CognAlign: A Multi-Agent Cognitive-Alignment Framework for Transparent, Bias-Aware Medical Triage Using Small Language Models

CognAlign presents a modular, multi-agentic system designed to enhance transparency, bias-awareness, and adaptability in Small Language Models (SLMs) for high-stakes decision-making, particularly medical triage. By integrating cognitive dual-process reasoning, bias sentinels, and structured rationale generation, CognAlign enables SLMs to dynamically allocate computational resources, detect and mitigate bias, and produce interpretable outputs, supporting clinician oversight and equitable triage decisions. The system was evaluated on clinically validated triage cases, including mis-triaged “tricky” cases, supplemented by development datasets exceeding 1,000 samples. Performance was assessed across clinical accuracy, bias absence, patient safety, resource optimization, transparency, and routing latency. Compared to baseline SLMs, CognAlign reduced clinical errors by 24%, eliminated patient-safety violations, increased transparency from 0 to 0.489, and improved overall performance by 9.7%, while maintaining efficiency and bias-absence. Tricky cases were more consistently routed to System 2, demonstrating effective recognition of uncertainty and deeper reasoning. Lightweight deployment on phi-3 SLMs maintained high clinical reliability and bias mitigation without requiring cloud resources. Feedback from professionals across multiple healthcare systems indicates practical utility, enabling nurses to shift from direct triage to oversight and supporting phased adoption in less formalized systems. CognAlign demonstrates the potential of cognitive-aligned, low-compute SLMs to provide interpretable, bias-aware, and safe AI outputs. Future testing on larger datasets, across diverse SLMs, and alongside real-time clinician comparisons could expand robustness. Beyond medical triage, its architecture offers a transferable framework for critical decision-making domains, enabling transparency, fairness, and efficiency.