Matching Artificial Intelligence Methods to Research Problems in Medical Informatics

Artificial intelligence is reshaping scientific research, yet its contribution may be associated with whether methods are matched to the types of research problems. This study investigates this question in medical informatics using 40,107 articles published during 2000–2024. Large language models are used to identify three AI method families—traditional machine learning, neural network methods, and generative AI—while a knowledge-distilled classifier groups studies into three problem types: comprehension-oriented, solution-oriented, and exploration-oriented. The study then examines how different problem–method combinations are associated with citation impact and recombinatorial novelty. AI use increased sharply after 2018 and became increasingly embedded in medical informatics. All three methods are positively associated with citation impact, with relatively stronger associations for generative AI. Associations vary by problem type: all three methods are positively associated with impact in solution-oriented studies; neural networks and generative AI are positively associated with impact in comprehension-oriented studies, whereas neural networks are negatively associated with impact in exploration-oriented studies. Traditional machine learning is positively associated with novelty in comprehension-oriented and solution-oriented studies, whereas neural networks are positively associated only in solution-oriented studies. These findings suggest that AI’s contribution to science is associated with specific problem–method combinations and that broad claims about AI’s role can obscure heterogeneity across research tasks.