Smarter Together: Combining Large Language Models and Small Models for Physiological Signals Visual Inspection

Hybrid AI systems outperform single approaches: The research demonstrates that combining large language models (LLMs) with specialized smaller models improves accuracy on medical tasks—for example, enhancing Qwen2-VL-7B's precision to 94.92% on arrhythmia detection, showing that no single AI model type is optimal for complex medical decisions.

Uncertainty quantification matters for clinical safety: The framework uses "conformal prediction" to provide statistical reliability guarantees on AI outputs, distinguishing between confident predictions (94.92%-96.82% precision) and uncertain ones (70.61%-78.10% precision)—a critical feature for responsible deployment where clinicians need to know when to trust AI recommendations versus seek human review.

Interpretability enables clinical adoption: The QTrans-Pooling mechanism identifies which specific segments of physiological signals (like heartbeat patterns) drive the AI's decisions, addressing a major barrier to responsible AI in healthcare where clinicians need to understand why the system made a recommendation before acting on it.

Specialized models address proprietary AI limitations: By combining smaller, customizable models with general-purpose LLMs, the approach overcomes challenges with proprietary models that cannot be fine-tuned on sensitive clinical datasets, offering a pathway for hospitals and clinics to implement responsible AI while protecting patient privacy.

Decision support, not replacement: The framework is explicitly designed as a "decision-support" tool rather than autonomous decision-maker, maintaining human clinicians in the loop—a key principle for responsible AI in high-stakes medical settings.