Validating AI-Based Detection of I-PASS Elements in Verbal Handoffs

Joshua Pankin, MD ¹ · Vishal Pallerla ² · Samantha Pendleton, DO ¹ · Hannah Miller, MD ¹ · Don Woodlock ² · Jonathan Teich, MD ² · Qi Li, MD ² · Christopher Landrigan, MD, MPH ³ · Molly Senn-McNally, MD ¹ · Amy Starmer, MD, MPH ¹

• ¹ Baystate Children's Hospital, Springfield, MA
• ² InterSystems Corporation
• ³ Boston Children's Hospital, Boston, MA

Overall accuracy

94.8%

Simulated handoffs

77

Element calls scored

385

Action List detection

100%

We built and validated a Whisper + GPT pipeline that automatically detects each of the five I-PASS handoff elements in verbal pediatric handoffs. Across 77 simulated cases (385 element-level assessments), the system reached 94.8% overall accuracy against a 10-reviewer consensus standard. 80% of errors were false negatives, the safer failure mode for educational feedback, since the tool defers rather than falsely confirms adherence. Action-List detection was perfect (100%); receiver-synthesis was the lowest performer (87%), driven by ASR errors and overlapping speech.