ChartOCR: A Longitudinal OCR Pipeline Using a Unified Periodontal Assessment Form

Presentation Type: Paper - Student

Click to View Presentation

2026 Amplify 25x5 Presentation

Presentation Time: 09:45 AM - 09:57 AM

Primary Track: Data Science/Artificial Intelligence


Periodontitis generates substantial clinical data, yet periodontal and plaque index charts remain predominantly paper-based, which limits reuse for clinical analysis and decision support. We present ChartOCR, a system that digitizes unified periodontal-plaque forms using template-based alignment, coordinate-based region-of-interest (ROI) extraction, a fine-tuned TrOCR model, and SHA-256 caching. ChartOCR supports common handwriting conventions and provides an interactive interface with real-time result streaming, in-place correction, and CSV export. We evaluated the system on 30 scanned handwritten charts. Periodontal charts achieved a mean accuracy and mAP of 0.987, while plaque index charts reached 0.942 accuracy and 0.884 mAP. Most digits performed well, although PPV was lower for higher plaque scores. On a MacBook Pro (M3 Pro), the average processing time per chart was approximately 70 seconds. These results demonstrate that our approach enables accurate, efficient digitization of handwritten dental charts and provides a scalable foundation for periodontal informatics research.

Speaker(s):
Yunseo Moon, BS
Seoul National University

Author(s):
Yunseo Moon, BS - Seoul National University; Minh Do Ngoc Luong, BS - Seoul National University; Seungjun Chong, BS - Seoul National University; Seung-Hee Ryu, RDH, MS - Seoul National University; Hye-Jin Hyun, RDH - Seoul National University; Hyun-Jae Cho, DDS, PhD - Seoul National University; Hyunggu Jung, PhD - Seoul National University;

---

Pre–Post Evaluation of Documentation Burden, Time Perception, and Cognitive Workload of Ambient AI Scribe Tools

Presentation Type: Paper - Student

Click to View Presentation

2026 Amplify 25x5 Presentation

Presentation Time: 09:57 AM - 10:09 AM

Primary Track: Data Science/Artificial Intelligence


Documentation burden and cognitive load contribute to clinician burnout, and ambient AI scribes offer a promising solution, though real-world evaluations are limited. This study assessed the impact of an ambient AI scribe on documentation burden, cognitive workload, and perceived burnout in clinical practice. Forty clinicians participated in a 60-day pilot where the AI scribe generated draft notes from encounter audio. Pre- and post-implementation surveys included a validated single-item burnout measure, workflow questions, NASA-TLX workload scores, and EHR audit data was assessed on after-hours documentation. Clinicians reported significant reductions in perceived documentation time (p < 0.001) and after-hours work (p < 0.001). NASA-TLX scores also decreased across mental demand, effort, time pressure (all three p < 0.001), and frustration (p = 0.005). Objective EHR documentation minutes did not change significantly (p = 0.148), but perceived time aligned more closely with actual categories. Ambient AI scribes may reduce perceived workload and improve calibration of documentation time.

Speaker(s):
Leopold Arko, MD, MS
University of Minnesota

Author(s):
Leopold Arko, MD, MS - University of Minnesota; Tony Huy Nguyen, MD - University of Minnesota Clinical Health Informatic Fellowship; Michelle Stoffel, MD, PhD - University of Minnesota and M Health Fairview; Jaya Kumar, MD, MBA - Fairview Health Services; Sameer Badlani, MD, FACP - M Health Fairview; Genevieve Melton-Meaux, MD, PhD, FACMI - University of Minnesota; Rebecca Markowitz, MD - M Health Fairview;

---

Large Language Models for Temporal PHI De-identification in Real-World Clinical Notes: Performance and Limitations

Presentation Type: Podium Abstract

Click to View Presentation

Presentation Time: 10:09 AM - 10:21 AM

Primary Track: Clinical Research Informatics


Large language models (LLMs) show strong potential for temporal PHI de-identification, but their reliability in real-world clinical notes remains uncertain. Using a sarcoma cohort, we assessed zero-shot LLM performance in extracting temporal PHIs and creating temporal surrogates. While LLMs maintained formatting well, they often have difficulty handling hybrid date-time formats and disrupted chronological order, especially in notes with dense temporal information. These results highlight the strengths and limitations of LLM-based temporal de-identification.

