Abstract
Fragility fractures due to osteoporosis are common and are associated with significant clinical, personal, and economic burden. Even after a fragility fracture, osteoporosis remains widely underdiagnosed and undertreated. Common fracture risk assessment tools, such as FRAX1 and Garvan,2 confer risk over the long term but do not provide short-term risk estimates necessary to identify very high-risk patients likely to fracture in the next 1–2 years. Furthermore, these tools utilize cross-sectional data representing a subset of all available clinical risk factors for risk prediction. Thus, these methods are generalized across patient populations and may not fully utilize patient histories commonly found in electronic health records (EHRs) that contain temporal information for thousands of unique features. The Optum® de-identified EHR dataset (2007–2018) provides an opportunity to use historical medical data to generate short-term, personalized fracture risk predictions for individual patients. We used the Optum® dataset to develop Crystal Bone, a method that applies machine learning techniques commonly used in natural language processing to the temporal nature of patient histories in order to predict fracture risk over a 1- to 2-year timeframe. Specifically, we repurposed deep-learning models typically applied to language-based prediction tasks in which the goal is to learn the meanings of words and sentences to classify them. Crystal Bone uses context-based embedding techniques to learn an equivalent “semantic” meaning of various medical events. Similar to how language models predict the next word in a given sentence or the topic of an overall document, Crystal Bone can predict that a patient’s future trajectory may contain a fracture or that the “signature” of the patient’s overall journey is similar to that of a typical fracture patient. We applied Crystal Bone to two datasets, one enriched for fracture patients and one representative of a typical hospital system. In both datasets, when predicting likelihood of fracture in the next 1–2 years, Crystal Bone had an area under the receiver operating characteristic (AUROC) score ranging from 72% to 83% on a test (hold-out) dataset. These results suggest performance similar to that of FRAX and Garvan, which have 10-year fracture risk prediction AUROC scores of 64.4% +/- 3.7%.3 Our results suggest that it is possible to use each patient’s unique medical history as it changes over time to predict patients at risk for fracture in 1–2 years. Furthermore, it is theoretically possible to integrate a model like Crystal Bone directly into an EHR system, enabling “hands-off” fracture risk prediction, which could lead to improved identification of patients at very high risk for fracture.
1Kanis JA, Osteoporos Int. 2012;23:2239–56. 2Rubin KH, J Bone Miner Res. 2013;28:1701–17. 3Leslie WD, Osteoporos Int. 2014;25:1–21.
Designing and Validating an Annotation Model of Dynamic Modality for English and Spanish: Issues and Problems