Predicting and Interpreting Patient Journeys Through Machine Learning and Process Mining

Abstract

Abstract The increasing availability of Electronic Health Records (EHR) presents a unique opportunity to model long-term patient journeys and disease trajectories. This research investigates the use of machine learning, particularly transformer-based architectures and process mining techniques, to interpret and predict patient pathways over extended periods. Using the MIMIC III dataset, this study develops and compares models for forecasting future hospital admissions and diagnostic outcomes using International Classification of Diseases (ICD) codes. The Temporal Fusion Transformer (TFT) demonstrated the strongest performance, achieving a Micro AUROC of 0.86 for multi-label ICD code prediction and a MAE of 43 days for next visit date prediction. In parallel, process mining was employed to visualise patient journeys and identify common diagnostic patterns, offering interpretability often missing from deep learning models. The results show that combining temporal prediction with explainable visualisations can support more informed, long-term clinical decision-making. However, limitations such as data sparsity and class imbalance among diagnostic categories impacted model performance, particularly for rare conditions. Additionally, predictions were restricted to a single future step, limiting their applicability for full trajectory forecasting. Future research could explore multi-step pathway prediction and the deployment of domain-specific fine-tuned models. Expanding visualisation capabilities to include dynamic filtering and root cause analysis may also further enhance clinical interpretability. This work contributes toward building tools that are both data-driven and practical for real-world healthcare planning.