expression of the methods of obtaining these findings.
8. A precise, unambiguous form of expression is not truly precise or unambiguous unless there is global agreement on its interpretation – hence the need for a standard.
9. A standard machine-interpretable way to define the evidence used for prescriptive analytics should improve transparency, accuracy, precision, explainability, reduced bias, validity, trustability, efficiency, and reproducibility.
10. A standard machine-interpretable way to define the evidence used for prescriptive analytics EXISTS, in early-maturity form as standards go, with the EBMonFHIR specification.

 

Brian S. Alper, MD, MSPH, FAAFP, FAMIA, FGIN

CEO, Computable Publishing LLC

bal...@computablepublishing.com

 

 

Making Science Machine-Interpretable
http://computablepublishing.com 

 

Lead, HEALTH EVIDENCE KNOWLEDGE ACCELERATOR

President, Scientific Knowledge Accelerator Foundation

 

Read about The Mission Change of a Lifetime

"It only takes a pebble to start an avalanche."

Schedule a meeting with me.

 

The views and opinions included in this email belong to their author and do not necessarily mirror the views and opinions of the company. Our employees are obliged not to make any defamatory clauses, infringe, or authorize infringement of any legal right. Therefore, the company will not take any liability for such statements included in emails. In case of any damages or other liabilities arising, employees are fully responsible for the content of their emails.