Prevalence and characteristics of potentially avoidable unplanned readmissions: a retrospective cohort study

Inclusion and exclusion criteria

The inclusion criteria were all adult (≥18 years) unplanned admissions within 30 days of discharge from an earlier index admission (IA). If a patient had multiple readmissions, then only the first readmission was included. Patients were also excluded if they were transferred to another hospital during their IA or if they left hospital against medical advice.

Data collection and variable definitions

Readmissions to all public hospitals in South Australia were captured, as 10–15% of patients may be readmitted to non-index hospitals.³ We recorded demographic data from electronic medical records. The number of comorbidities was assessed using the Charlson comorbidity index (CCI).¹⁵ The socioeconomic status was determined using the index of relative socio-economic disadvantage (IRSD).¹⁶ The frailty status was determined using the Hospital Frailty Risk Score (HFRS), and patients with a HFRS ≥5 were classified as frail.¹⁷ Nutrition status was determined using the Malnutrition Universal Screening Tool (MUST), and patients with a MUST score >1 were classified as malnourished.¹⁸ The number of usual medications at the time of admission was recorded, and polypharmacy was defined as the use of five or more medications. We also collected data on routine laboratory tests during IA, including admission haemoglobin, white cell count, neutrophil count, lymphocyte count, platelet count, creatinine, albumin, and C-reactive protein (CRP). The neutrophil-to-lymphocyte ratio (NLR), an inflammatory haematological index increasingly recognised as a predictor of adverse clinical outcomes,^19,20 was determined by dividing the absolute neutrophil count by the absolute lymphocyte count at admission.

HOSPITAL score

The HOSPITAL score was calculated for each hospital discharge using seven variables: last available Haemoglobin before discharge (positive if <120 g/L), discharge from an Oncology service, last available Sodium concentrations (positive if <135 mmol/L), any ICD-10 AM (International Statistical Classification of Diseases and Related Health Problems, Tenth Revision, Australian Modification) coded procedure performed during hospitalisation, IA type (emergent or urgent vs elective), number of admissions in the previous 12 months, and Length of stay (LOS, positive if ≥5 days).²¹ The HOSPITAL score ranges from 0 to a maximum of 13 points, with higher scores associated with a greater risk of 30-day PAURs.²² The risk categories were defined as low (0–4), intermediate (5–6), and high (≥7).

LACE index

The LACE index²³ was determined from the following variables: LOS, Acute or urgent admission, CCI, and emergency department visits in the previous 6 months and can be used to quantify the risk of death or unplanned readmission within 30 days of hospital discharge. The LACE index can range from 0 to 19, and a score of ≥10 has been shown to indicate a high risk of readmission (expected probability of readmission of 12.2–47%) as per previous research.²⁴

Ethics

This study was deemed exempt from ethical approval by the Southern Adelaide Human Clinical Research Ethics Committee (SAH REC).

Determination of avoidability of a readmission

All readmitted patients over the 3 months from 1 July to 30 September 2022 were assessed for avoidability by two senior consultant physicians. The physicians independently reviewed the IA and readmission by thoroughly examining the electronic medical records of each patient. In case of disagreement regarding the classification of a readmission as a PAUR, a third physician independently reviewed the records to resolve the discordance. The inter-rater reliability of the classification of the two reviewers was determined using the Cohen’s kappa.

A readmission was defined as a PAUR if two conditions were fulfilled (Fig. 1) according to previously published literature.²

Fig. 1.

Criteria for identification of an avoidable readmission.

Statistical analysis

Summary descriptive statistics were generated for clinical and demographic characteristics and compared between potentially avoidable and non-avoidable unplanned readmissions. Continuous variables were assessed for normality using the Shapiro–Wilk test and then assessed for differences between groups using either an independent t-test or the rank sum test as appropriate. Categorical variables were compared using the chi-squared statistics or the Fisher’s exact test. Univariate logistic regression was used to identify potential factors associated with PAURs with variables significant at P ≤ 0.25 considered for inclusion in a backwards stepwise multivariate logistic regression model, and variables significant at P < 0.05 were retained in the final model. To account for multiple testing in the multivariable model, we applied a Bonferroni correction. Given that 14 predictors were included, the adjusted significance threshold was set at P < 0.0036 (0.05/14). This correction was used to determine statistical significance in the context of multiple comparisons.

