The 30-day post-discharge readmission rate is a quality indicator that may reflect suboptimal care. The computerised algorithm SQLape® can retrospectively identify a potentially avoidable readmission (PARA) with high sensitivity and specificity. We retrospectively analysed the hospital stays of patients readmitted to the Department of Internal Medicine of the CHUV (Centre Hospitalier Universitaire Vaudois) in order to quantify the proportion of PARAs and derive a risk prediction model.
All hospitalisations between January 2009 and December 2011 in our division of general internal medicine were analysed. Readmissions within 30 days of discharge were categorised using SQLape®. The impact on PARAs was tested for various clinical and nonclinical factors. The performance of the developed model was compared with the well-validated LACE and HOSPITAL scores.
From a total of 11 074 hospital stays, 777 (7%) were followed with PARA within 30 days. By analysing a group of 6729 eligible stays, defined in particular by the patients' returning to their place of residence (home or residential care centre), we identified the following risk factors: ≥1 hospitalisation in the year preceding index admission, Charlson score >1, active cancer, hyponatraemia, length of stay >11 days, prescription of ≥15 different medications during the stay. These variables were used to derive a risk prediction model for PARA with a good discriminatory power (C-statistic 0.70) and calibration (p = 0.69). Patients were then classified as low (16.4%), intermediate (49.4%) or high (34.2%) risk of PARA. The estimated risk of PARA for each category was 3.5%, 8.7% and 19.6%, respectively. The LACE and the HOSPITAL scores were significantly correlated with the PARA risk. The discriminatory power of the LACE (C-statistic 0.61) and the HOSPITAL (C-statistic 0.54) were lower than our model.
Our model identifies patients at high risk of 30-day PARA with a good performance. It could be used to target transition of care interventions. Nevertheless, this model should be validated on more data and could be improved with additional parameters. Our results highlight the difficulty to generalise one model in the context of different healthcare systems.