International Telemedicine Program: Physician Versus AI Responses

Abstract

Introduction: Asynchronous, provider-to-provider telemedicine is increasingly used to extend specialist expertise to frontline clinicians in low-resource settings,[1–4] yet the role of artificial intelligence (AI) clinical decision support in these workflows remains unclear.[5,6] The Telemedicine Health and Empowerment Program of a US based NGO links U.S.-based volunteer physicians with African providers through a secure, store-and-forward platform.[3] Open Evidence is an AI decision support tool that generates differentials and management suggestions from structured clinical narratives.[5] This study compared how Open Evidence and human volunteers responded to the same real-world telemedicine consultations.

Methods: A retrospective comparative case series of 50 telemedicine consultations from African providers that had de-identified clinical narratives and written responses from volunteer physicians was conducted. The same narratives were entered into OpenEvidence, and AI outputs were captured without additional prompting. For each physician and AI, the response was recorded for the presence of four elements: (1) explicit diagnostic list or working diagnosis, (2) recommended next diagnostic steps (history, examination, laboratory tests, or procedures), (3) immediate treatment recommendations, and (4) identification of red flag features or follow-up concerns. Elements were compared descriptively between AI and physician responses. 

Results: Seven consultations (14%) asked whether an initial assessment was correct, 9 (18%) requested guidance on further workup, and 34 (68%) sought a broader differential diagnosis. AI responses included a diagnostic list in 41 cases (82%), suggested next diagnostic steps in 50 (100%), immediate treatment recommendations in 50 (100%), red flag features in 26 (52%), and follow-up concerns in 37 (74%). Physician responses included an explicit diagnostic list in 12 cases (24%), suggested next diagnostic steps in 42 (84%), immediate treatment recommendations in 29 (58%), red flag features in 4 (8%), and follow-up concerns in 9 (18%).

Conclusions: In this African–U.S. telemedicine program, AI-generated responses were more structurally complete and safety-oriented, whereas physicians offered more selective, context-focused guidance. These findings support a hybrid telehealth workflow in which AI provides systematic completeness and safety checks while human consultants supply contextual judgment and local prioritization. [5–7]