Large Language Models for Interpreting Standardized Neuropsychological Summaries in Idiopathic Normal Pressure...

Area of Science:

• Artificial Intelligence in Medicine
• Neurology
• Medical Diagnostics

Background:

• Idiopathic normal pressure hydrocephalus (iNPH) diagnosis and prognosis are challenging clinical issues.
• Current diagnostic and prognostic methods for iNPH require enhancement for improved patient outcomes.

Purpose of the Study:

• To evaluate the capability of large language models (LLMs) in classifying neuropsychological test (NPT) summaries for iNPH identification.
• To explore the potential of LLMs in predicting postoperative shunt responsiveness in iNPH patients.

Main Methods:

• A retrospective study involving 42 shunt-treated iNPH patients and 53 controls.
• Standardized NPT summaries were analyzed by three LLMs (ChatGPT-5, Gemini 2.5 Flash, DeepSeek) using zero-shot prompting.
• Diagnostic and prognostic performances were assessed and compared against clinical diagnoses and outcomes.

Main Results:

• ChatGPT-5 achieved the highest diagnostic accuracy (78%) and prognostic accuracy (83%) for iNPH.
• LLMs demonstrated a measurable diagnostic signal, with AUC values up to 0.84 for ChatGPT-5.
• Exploratory analysis indicated prognostic potential, with positive predictive values ranging from 87% to 96%.

Conclusions:

• LLMs applied to standardized NPT summaries show potential for assisting in iNPH diagnosis and prognosis.
• The findings suggest feasibility for NPT-based LLM assistance in clinical settings.
• Larger, externally validated cohorts are necessary before widespread clinical implementation.