Public articles linked to the same research event.
JMIR Medical Informatics This study deployed the open-source Qwen3-14B model on a hospital intranet server using the Ollama framework, supported it with lightweight knowledge augmentation through exact-match injection from a structured knowledge base derived from drug package inserts, and used a 2-period crossover design in which 2 pharmacists independently reviewed the same 213 outpatient prescriptions under unaided and AI-assisted conditions; human-AI collaborative review achieved 97.2% (207/213) accuracy versus 82.6% (176/213) for pharmacist-alone review, sensitivity was 98% (61/62) versus 55% (34/62), the false-negative rate fell from 45% to 2%, knowledge augmentation reduced the model hallucination rate from 19.7% (42/213) to 4.7% (10/213), mean per-prescription review time fell from 2.33 to 1.
This study deployed the open-source Qwen3-14B model on a hospital intranet server using the Ollama framework, supported it with lightweight knowledge augmentation through exact-match injection from a structured knowledge base derived from drug package inserts, and used a 2-period crossover design in which 2 pharmacists independently reviewed the same 213 outpatient prescriptions under unaided and AI-assisted conditions; human-AI collaborative review achieved 97.2% (207/213) accuracy versus 82.6% (176/213) for pharmacist-alone review, sensitivity was 98% (61/62) versus 55% (34/62), the false-negative rate fell from 45% to 2%, knowledge augmentation reduced the model hallucination rate from 19.7% (42/213) to 4.7% (10/213), mean per-prescription review time fell from 2.33 to 1.
This study deployed the open-source Qwen3-14B model on a hospital intranet server using the Ollama framework, supported it with lightweight knowledge augmentation through exact-match injection from a structured knowledge base derived from drug package inserts, and used a 2-period crossover design in which 2 pharmacists independently reviewed the same 213 outpatient prescriptions under unaided and AI-assisted conditions; human-AI collaborative review achieved 97.2% (207/213) accuracy versus 82.6% (176/213) for pharmacist-alone review, sensitivity was 98% (61/62) versus 55% (34/62), the false-negative rate fell from 45% to 2%, knowledge augmentation reduced the model hallucination rate from 19.7% (42/213) to 4.7% (10/213), mean per-prescription review time fell from 2.33 to 1.
This study deployed the open-source Qwen3-14B model on a hospital intranet server using the Ollama framework, supported it with lightweight knowledge augmentation through exact-match injection from a structured knowledge base derived from drug package inserts, and used a 2-period crossover design in which 2 pharmacists independently reviewed the same 213 outpatient prescriptions under unaided and AI-assisted conditions; human-AI collaborative review achieved 97.2% (207/213) accuracy versus 82.6% (176/213) for pharmacist-alone review, sensitivity was 98% (61/62) versus 55% (34/62), the false-negative rate fell from 45% to 2%, knowledge augmentation reduced the model hallucination rate from 19.7% (42/213) to 4.7% (10/213), mean per-prescription review time fell from 2.33 to 1.