MedicalRec: Transformer-based recommender picks best AI model without retraining

A team of eight researchers introduced MedicalRec, a transformer-based recommender system that predicts the best deep learning model for medical image classification tasks without needing retraining. The system relies on MedicalRec-Bench, a new public dataset parsed from 3,000 peer-reviewed papers, containing over 5,000 records of models tested on tasks like skin cancer, tumor, wound, breast cancer, and MRI classification. The dataset supports four configurations—MedicalRec I (5 features), II (9 features), III (11 features), and IV (18 features)—but suffers from significant missing values due to inconsistent reporting in source papers. MedicalRec uses a transformer architecture for item recommendations and achieved a maximum HitRate@100 of 75.5% across evaluations with 12 base models.

The core innovation is eliminating the energy-heavy trial-and-error process researchers currently use to pick models. By recommending a suitable architecture and hyperparameters upfront, MedicalRec reduces computational waste, e-waste, and carbon emissions from repeated training. The authors made both the dataset and implementation openly available on GitHub, enabling other teams to extend the system to new imaging tasks or modalities. This approach directly addresses the growing environmental cost of machine learning in healthcare—a field that demands high accuracy but is increasingly scrutinized for its carbon footprint. The paper is published on arXiv and positioned as a step toward sustainable AI medical diagnostics.