5492. Improving Learning Outcomes with an Adaptive Intelligent Tutor: First Pilot Study
Authors * Denotes Presenting Author
  1. Julian Lopez Rippe *; Children's Hospital of Philadelphia
  2. Manasa Reddy; Children's Hospital of Philadelphia
  3. Maria Velez-Florez ; Massachusetts General Hospital
  4. Avin Khera; Children's Hospital of Philadelphia; Perelman School of Medicine
  5. Ami Gokli; Children's Hospital of Philadelphia; Staten Island University Hospital
  6. Michael Francavilla; Children's Hospital of Philadelphia
  7. Janet Reid; Children's Hospital of Philadelphia
Artificial intelligence (AI) use in medical imaging is rapidly expanding, but most uses in radiology focus on lesion detection and age-related normal anatomy. We present progress in implementing and assessing the Intelligent Tutor (IT), a novel machine learning (ML)-derived recommender algorithm that delivers knowledge to enhance image interpretation while contemporaneously creating an individualized curriculum to support precision education.

Materials and Methods:
Four phases include 1) completion and integration, 2) heatmaps, 3) rollout, 4) assessment (internal and external). We report phase 4 that evaluated radiology trainees at Children's Hospital of Philadelphia via control group and intervention group (with/without exposure to IT, respectively) at the start and end of a month’s rotation. The pre-assessment evaluated educational experience, background, and reporting skills for three pediatric radiology cases. The post-assessment evaluated their report of three similar cases, their technology acceptance, and cognitive load using NASA-TLX. Descriptive statistics were reported for surveys. Mann-Whitney and Kruskal-Wallis test analyzed difference in reports (<em>p</em> < 0.05).

We recruited nine fellows (23.1%) and 30 residents (76.9%) who rely on digital learning tools for 75.1% of their learning. Only 21.3% of their searches were typically done on learning management systems. Time spent was case-dependent, with no significant differences in resource usage or reporting time for cases of lower complexity (mean dif. 154 s, <em>p</em> = 0.103). For more complex cases, the intervention group took significantly less time on resourcing and reporting (mean dif. 700 s, <em>p</em> < 0.001), with higher accuracy (score % mean dif 0.379, <em>p</em> < 0.001). IT technology acceptance was good: 70% reported it easy to use with increased job productivity. Both groups showed similar cognitive load (mean dif. 2.42 pts, <em>p</em> = 0.456).

This ML-based tool enhances learners' case reporting, particularly in complex cases. IT is user-friendly and can enhance productivity and may reduce cognitive load over a longer period of use. We plan to externally validate the tool's performance at the national and international levels. With powerful computer infrastructure and the advent of machine learning, we are now able to move from the "information era" to a "knowledge era" to create true precision learning in radiology.