Predictive normative modeling of neurocognitive performance in Irish jockeys using regression and neural quantile methods

This study presents a comprehensive evaluation of regression-based methods for generating normative data for neuropsychological test scores, using retrospective cross-sectional data from Irish jockeys between 2010 and 2024. The analysis focuses on four key cognitive assessments used in concussion protocols: Digit Span Forward and Digit Span Backward (assessing working memory), Speed of Comprehension from the Speed and Language Processing battery (assessing language processing speed), and the Symbol Digit Modalities Test (assessing attention and processing speed). In addition to the widely used linear regression (LR) approach, we evaluate conditional percentile estimation using both linear quantile regression (LQR) and a neural network-based quantile regression (NNQR) framework with monotonicity constraints across quantile levels which has not been previously applied in the context of normative modeling to the best of our knowledge. Results highlight that percentile thresholds vary substantially depending on the chosen method and the covariates for corresponding test score under consideration. This variation underscores the importance of selecting an appropriate modeling approach based on the clinical scenario and the specific percentile range of interest. To make this work more viable in a practical sense, we provide an interactive dashboard that estimates normative percentiles using pre-trained LR, LQR, and NNQR models based on age, education, and sex. It automatically highlights the best-performing method based on coverage offset analysis and supports detailed report generation. While the dashboard operates on pre-trained models for ease of use, the underlying code is open-source and available via GitHub, allowing researchers and clinicians to customize it with their own datasets and model configurations for more tailored normative estimation.