Enhancing Oncology Documentation and Diagnosis Capture With Integrated AI Scribe in Electronic Health Record

Citation:

Abstract 2170038

Accurate clinical documentation is critical in oncology, influencing decision-making, care coordination, billing, and research. However, manual entry into electronic health records (EHRs) is time-intensive and prone to omissions, particularly for comorbidities. Complex patient histories and treatment plans further exacerbate oncology documentation challenges, contributing to workflow inefficiencies, clinician burnout, and impaired physician-patient interactions.  

Artificial intelligence (AI) scribe tools offer a promising solution by generating real-time clinical notes and diagnosis suggestions. This study evaluated the impact of integrating DeepScribe’s oncology-specific AI scribe into Flatiron Health’s OncoEMR system on documentation efficiency, diagnosis completeness, and clinician experience.

A pre-post implementation study was conducted across 17 community oncology practices comprising 258 clinic sites. The study included 215 physicians and covered 405 210 patient encounters from 180 days before each site’s launch date through June 22, 2025. Diagnosis-related outcomes were assessed in a subset of 4 practices and 71 providers who reviewed AI-suggested diagnoses.

AI-assisted and non-AI-assisted patient encounters were compared across pre- and post-implementation periods. Quantitative metrics included provider adoption and usage, 72-hour note closure rates, and the number and complexity of diagnoses per encounter and patient. Qualitative feedback was derived from clinician interviews regarding usability, impact on workflow, and satisfaction.

AI scribe adoption reached 98%, with 47% of clinicians using it for more than half of their encounters. The 72-hour note closure rate for AI-assisted encounters was nearly the same (84.2%) compared to non-AI-assisted encounters (83.5%). Despite this, clinicians reported reduced time spent on documentation, improved clinic flow, enhanced note quality, decreased cognitive burden, and high usability.

Amongthe reviewed AI-suggested diagnoses, 35% were added to patient charts. AI-assisted encounters were associated with a 16% increase in diagnoses per visit and higher proportions of comorbidities (61% vs 50%) and social determinants of health (SDOH) (16% vs 11%) compared with manually added diagnoses. Additionally, 45% of AI-assisted diagnoses included ICD-10 codes with 5 or more characters vs 37% for manually entered diagnoses.  

At the patient level, those who had at least one AI-suggested diagnosis had nearly twice as many recorded diagnoses (mean 5.11 vs 2.73) and 17% more ICD codes linked to evaluation and management (E/M) charges.

Integration of an AI scribe into the EHR demonstrated high adoption and favorable clinician feedback. There was not a dramatic difference in AI-assisted note closure rates. Qualitative feedback suggests this may reflect reduced urgency to finalize notes due to decreased cognitive load. Clinicians described a "mental offloading" effect, enabling more thoughtful and less stressful documentation after the visit.    

The accuracy of the AI-generated notes was not explicitly evaluated. Instead, providers were entrusted to ensure accuracy and clinical quality before finalization as an exercise in real-world usability and workflow integration. AI-assisted documentation increased the number, length, and diversity of recorded diagnoses—particularly for comorbidities and SDOH—which may improve the quality of clinical care, outcomes, and billing accuracy.     

These findings support the clinical utility and scalability of AI-integrated documentation tools in complex care settings. By reducing real-time documentation demands and improving data abstraction, AI scribes may enhance provider well-being and enable more sustainable clinical practice.

Authors:

Eileen Campetti, PA-C, MS¹; Saranga Arora, MS, MBA¹; Helena Hay¹; Jennifer Siggia, RN, MSN²; Stephen Speicher, MD, MS¹

Affiliations:

¹Flatiron Health, New York, NY; ²DeepScribe, San Francisco, CA

1. O'Malley AS, Peikes D, Ghosh A, et al. EHR Workload and Inbox Burden Among Oncology Clinicians. J Natl Compr Canc Netw (JNCCN). 2024;22(1):45-52. doi:10.6004/jnccn.2023.7105
2. Zhou L, Lyles A, Linder J, et al. Time to Chart Closure Predicts Burnout in Outpatient Physicians: A Cross-Sectional Study of 242,432 Encounters. J Am Med Inform Assoc (JAMIA). 2024;31(2):207-215. doi:10.1093/jamia/ocad210
3. Shanafelt TD, Dyrbye LN, Sinsky C, et al. Electronic Health Record Use and the Work of Oncologists: Implications for Burnout and Well-Being. JCO Oncol Pract. 2023;19(6):365-373. doi:10.1200/OP.22.00674