2018/08/09 by Daisy Yi Ding, Ding, Daisy Yi, Chloé Simpson +9 · 3 citations
Biochemistry, Genetics and Molecular Biology · Computer Science · Health Professions · #Artificial Intelligence in Healthcare #Biomedical Text Mining and Ontologies #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Machine Learning in Healthcare #Topic Modeling
paper · pdf · doi:10.48550/arxiv.1808.03331
openalex publication_date 2018/08/09 · openalex created_date 2022/08/04 · openalex updated_date 2026/07/28
Electronic phenotyping is the task of ascertaining whether an individual has\na medical condition of interest by analyzing their medical record and is\nfoundational in clinical informatics. Increasingly, electronic phenotyping is\nperformed via supervised learning. We investigate the effectiveness of\nmultitask learning for phenotyping using electronic health records (EHR) data.\nMultitask learning aims to improve model performance on a target task by\njointly learning additional auxiliary tasks and has been used in disparate\nareas of machine learning. However, its utility when applied to EHR data has\nnot been established, and prior work suggests that its benefits are\ninconsistent. We present experiments that elucidate when multitask learning\nwith neural nets improves performance for phenotyping using EHR data relative\nto neural nets trained for a single phenotype and to well-tuned logistic\nregression baselines. We find that multitask neural nets consistently\noutperform single-task neural nets for rare phenotypes but underperform for\nrelatively more common phenotypes. The effect size increases as more auxiliary\ntasks are added. Moreover, multitask learning reduces the sensitivity of neural\nnets to hyperparameter settings for rare phenotypes. Last, we quantify\nphenotype complexity and find that neural nets trained with or without\nmultitask learning do not improve on simple baselines unless the phenotypes are\nsufficiently complex.\n