Scoring systems are widely used in clinical practice thanks to their simplicity and transparency, making them particularly well-suited for classification problems involving tabular data. However, their application to regression tasks has remained limited. This study addresses that gap with the introduction of RegScore — a sparse, interpretable scoring system purpose-built for continuous outcome prediction.
Unlike traditional scoring systems that rely on integer-valued coefficients, RegScore combines beam search with k-sparse ridge regression to relax this constraint, improving predictive flexibility without sacrificing interpretability. The authors further extend the method to a bimodal deep learning setting, integrating tabular data with medical imaging. By incorporating the classification token from the TIP (Tabular Image Pretraining) transformer, the framework generates personalized linear regression parameters and individualized scoring — enabling patient-specific predictions rather than population-level estimates.
RegScore is validated on the task of estimating mean Pulmonary Artery Pressure from tabular data, with additional refinement achieved by incorporating cardiac MRI images. Across experiments, RegScore and its personalized bimodal variants match or outperform state-of-the-art black-box models, while retaining the transparency and clinical trustworthiness that make scoring systems valuable in practice. Code is publicly available.
Autors: Michal K. Grzeszczyk, Tomasz Szczepański, Pawel Renc, Siyeop Yoon, Jerome Charton, Tomasz Trzciński, Arkadiusz Sitek
Read the article