Applying models of care for total hip and knee arthroplasty: External validation of predictive models to identify extended stay prior to lower-limb arthroplasty

Introduction/AimsMultiple screening tools exist for identifying patients at risk of extended stay following lower limb arthroplasty. Use of these models at other hospital sites requires verification of appropriate data coverage and evidence of validity in a new population. The aim of this study was to adapt and assess 1) data compatibility, 2) discrimination, and 3) calibration of three published models for identifying patients at risk of an extended (5+ day) stay, or those likely to stay for the target 3 or fewer days following lower limb arthroplasty. MethodsRetrospective study, utilising a randomly selected (N=200 of a total 331 available in the electronic medical record) cohort of lower-limb Total Joint Arthroplasty (TJA) patients, to externally validate an adaptation of predictive tools and regression models published by three independent groups: Winemaker et al (2015)1, Oldmeadow et al (2003)2 and Gabriel et al (2018)3. Electronic medical records of a single, medium-sized public hospital were accessed to extract data required for the models and respective predictive tools, and model characteristics (included predictors, data coding, sample sizes) were modified according to the available data. ResultsThe study cohort comprised 200 patients (60% female) at a median 70yrs of age (IQR 62-75). Approximately 58% received total knee arthroplasty (TKA) and 42% underwent total hip arthroplasty (THA). The two prediction tools and three regression models all required modifications due to data items being unavailable in the electronic records. A modification of the RAPT tool applied to 176 eligible patients resulted in sensitivity of 85.71% (95%CI 71.46-94.57) and poor specificity 32.09% (24.29-40.70), with 68% of short-stay patients classified in the high risk group. Adaptation of the second tool to 85 eligible patients resulted in unreliable estimates of sensitivity due to limited data. The three adapted regression models performed similarly well with regard to discrimination when used to predict patients staying for 5 days or longer (concordance index: Winemaker et al:, 0.79, n=198; Oldmeadow et al: 0.79, n=176), or those staying 3 days or less (Gabriel et al: 0.70, n=199). Estimates of calibration suggested the models were relatively well calibrated (spiegelhalter Z -0.01-0.29, p>0.05), although calibration plots indicated some variation remained unaccounted for, particularly with patients considered at intermediate risk. ConclusionThe three resulting regression models performed adequately in terms of discrimination and calibration for identification of patients at risk of an extended stay. However, comparison with published models was hampered by systemic issues with data compatibility. Further evaluation of such models in a specific hospital setting should incorporate improvements in data collection, and establish key thresholds for use in targeting resources to patients in need of greater support.