Study groups to Received actions labeled as reciprocal determine - PowerPoint PPT Presentation

─ Received their first-ever action from a board Group 1 ─ Represent 45% of disciplined physicians in 2017 ─ Have 1.3 board orders and 2.5 actions per physician Study groups to ─ Received actions labeled as reciprocal determine if board Group 2 ─ Represent 13% of disciplined physicians in 2017 ─ Have 4.6 board orders and 7.7 actions per physician order narratives provide additional ─ Received multiple actions all from the same board insight beyond the Group 3 ─ Represent 26% of disciplined physicians in 2017 ─ Have 3.3 board orders and 5.8 actions per physician Conclusions of Law. ─ Received multiple actions from different boards Group 4 ─ Represent 45% of disciplined physicians in 2017 ─ Have 6.1 board orders and 9.8 actions per physician

 Board order narratives usually give more complete information than ‘Conclusions of Law’ therefore it is appropriate to move ahead with a categorization project.  Documents received from boards vary in quality, format and detail. This makes automation non-feasible at this time. Observations  Challenges with reciprocal actions remain. For statistical purposes, if an action is known to be reciprocal, it will be classified as such, regardless of associated narrative.  ‘Not applicable’ basis code will be classified as ‘No reason given.’

Bayesian method yielded the highest accuracy for entire documents. fastText provides the highest accuracy for slices which could be used to assist a human evaluator. Results Accuracy on three classes: averages 40%, up to 60% on certain training sets. These findings seem low – but are statistically significant compared to random guessing. These models were trained on limited data pools and could improve with more data.