Diagnostic performance of artificial intelligence in the syncope unit
S Van Zanten, J S Y De Jong, T T Boel, B Bais, F Giele, C Geertsma, M G Scheffer, J R De Groot, F J De LangeAbstract
Background
Transient loss of consciousness (T-LOC) has diverse causes. Artificial Intelligence (AI, GPT-5) holds promise for syncope evaluation, but its role in clinical decision making in the Syncope Unit (SU) remains unknown.
Methods
Fifty-five T-LOC patients were evaluated in SUs. The diagnostic evaluation comprised four phases: Phase-1: SU-expert assessment by SU−: guideline based initial-evaluation and SU+: initial evaluation+additional diagnostic tests. Phase-2: GPT-5 assessment using SU− and SU+, with data from Phase-1. Phase-3: follow-up, and Phase-4: diagnostic adjudication by a multidisciplinary committee. GPT-5’s diagnostic performance was assessed as defined. A fivefold reanalysis was assessed for within-case diagnostic consistency.
Results
GPT-5 achieved a diagnostic yield of 100% (SU-) and 96,3% (SU+), for SU-expert this was 94,0% for both. Diagnostic performance for GPT-5 was 51,9% (SU-) and 57,4% (SU+), for SU-expert this was 66,7% for both. GPT-5 correctly identified cardiac syncope in 11,1% (SU-) and 20,0% (SU+), for the SU-expert this was 27,3% for both. Of the four adjudicated cardiac diagnoses, GPT-5 correctly identified 25,0% (n=1), while the syncope expert identified 75,0% (n=3). The diagnostic consistency of GPT-5 revealed that 42,6% of cases produced a single diagnosis consistently across all five runs.
Conclusion
Although GPT-5 achieved a high diagnostic yield, the diagnostic accuracy and safety is low. Additional diagnostic tests (SU+) did not enhance diagnostic performance significantly (p=0,453). Moreover GPT-5 is randomly inconsistent. Until generative artificial intelligence (AI) can integrate robust clinical reasoning, probabilistic weighting, and calibrated uncertainty estimation, its role in clinical syncope decision making should remain supportive rather than autonomous.Diagnostic Yield and PerformanceDiagnostic Consistency