A Configurable Decision-Support Engine for Sequential Clinical Workflows

Machine-learning models in decision-support systems usually require fixed inputs, but in clinical workflows information arrives sequentially, leaving the model unusable or reliant on imputation for early decisions. This paper presents a configurable decision-support engine, designed as a module of a patient digital twin, that adapts to the information available. Its central element is the configuration: an input set paired with a model and its cross-validated performance record. Configurations are versioned in a repository and assigned to each case by a Business Process Model and Notation procedure covering exact matches, similarity-based fallback for unseen inputs, and test recommendation when no estimate is reliable. The engine is specified independently of learning algorithm and pathology, with a Unified Modeling Language package diagram, a formal selection rule and seven design requirements. A prototype was assessed on a public cohort of 777 children with suspected appendicitis. With incomplete information, the engine discriminated better than a single full-feature model (area under the curve 0.81 vs. 0.73–0.76 at triage) and was better calibrated, with no loss on complete data; 41% of patients reached a high-confidence estimate before imaging, versus 14% with a static model. Managing feature-set-specific models as architectural artifacts keeps decision support useful under incomplete information.