The application of artificial intelligence (AI) in clinical settings is gaining traction, particularly in augmenting decision-making processes. However, its effectiveness in diagnosing pediatric conditions, especially rare diseases, has been less clear. A study published in Pediatric Investigation sheds light on this, revealing that AI models have surpassed human clinicians in diagnostic accuracy for rare conditions, with a collaborative human-AI approach yielding the best results overall. These findings suggest AI’s potential as a valuable tool in enhancing diagnostic precision and improving patient care outcomes.
Pediatric diagnosis often presents unique challenges due to the subtlety and overlap of symptoms, particularly in rare diseases, leading to potential delays in treatment. Although AI has shown promise in healthcare, many prior studies have utilized simplified datasets rather than complex real-world clinical data, leaving a gap in understanding AI’s effectiveness in everyday clinical environments. This study, led by Dr. Cristian Launes from Hospital Sant Joan de Déu in Barcelona, Spain, aimed to address this by evaluating AI performance using actual pediatric clinical cases. The research involved comparing four advanced AI language models with 78 pediatric clinicians across 50 real-world cases, encompassing both common and rare conditions.
Dr. Launes, an expert in pediatric infectious diseases, guided the study by using patient summaries from the initial 72 hours of presentation to reflect real clinical scenarios. The study assessed diagnostic accuracy based on whether the correct diagnosis was among the top predictions. Results indicated that AI models demonstrated superior accuracy, particularly in diagnosing rare diseases. However, clinicians showed strengths in certain complex scenarios, highlighting the complementary nature of human and AI diagnostic reasoning.
The study did not test a real-time interactive “human-plus-AI” diagnostic workflow but estimated complementarity through a “union” approach, asking whether the correct diagnosis appeared within the top five suggestions from either clinicians or AI models. The best-performing combination achieved a 94.3% Top-5 union accuracy, indicating the potential for AI to serve as a second opinion, especially in complex cases involving rare diseases. Dr. Launes emphasized that AI could broaden differential diagnoses and reduce missed diagnoses, provided its outputs are critically interpreted within robust oversight frameworks.
From a governance standpoint, diagnostic AI systems are considered high-risk applications under the European Union AI Act, necessitating stringent risk management, data governance, and human oversight. The study also highlighted that additional clinical data, such as lab results, improved diagnostic accuracy for both AI and clinicians, suggesting that AI systems are most effective when integrated into dynamic, data-rich clinical workflows. Overall, the study underscores the promising role AI can play in pediatric healthcare, particularly in diagnosing rare conditions, when used alongside human expertise.