<< Back to the abstract archive
Small but Mighty: Local AI Models for Secure and Accessible Medical Data Extraction
Christopher Slater B.S., Min-Jeong Cho M.D.
The Ohio State University College of Medicine
2025-01-10
Presenter: Christopher Slater
Affidavit:
I certify that the material proposed for presentation in this abstract has not been published in any scientific journal or previously presented at a major meeting. The program director is responsible for making a statement within the confines of the box below specific to how much of the work on this project represents the original work of the resident. All authors/submitters of each abstract should discuss this with their respective program director for accurate submission of information as well as the program director's approval for inclusion of his/her electronic signature.
Director Name: Min-Jeong Cho
Author Category: Medical Student
Presentation Category: Basic Science Research
Abstract Category: General Reconstruction
Purpose: Artificial intelligence (AI) offers transformative potential in medical data extraction. However, large models relying on third-party servers pose significant data security risks, with 88.4% of last year's 121 million medical record breaches involving server vulnerabilities. Small, locally deployed AI models provide a safer, more accessible alternative, capable of running on standard consumer-grade hardware while maintaining high accuracy and security.
Methods: A local AI model with 8 billion parameters analyzed 357 breast tissue expander placement notes, extracting structured data on surgical laterality, tissue expander brand, and tissue expander plane-a well-defined but variably described data point in operation notes. Manual data extraction served as a benchmark for accuracy and efficiency, with time required for data collection recorded.
Results: The AI model achieved 96.6% accuracy in identifying surgical side and 97.8% for tissue expander brand. For the tissue expander plane, it achieved 94.9% accuracy, statistically comparable to manual review at 93.8% (p=0.515). The AI model significantly outpaced manual methods in efficiency, averaging 0.507 seconds per item for laterality and brand data and 2.1 seconds per chart for tissue expander plane data, compared to 8.14 seconds and 45.8 seconds, respectively, for manual review (p<0.05). These results align with published human standards (93.4%) and rival the performance of larger AI models (90-100%).
Conclusion: Small, local AI models offer secure, rapid, and accurate data extraction, reducing barriers for researchers and promoting inclusivity. They present a transformative opportunity to enhance research efficiency, data security, and equity in advancing medical fields such as plastic surgery.
