Embedding a standardized measure of physiologic stress in electronic health records (EHR) could be instrumental to preventative care. The allostatic load index (ALI), calculated from ten biomarkers across three body systems, offers a promising snapshot of cumulative strain. The ALI can be calculated from EHR data, but many components are missing, since not all patients undergo all tests. Using statistical modeling and machine learning, EHR data for $707$ patients from a large academic health system ($17\%$ of whom were hospitalized), were used to predict in-patient hospitalization from ALI, controlling for age and sex. Various methods were evaluated to fill in information gaps for patients' missing ALI components, including summary measures combining components or using them separately. Performance was measured using various discrimination, calibration, and hybrid metrics. Discriminative performance was fair to low and calibration low to moderate for both model types. Brier scores indicated good overall accuracy across models. Summary measure approaches performed similarly with respect to most metrics. Separate component methods, especially pattern submodels, showed potential improvement over summary measures in the full sample, but their advantages did not persist under cross-validation. Most summary measure models had moderate discrimination and calibration ability. When cross-validated, separate component models appeared overfit and performed similarly to (or worse than) summary measures. Researchers may want to prioritize whichever summary measure was most interpretable or appropriate for their beliefs about missingness. Next steps include developing refined prediction models from EHR-derived ALI to improve performance and enhance clinical relevance.

翻译：暂无翻译