Deep Learning With Electronic Health Records for Short-Term Fracture Risk Identification: Crystal Bone Algorithm...

Area of Science:

• Artificial Intelligence in Healthcare
• Machine Learning for Medical Prediction
• Natural Language Processing in Clinical Data Analysis

Background:

• Osteoporosis-related fractures pose a significant burden, yet high fracture risk is often underdiagnosed and undertreated.
• Current fracture risk assessment tools have limitations, including reliance on manual data entry and long-term predictions.
• There is a critical need for tools that provide short-term fracture risk estimates for timely intervention.

Purpose of the Study:

• To develop and evaluate an algorithm for identifying patients at high risk of fracture within a 1- to 2-year period.
• To overcome limitations of existing tools by focusing on short-term risk, automated data entry, and longitudinal patient data.
• To create a predictive model that aids in the early identification and management of individuals susceptible to fractures.

Main Methods:

• Developed Crystal Bone, an algorithm using machine learning and natural language processing on retrospective electronic health record data from over 1,000,000 patients.
• Applied temporal analysis of patient histories to generate short-term fracture risk predictions, analogous to language models predicting future text.
• Validated the algorithm's accuracy using a holdout set of 192,590 patients, comparing performance against baseline models and approximated human performance.

Main Results:

• The Crystal Bone algorithm accurately predicted 1- to 2-year fracture risk in patients over 50 with an AUROC of 0.81.
• The model significantly outperformed experimental baseline models (AUROC 0.67).
• It correctly identified 70% of at-risk patients who had not received preventative interventions, highlighting its potential for clinical application.

Conclusions:

• A patient's evolving medical history can be effectively utilized to predict short-term fracture risk.
• Implementing and validating such a tool in healthcare systems can enable automated, widespread fracture risk prediction.
• This technology may significantly improve the identification of patients at very high risk for fracture, facilitating proactive care.