GenHPF: General Healthcare Predictive Framework with Multi-task Multi-source Learning

• DOI: 10.48550/arxiv.2207.09858
• Publication Date: 2022-01-01
• Authors: Kyunghoon Hur; Jungwoo Oh; Junu Kim; Jiyoun Kim; Min Jae Lee; Eunbyeol Cho; Seong-Eun Moon; Young‐Hak Kim; Louis Atallah; Edward Choi
• Keywords: Feature Engineering; Supervised Learning; Transfer of learning

Despite the remarkable progress in development of predictive models for healthcare, applying these algorithms on a large scale has been challenging. Algorithms trained particular task, based specific data formats available set medical records, tend to not generalize well other tasks or databases which fields may differ. To address this challenge, we propose General Healthcare Predictive Framework (GenHPF), is applicable any EHR with minimal preprocessing multiple prediction tasks. GenHPF resolves heterogeneity codes and schemas by converting EHRs into hierarchical textual representation while incorporating as many features possible. evaluate efficacy GenHPF, conduct multi-task learning experiments single-source multi-source settings, three publicly datasets different 12 clinically meaningful Our framework significantly outperforms baseline that utilize domain knowledge learning, improving average AUROC 1.2%P pooled 2.6%P transfer also showing comparable results when single dataset. Furthermore, demonstrate self-supervised pretraining using effective combined resulting 0.6%P improvement compared without pretraining. By eliminating need feature engineering, believe work offers solid can be leveraged speed up scaling usage healthcare.