AI-powered real-time clinical documentation in oncology and internal medicine: A pilot study of automated SOAP note.

Abstract

735 Background: Administrative burden from clinical documentation consumes an estimated 2–3 hours per physician per day, reducing time available for direct patient care. In oncology, visit complexity is high and thorough notes are critical for high-value care, often pushing documentation ahead of patient care. Artificial intelligence (AI)-assisted documentation tools offer a potential solution, yet real-world pilot data across oncology and primary care settings remain limited. Methods: We conducted a pilot randomized controlled study evaluating PracticeOS, an AI-powered clinical documentation platform, across 10 standardized simulated clinical scenarios spanning oncology, internal medicine, and primary care. The platform uses real-time audio recording, automatic speech recognition (OpenAI Whisper), and large language model-based (Claude, Anthropic) structured SOAP (Subjective, Objective, Assessment, Plan) note generation. We evaluated 24 simulated patient encounters, with 12 in each study arm. Simulated clinicians independently authored timed SOAP notes (control); the same scenarios were then processed through PracticeOS and timed from recording initiation to note generation (intervention). All notes were independently scored by a reviewer on a validated 1-5 scale assessing completeness, accuracy, clinical appropriateness, and structure, blinded to note origin. Results: Mean documentation time was reduced from 15.2 ± 2.9 minutes (manual) to 15.7 ± 2.9 seconds (PracticeOS), a 98.3% reduction (t=16.57, p<0.001, Cohen's d=5.24). Note quality scores improved from 2.90 ± 0.70 to 4.85 ± 0.34 out of 5.0, a 1.95-point mean improvement (t=8.51, p<0.001, Cohen's d=2.69). AI-generated notes achieved near-perfect quality scores across all 10 scenarios. The platform supported multilingual encounters including English and regional South Asian languages. Note-editing frequency and language-support utilization were tracked as secondary outcomes. Conclusions: This pilot demonstrates feasibility of real-time AI-assisted SOAP note generation across oncology and internal medicine settings, with statistically significant improvements in both documentation speed and note quality. Effect sizes were very large for both outcomes (Cohen's d > 2.5), suggesting robust and clinically meaningful differences. Larger prospective studies integrating real clinical encounters, blinded physician scoring, and EHR workflow integration are warranted. Key statistics summary. Metric Manual PracticeOS Change p-value Cohen's d Documentation time 15.2 ± 2.9 min 15.7 ± 2.9 sec −98.3% <0.001 5.24 Note quality score (1–5) 2.90 ± 0.70 4.85 ± 0.34 +1.95 pts <0.001 2.69