Weakly supervised treatment selection: Machine learning models for appropriate surgical planning of submandibular stones.

There is a gap in real-world clinical adoption of machine learning (ML) solutions due to the inherent uncertainty and variability in treatment outcomes. To bridge this gap, we present a novel approach to the problem of medical treatment selection using ML models and we apply it to the case of submandibular sialolithiasis treatment. The study introduces a weakly supervised learning framework which allows for the inclusion of imprecise, incomplete, or noisy ground truth data. By applying this methodology to the specific medical problem of submandibular stone treatment, we demonstrate the potential of encoding treatment outcomes as credal sets-collections of probability distributions reflecting the uncertain nature of the optimal treatment-to improve surgical planning and decision-making. We validated our model using real-world patient data, showcasing its ability to offer personalized treatment recommendations based on radiological features of submandibular stones. Our study underscores the importance of incorporating proper uncertainty management into ML for clinical practice to support clinical decision-making, by showing a promising solution to improve the treatment of sialolithiasis.