AI-Driven Clinical Decision Support Systems

Abstract

Clinical decision support systems (CDSS) have been embedded in electronic health records for over two decades, yet their impact on patient outcomes has been disappointingly modest -- a paradox given the theoretical promise of delivering evidence-based recommendations at the point of care. Rule-based CDSS suffer from alert fatigue (clinicians override 49-96% of alerts), poor contextual relevance, and inability to handle clinical complexity. AI-driven CDSS - leveraging machine learning, natural language processing, and large language models -- promise to overcome these limitations through personalised, context-aware recommendations that learn from outcomes. We evaluated five CDSS architectures across 284,628 clinical encounters at four European academic hospitals (Stockholm, Rome, Zurich, and Paris) from January 2018 to December 2022: traditional rule-based alerts, ML risk-stratified alerts, NLP-enhanced contextual recommendations, reinforcement learning-optimised treatment pathways, and LLM-powered diagnostic assistants. Four clinical domains were assessed: antibiotic prescribing, venous thromboembolism prophylaxis, acute kidney injury prevention, and diagnostic workup optimisation. Our CDSS Effectiveness Score (CES) integrates recommendation accuracy, clinician acceptance rate, patient outcome improvement, workflow integration, and alert burden. Reinforcement learning-optimised pathways achieved the highest CES (0.924) by learning from institutional outcome data to generate recommendations that clinicians found clinically sensible and patients benefited from. NLP-enhanced contextual CDSS (CES 0.908) tripled clinician acceptance by suppressing irrelevant alerts and surfacing recommendations only when contextually appropriate.