使用半合成数据构建医疗保健领域的人工智能模型.pdf

1、Building AI ModelsIn Health CareUsing Semi-Synthetic DataWaiting?Go to:https:/ and play with it!This talk is not legal adviceIf you are training models in a regulated domain talk to your lawyerSeriouslyEspecially in any places where youre working with vendorsContent warnings&biasesContent warnings:M

2、ention of crash,broken bones,american healthcare,and gender affirming careIf its not your vibe no judgement prioritize yourself please.Biases:Doing this full-time so strong opinionsex-Netflix,Amazon,Apple,Google,etc.Spark&Ray nerd:DThe journey aheadWhy Im working on this and why you should careFine

3、tuning AI models in privacy sensitive domainsDifferent approaches to preserving privacyPutting them all together to fine tune a modelFind yourself zoning out?Talk with doughnut https:/ https:/ out the github https:/ some personal motivationRule based expert systems dont scale&get ignoredGeneral purp

4、ose generative dont have restricted dataReduce costMore control over dataRetrieval augmentation only goes so farReasoning(ish)Why fine tune generative AI?Initial dataset are difficult to come by(bootstrap)Information leakage is even more critical to be avoidedKeeping fresh is hardProblems fine tunin

5、g in Healthcare(and related)Bootstrapping:How to get started?Request them from someone who has themThats insurance companies&doctors officesGoogle/Reddit search for denialsThen write some appealsFind some other data we can produce appeals/denials fromOooh!Insurance commissionersIndependent Medical R

6、eview BoardsCalHHS Open Data https:/wpso.dmhc.ca.gov/imr/Texas:Has the data at the state level but couldnt find a quick open version(also california was enough to start)NY:Huzzah!WA:Also huzzahExample CA IMR RecordFindings:The physician reviewer found that Nature of Statutory Criteria/Case Summary:T