Machine learning functional impairment classification with electronic health record data

Area of Science:

• Gerontology
• Health Informatics
• Machine Learning

Background:

• Functional status is a key indicator of morbidity but is often not assessed in clinical practice.
• Electronic Health Records (EHR) offer a potential data source for assessing functional status.
• Developing scalable methods to identify functional impairment is crucial for patient care.

Purpose of the Study:

• To develop and evaluate a machine learning algorithm for identifying functional impairment using EHR data.
• To create a scalable process for detecting patients with varying levels of functional status.
• To assess the accuracy of the algorithm in differentiating normal, mild to moderate, and severe functional impairment.

Main Methods:

• A cohort of 6484 patients with functional status screening data (2018-2020) was analyzed.
• Unsupervised learning (K-means, t-SNE) classified patients into normal function (NF), mild to moderate (MFI), and severe (SFI) states.
• An Extreme Gradient Boosting model was trained on 832 EHR features to predict functional status, with SHAP analysis for feature importance.

Main Results:

• The model achieved high predictive accuracy with AUROC values of 0.92 (NF), 0.89 (MFI), and 0.87 (SFI).
• Key predictors included age, falls, hospitalizations, home health use, lab values (albumin), comorbidities (dementia, heart failure, CKD), and social determinants (alcohol use).
• The algorithm demonstrated strong performance in distinguishing between different functional status categories.

Conclusions:

• Machine learning algorithms utilizing EHR data can effectively differentiate functional status in clinical settings.
• These algorithms show promise in complementing traditional screening methods for functional status.
• Further validation could lead to population-based strategies for identifying patients requiring additional health resources due to poor functional status.