Authors :
Presenting Author: Amerta Bai, MD, MS – University of Minnesota
Gayal Kuruppu, B.Sc – University of Minnesota
Aisha Abdul-Razaq, MD – University of Minnesota
Zhiyi Sha, MD, PhD – University of Minnesota
Abdulrahman Alwaki, MD – Department of Neurology, McGovern Medical School, The University of Texas Health Science Center at Houston, Houston, TX, USA
Jay Gavvala, MD – UTHealth Houston
Yogatheesan Varatharajah, PhD – University of Minnesota
Sandipan Pati, MD – University of Minnesota
Rationale:
Background: Responsive Neurostimulation (RNS) produces rich patient data metrics (PDMS) critical for optimizing therapy in drug-resistant epilepsy. Yet, these reports remain siloed—buried in secure servers, absent from clinical documentation, and inaccessible to non-specialists. This gap impedes care coordination and misses billing opportunities, leaving vital data underutilized.
Objective: We developed and validated a novel fully automated software tool to transform complex RNS PDMS reports into structured, clinician-ready summaries, seamlessly integrated into medical records for broader care team use.
Methods:
Our automated pipeline combines structured data extraction and logic-driven narrative generation to convert RNS summary data available within portable document format (PDF) reports into textual summaries that can be readily integrated within electronic health records (EHR). Eleven clinicians (6 neurology residents, 5 epilepsy attendings) evaluated 30 AI-generated reports across six domains—accuracy, completeness, coherence, relevance, bias, and toxicity—using a 5-point Likert scale (330 total assessments). Group differences were analyzed via Mann-Whitney U tests; inter-domain variation used Friedman/Wilcoxon tests with Bonferroni correction.
Results:
Automated software-generated summaries achieved near-perfect median scores for accuracy, completeness, coherence, and relevance (4.00/5.00), with negligible bias/toxicity (1.00/5.00). Attendings were slightly more critical than residents (p< 0.001–0.011). Relevance and coherence ranked highest (χ²[5]=1424.669, p< 0.001), confirming clinical utility.