Intravenous immunoglobulin weaning evaluation with zero-shot large language model classification.

OBJECTIVES

The aims of this study were to determine the performance of a large language model (LLM) with a zero-shot strategy for the classification of several factors relevant to the consideration of intravenous immunoglobulin (IVIg) weaning.

BACKGROUND

In many cases, IVIg should be weaned to prevent excessive resource utilisation and adverse effects.

METHODS AND MATERIALS: A cohort study was conducted examining neurology outpatients receiving regular IVIg in a 2-month period. Prespecified criteria were used to determine how many individuals were suitable for IVIg weaning. A LLM was applied to determine the level of performance with which the model could provide answers to the prespecified criteria.

RESULTS

In the inclusion period, 14 individuals were identified and four patients fulfilled the criteria for possible IVIg weaning. The total annual cost saving with IVIg weaning was conservatively estimated to be $84702.20 (1433.9 g of IVIg annually). The LLM achieved an overall classification accuracy of 78.6% (11/14) when a rule-based approach was applied to the individual criteria that it extracted from notes.

CONCLUSION

Further research is indicated to determine the frequency with which patients suitable for IVIg weaning are identified at other centres and the degree to which LLM may be able to assist with this process.