PRESTIGE: Prevalence estimation for interactive guesstimates

Area of Science:

• Medical Informatics
• Natural Language Processing
• Radiology

Background:

• Estimating the prevalence of radiological findings from clinical text is crucial for research and patient care.
• Traditional methods often rely on computationally intensive similarity comparisons or manual chart reviews.
• There is a need for efficient, scalable solutions for real-time prevalence estimation.

Purpose of the Study:

• To introduce PRESTIGE, a direct prediction approach for estimating radiological finding prevalence from text.
• To evaluate PRESTIGE's performance in both global and local settings.
• To assess the impact of low-rank adaptation (LoRA) on PRESTIGE's efficiency and accuracy.

Main Methods:

• PRESTIGE utilizes a direct prediction model, bypassing similarity-based comparisons for faster inference.
• The method takes a single query example to estimate prevalence in a larger dataset.
• Low-rank adaptation (LoRA) was incorporated to further optimize model size and training speed.

Main Results:

• PRESTIGE demonstrated high accuracy in predicting prevalence from radiology reports.
• The approach significantly reduced inference time and storage requirements compared to naive and approximate nearest neighbor methods.
• LoRA integration maintained strong performance while substantially lowering hardware demands, with a potential minor accuracy trade-off.

Conclusions:

• PRESTIGE provides an efficient and scalable solution for real-time, example-based prevalence estimation of radiological findings.
• The method facilitates interactive dataset exploration and streamlines research on observation frequency in large document collections.
• PRESTIGE empowers clinicians and researchers to quickly assess finding frequencies in local patient populations.