Predicting the Length of Stay in Neurosurgery with RuGPT-3 Language Model.

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Natural Language Processing

Background:

• Evaluating AI language models for medical applications is crucial.
• Accurate prediction of Length of Stay (LOS) impacts healthcare resource management.
• Previous studies explored recurrent neural networks and FastText for LOS prediction.

Purpose of the Study:

• To assess the performance of the Russian GPT-3 (ruGPT-3) language model in predicting neurosurgical Length of Stay (LOS).
• To compare ruGPT-3's LOS predictions against neurosurgical doctors' and patients' expectations.
• To analyze the quality of ruGPT-3's performance using narrative clinical texts.

Main Methods:

• Utilized the Russian GPT-3 (ruGPT-3) language model for Length of Stay (LOS) prediction.
• Compared model predictions with Mean Absolute Error (MAE) against actual outcomes and human expectations.
• Analyzed performance based on narrative medical records in neurosurgery.

Main Results:

• Doctors demonstrated the most realistic LOS expectations (MAE = 2.54).
• ruGPT-3's predictions (MAE = 3.53) were closer to patients' (MAE = 3.47) but statistically inferior (p = 0.011).
• ruGPT-3 showed solid performance quality on narrative clinical texts.

Conclusions:

• The ruGPT-3 model shows promise in LOS prediction from clinical texts but is currently inferior to patient expectations.
• Further improvements in AI language models for LOS prediction are possible.
• This study updates previous findings and highlights areas for future research in medical AI.