Archived File The file below has been archived for historical reference purposes only. The content and links are no longer maintained and may be outdated. See the OER Public Archive Home Page for more details about archived files.
Automated Referral Workflow System Facilitating Referral Through Text Mining National Institutes of Health Department of Health and Human Services 2
Role of Referral • CSR is portal for all incoming competing applications § > 70, 000 in FY 2006 • Referral to CSR Integrated Review Group (IRG) or Institute (IC) Review Branch • Referral to IC for potential consideration • Factors include § PI requests in cover letters § Written guidelines Automated Referral Workflow System 3
Benefits of Automated Referral • Shortened CSR review cycle • Improved § § Speed Efficiency Transparency Consistency Automated Referral Workflow System 4
How? • Automated Cover Letter Mining § § 47% of PIs Request a SRG CSR usually initially refers PI requests to requested IRG and 89% remain there • Machine Learning Algorithms • Referral by Experts Will Continue for Difficult Cases Automated Referral Workflow System 5
After identifying the top matching grants from among 1000 s… DIG HEME Target Grant DIG HEME HEME … Machine learning algorithm assigns the target grant to the HEME IRG because it is the most common IRG in the top N matches. Automated Referral Workflow System 6
Machine Learning Experiment • All R 01 s reviewed by standing CSR review groups in 1 -year period (about 20, 000) • About 15, 000 applications in historical set (some could not be accessed), about 4500 were also used as test applications (initial submissions without PI requests) § Electronic submission will dramatically facilitate this approach • Abstract & Specific Aims • Predictions based on top 15 matches Automated Referral Workflow System 7
Dependent Variables • Agreement: Concurrence between automated referral prediction and historical assignment made by human experts § § Imperfect proxy for real-world “acceptance” IRG referral guidelines have overlap • Yield: Percentage of applications for which prediction may be made at a given accuracy level Automated Referral Workflow System 8
IRG Assignment Prediction Automated Referral Workflow System 9
Implications for IRG Referral Model Referral Method Percentage PI assignment request automatically mined from cover letter 47 Machine learning prediction when no PI request and > 80% likelihood of agreement with human experts Remainder referred by human experts 39 Automated Referral Workflow System 14 10
IC Assignment Prediction Automated Referral Workflow System 11
IC Referral Considerations • Programmatic areas of overlap or disagreement among ICs • What is appropriate consideration of prior grant or training support from an IC? • Protecting interests of smaller ICs – ensuring algorithms do not inappropriately favor larger ICs • Stakeholders, including ICs, will be consulted • Steering body will include IC representatives Automated Referral Workflow System 12
Next Steps – IMPAC II Database Integration • ARWS will be web service maintained by CSR ARWS will be ready to make some automated referrals in February but as yet there is no way to automatically pass referrals into IMPAC II (enterprise grants database) Interface to IMPAC database is critical • • § • Working with IMPAC to prioritize and hopefully have interface built Savings that could be realized by implementing interface § § Fewer referral officers Review meetings 2 -3 weeks earlier Automated Referral Workflow System 13
Acknowledgements • Support § Office of the Director § Extramural Affairs Working Group • ARWS Project Team § § § CSR Staff IC Staff e. RA Staff Discovery. Logic Emergint (assistance with related, earlier pilot study) 14
Скачать презентацию Archived File The file below has been archived
d323077092b595e8e38824c8124a2b22.ppt