Speaker(s):
Xiaomeng Wang, Master of Science
University of Texas Health Science Center at Houston

Author(s):
Xiaomeng Wang, Master of Science - University of Texas Health Science Center at Houston; Liwei Wang, MD, PhD - UTHealth; Rui Li, Phd - UT health; Andrew Wen, MS - University of Texas Health Sciences Center at Houston; Wanjing Wang, MS - University of Texas Health Science Center at Houston; Huipeng Liu, MS - University of Texas Health Science Center at Houston; Qiuhao Lu, Ph.D. - University of Texas Health Science Center at Houston; Heather Lyu, MD, MBI - University of Texas, MD Anderson Cancer Center; Hongfang Liu, PhD - University of Texas Health Science Center at Houston;

---

RadAnnotate: Large Language Models for Efficient and Reliable Radiology Report Annotation

Presentation Type: Paper - Student

Click to View Presentation

2026 Amplify 25x5 Presentation

Presentation Time: 10:21 AM - 10:33 AM

Primary Track: Data Science/Artificial Intelligence


Radiology report annotation is essential for clinical NLP, yet manual labeling is slow and costly. We present
RadAnnotate, an LLM-based framework that studies retrieval-augmented synthetic reports and confidence-based selective automation to reduce expert effort for labeling in RadGraph. First, we train entity-specific classifiers on
gold-standard reports and characterize their strengths and failure modes across anatomy and observation categories, with uncertain observations hardest to learn. Second, we generate RAG-guided synthetic reports and show that synthetic-only models remain within 1–2 F1 points of gold-trained models, and that synthetic augmentation is especially helpful for uncertain observations in a low-resource setting, improving F1 from 0.61 to 0.70. Finally, by learning entity-specific confidence thresholds, RadAnnotate can automatically annotate 55–90% of reports at 0.86–0.92 accuracy while routing low-confidence cases for expert review.

Speaker(s):
Saisha Shetty, Masters of Science in Computer Science
University of California, Davis

Author(s):
Saisha Shetty, Masters of Science in Computer Science - University of California, Davis; Roger Goldman, MD, PhD - UC Davis; Vladimir Filkov, PhD - UC Davis;

---

Exploring Thematic Differences in Ambient Scribe vs Traditional Clinical Notes Using Topic Modeling

Presentation Type: Podium Abstract

Click to View Presentation

2026 Amplify Health Equity Presentation

2026 Amplify 25x5 Presentation

Presentation Time: 10:33 AM - 10:45 AM

Primary Track: Data Science/Artificial Intelligence


Large language model–based “ambient scribe” systems are increasingly used to generate clinical notes, yet their linguistic differences from traditionally written notes remain unclear. We analyzed thematic patterns in HPIs written with and without ambient scribe in a cohort of patients using topic modeling with BERTopic. Topic prevalence differed significantly between groups (p<0.05), revealing distinctive clusters of generated versus manual narratives. Qualitative review highlights stylistic shifts in clinical documentation associated with LLM assistance.

Speaker(s):
Erik Holbrook, MD
Mass General Brigham

Author(s):
Erik Holbrook, MD - Mass General Brigham; Winston Guo, MD - Brigham and Women's Hospital; Suzane Blackley, MA - Mass General Brigham; Li Zhou, MD, PhD - Brigham and Women's Hospital;

---

End-to-end extraction of temporal information from psychiatric discharge summaries

Presentation Type: Paper - Regular

Click to View Presentation

Presentation Time: 10:45 AM - 10:57 AM

Primary Track: Data Science/Artificial Intelligence


Among the many types of information embedded in electronic health records that may yield insights into patients' health trajectories, events and the relations between them hold particular promise. But this area is poorly explored, particularly within the psychiatric domain.
To fill this gap, we trained five new models to extract events, time expressions, and relations from psychiatric notes. We evaluated these models alongside a similar model trained to do the same task on colon cancer notes. We found that each of the psychiatrically fine-tuned models outperformed the pre-existing model; among the new models, some training choices gave small performance improvements over others.

Speaker(s):
Tim Miller, PhD
Children's Hospital Boston/Harvard Medical School

Author(s):
Spencer Thomas, MS - Boston Children's Hospital; Gaby Dinh, MS - Boston Children's Hospital; WonJin Yoon, PhD - Boston Children's Hospital / Harvard University; Boyu Ren, PhD - McLean Hospital; Guergana Savova, PhD - Boston Children's Hospital and Harvard Medical School; Mei-Hua Hall, PhD - McLean Hospital/Harvard Medical School; Tim Miller, PhD - Children's Hospital Boston/Harvard Medical School;

---