Model discrimination was assessed using the area under the receiver operating characteristic curve. The predictive ability of the model was determined using the Brier score (Brier score <0.25 indicating good prediction).²⁵ As a sensitivity analysis, we also performed regularised logistic regression using a LASSO (least absolute shrinkage and selection operator) model²⁶ as an alternative method of variable selection to stepwise regression. To assess the robustness of the observed associations to unobserved confounding, we calculated the E-values²⁷ to estimate the magnitude of confounding required to overcome the observed associations. All tests were two-sided, and a P-value <0.05 was regarded as statistically significant. All statistical analyses were performed using STATA software version 18.0 and Python 3.10.8, utilising the scikit-learn library version 1.5.1 for machine learning and data analysis.

Avoidable unplanned readmissions

The most common reason for a PAUR was a relapse of the condition treated during the IA, which accounted for 44.4% of cases. This was followed by complications related to the disease or its treatment, observed in 22.2% patients (Fig. 3). The most frequent avoidable factor was the quality of care received during the IA in 43.8% of cases, followed by issues with post-discharge follow-up care in 30% cases (Fig. 4). The inter-rater reliability for the classification of readmission showed substantial agreement (k = 0.76, 95% confidence interval (CI) 0.63–0.85, P < 0.001).

Fig. 3.

Relationship of readmission with index admission.

Fig. 4.

Avoidable factor identified for readmission.

Comparison between avoidable and non-avoidable unplanned readmissions

Table 1 compares the characteristics of patients with potentially avoidable and non-avoidable unplanned readmissions. Patients with a PAUR were significantly older, more likely to be readmitted within a week of discharge, and had a higher prevalence of coronary artery disease (CAD) and congestive heart failure (CHF) compared to those who had a non-avoidable readmission (P < 0.05). Additionally, patients with a PAUR had significant elevations in inflammatory markers, including neutrophil count, CRP, and NLR during their IA, and lower platelet counts and albumin levels (P < 0.05) when compared to those who had a non-avoidable unplanned readmission. There were no significant differences in the HOSPITAL score, the LACE index, or other characteristics between the two groups (P > 0.05) (Tables 1, 2).

Multivariable logistic regression

In multivariable logistic regression, after applying the Bonferroni correction, CHF (adjusted odds ratio (aOR) 2.46, 95% CI 1.28–4.71, P = 0.002), a high NLR on admission (aOR 1.05, 95% CI 1.02–1.08, P < 0.001), and a low CCI (aOR 0.84, 95% CI 0.72–0.98, P = 0.001) were independently associated with PAURs (Table 3 and Fig. 5). The model demonstrated modest discriminative ability with an area under the curve of 0.68 (95% CI 0.62–0.75) (Fig. 6) and a Brier score of 0.15, indicative of accurate prediction. In sensitivity analysis, the LASSO regression identified only CHF and NLR as being significantly associated with PAURs.

Fig. 5.

Forest plot showing predictors associated with avoidable readmissions after multivariable logistic regression.

Fig. 6.

Receiver operating characteristic curve for avoidable readmissions.

E-value

The calculated E-value was 1.54 for CHF (1.17 for lower 95% CI) and 1.09 for NLR (1.03 for lower 95% CI), suggesting that unmeasured binary confounding variables with effects at least this large would be required to completely remove the observed association between these variables and PAURs.

Implications for practitioners

While improving discharge summaries is one important strategy, previous research has highlighted the broader need for a multifaceted approach to reduce PAURs. Future studies and quality improvement initiatives should consider assessing patient and family readiness for discharge, the availability and coordination of community and post-discharge support, the timing of discharge, and the use of patient education strategies such as teach-back methods.^40–42 Integrating these elements into routine discharge planning may help address modifiable factors contributing to avoidable readmissions.

Limitations

While reviewing medical records is regarded as the gold standard in judging the cause of a readmission, reviewer bias can occur as clinicians may adjudicate a complication to decompensation of a patient’s illness rather than issues with the quality of care received during the IA.⁴³ We were unable to determine the influence of functional state or cognition of patients, which have been identified as significant risk factors for readmissions.⁴⁴ This study was limited by the relatively short 3 month sampling period, which may restrict the generalisability of findings. However, this timeframe was intentionally selected to facilitate detailed, clinician-led adjudication of each readmission’s avoidability – a labour-intensive process requiring in-depth chart review and consensus. While this approach reduced the sample size compared to studies using administrative data over longer durations, it enhanced the clinical validity and specificity of the findings. Another strength of this study was capturing readmissions to all public hospitals rather than limiting the analysis to the index hospital.