Background Guidelines recommend routine identification of frailty to provide proactive care. In the UK, the electronic Frailty Index (eFI) is built into primary care electronic health records (EHR) to categorise patients according to their frailty score. The eFI uses the ‘cumulative deficit model’, where the frailty score is the proportion of deficits (signs, symptoms, diagnoses, impairments) experienced out of a total number of deficits. The eFI may be limited by the equal weighting of deficits and may be improved through the inclusion of additional deficits such as those representing mental health. The aim of this study was to develop and validate the electronic Frailty Index 2 (eFI2) using a clinical prediction model that uses information in EHR. The deficits in the eFI2 will be weighted according to the strength of association with frailty-related outcomes.

Methods The eFI2 model was developed using linked primary and secondary care data from ‘Connected Bradford’. External validation was performed in the Secure Anonymised Linkage (SAIL) databank, containing linked EHR from patients in Wales. Patients aged ≥65 on April 1^st2018 were included with a look-back period of primary care EHR from first registration. Cox regression was performed with a composite outcome of new or increased home care package, hospital admission for fall or fracture, nursing home admission or mortality within one year. Predictors were the 36 deficits included in the original eFI and those identified in a systematic review, targeted scoping reviews and consultations with clinical practitioners. Model performance was assessed using the C-statistic and Calibration plots.

Results There were 77,809 patients (5309 (6.8%) with composite outcome) in Connected Bradford and 660,417 (56,408 (8.5%) with composite outcome) in the SAIL databank. There were 41 predictors in the model, including dementia, underweight BMI, self-harm and palliative care. Internal validation showed excellent discrimination with a bootstrapped C-statistic of 0.81. In external validation, the model had good discrimination with a C-statistic of 0.73 (95%CI: 0.73, 0.74). Calibration in the external validation was good, although there was over-prediction for the 5% at greatest risk. The one-year risk of the composite outcome was 3.4%, 5.6%, 10.3%, 23.0%, 38.7% for patients categorised as fit, pre-frail, mild, moderate and severe frailty.

Conclusion The eFI2 builds on the eFI by incorporating additional deficits and assigning weights to deficits based on the association with key outcomes. The eFI2 has robust validity for predicting frailty-related outcomes and can be implemented into primary care EHR to enable identification of patients with frailty.