Background: Complex patients represent a challenge in the health care system, as they are frequently associated with high health care utilization, longer hospital stays, and higher costs. Nonetheless, the characteristics of the complex patients in the hospital setting are not well known, and tools to help hospitals to identify them early in the process of care are missing. Therefore, we aim to develop a novel prediction model to identify the complex hospitalized medical patients.
Methods: We prospectively included all consecutively discharged patients from the department of medicine of a large teaching hospital between October, 2016 and February, 2017. Patients could be included only once in the study. Complexity was assessed at time of hospital discharge by the resident in charge of the patient, and defined as patients who needed more time, effort, and/or resources during the hospital stay. We tested 34 potential predictors readily available at time of admission from the electronic health records (demographics, health care utilization, diagnosis at admission, medication classes at admission, and lab values at admission). Predictors were selected using a multivariable logistic regression with backward selection. The remaining significant predictors were then used to create a scoring system. The performance of the score was assessed through its discriminatory power (area under the receiver operating curve, AUROC), its calibration, and its misclassification error.
Results: During the study period, 1,535 adult patients have been discharged. The complexity assessment was missing for 14 discharges (0.9%), and 114 (7.4%) were excluded because they were returning patients already included once in the study. The remaining population included 1,407 patients, with a proportion of complex patients of 31.8% (n=447). The mean age was 72 years old (interquartile range 57-82) , and 53.2% (n=748) were male. Most of the admissions were urgent (85.9%, n=1,167), and the median number of medications at time of admission was 8 (interquartile range 4-12). The mean Charlson comorbidity index was 2 (interquartile range 0-4). The most frequent reported components for patient complexity were the acute disease and/or comorbidities of the patient (94%), the patient characteristics (63%), the medical complications (58%), and the care coordination (53%). Nine predictors were included in the final prediction model (Table 1). The score points ranged from 0 to a maximum of 19. Using a cutoff of 8 or more points for high-risk patients for complexity, 286 patients (20%) were identified as high-risk (Table 2). The corrected AUROC was 0.67. The specificity was 89% and the negative predictive value 74%. The calibration was excellent with a perfect match between observed and expected risk in each risk group (Table 2) The misclassification error was 29%.
Conclusions: This novel prediction model including simple predictors allows to help identify at time of admission the medical patients at higher risk of need for complex care during their hospitalization. Identification of the complex patients may be valuable in order to better allocate the resources during and after the hospitalization.
.jpg)
.jpg)