Developing and validating an electronic health record-based frailty index in pre-operative settings using machine learning

Frailty is associated with poor post-operative outcomes. However, frailty assessments in clinical practice are challenging due to the need for more resources and pragmatic complexities. We aimed to create a pre-operative frailty ascertainment using machine learning (ML) from electronic health record (EHR) data that is built on the Fried frailty phenotype. We leveraged a research database of 8,999 individuals aged 65 years and above who underwent a surgery. Healthcare providers administered a pre-operative frailty assessment using the Fried frailty phenotype. We built ML models to predict pre-operative frailty as a whole and by surgical service. We used the SHapley Additive exPlanations (SHAP) to interpret the results. Comparisons with the accumulated deficit approach were made using Pearson’s correlation coefficients and McNemar’s test. The ML model achieved an AUC of 0.74 in predicting pre-operative frailty phenotype. The ML models’ predictive power varied by surgical services, with AUC ranging from 0.63 to 0.81. SHAP showed that advanced age, anemia, chronic pulmonary disease, and literacy deficit were the most important features of pre-operative frailty. The ML model had similar performance to the accumulated deficit approach, with positive correlation (r: 0.36-0.58, P> 0.001) between frailty risk scores predicted by ML and frailty indices obtained from the accumulated deficit approach. Our approach provided new insights into the importance of specific EHR features in pre-operative frailty assessment. ML may offer an alternative method to predict frailty in pre-operative settings. Further validation on external data sources is required to generalize our ML models.