Predictive Modelling of Post-Surgery Complications using Multi-Modal Machine Learning Approaches

Adverse outcomes post-surgery remains a key predictor for extended hospital stay in association with raised patient morbidity, indicating requirement of reasonably accurate and timely risk prediction frameworks. This paper describes a new multi-modal Machine Learning architecture that combines the structured clinical notes, Radiological imaging and the Physiological Time-Series signals with a view to predicting the post-operative complications. Modality-dependent deep neural encoders MLP for clinical data, CNN for imaging, and BiLSTM for vital signs, and a risk representation of following attention-based fusion layer are employed by the system. The proposed model achieves the accuracy of 91.4% and the ROC-AUC of 0.94 on a real-world surgical dataset outperforming unimodal and baseline models. The framework shows good generalizability and clinical relevance and may be a useful decision-support system for proactive post operative care management.