Hybrid recommendation of healthcare interventions for personalised patient engagement: a leakage-free offline evaluation on synthetic electronic health records

Abstract Digital health platforms often deliver the same content to every patient, which limits engagement. Recommender systems could tailor care plans, medications and procedures to the individual, but evaluations in health settings often lack strong baselines, select hyperparameters on test data, or rely on a single random split. We built a catalogue of 300 recommendable interventions from the April 2020 Synthea sample of 1,171 synthetic patients and compared nine recommenders: random and popularity baselines, content-based filtering on clinical reason codes (CBF), CBF with data-derived item profiles (CBF-D), user-based nearest-neighbour collaborative filtering (kNN), the linear autoencoder EASE, and three weighted hybrids. All hyperparameters were tuned on validation data carved from training only. We report patient-level bootstrap confidence intervals, paired tests, random and temporal splits, ten split seeds, cohort sizes from 113 to 1,171 patients, subgroup performance and a cold-start scenario. The CBF + kNN hybrid outperformed both of its components (NDCG@10 0.692 versus 0.643 for kNN and 0.319 for CBF) with a selected content weight of 0.3. However, the tuned EASE baseline scored higher (0.754), and the best model combined CBF-D with EASE (0.787, 95% CI 0.773-0.802). Under the temporal split, every hybrid outperformed every single-signal model, including EASE. A popularity baseline reached a hit ratio of 0.682 while covering only 9.3% of the catalogue. Hybrid recommendation is a sound basis for personalised engagement, but strong baselines and leakage-free tuning are needed to measure its benefit. Validation on real records with engagement outcomes is the next step.