Prediction of 30-day unplanned hospital readmission among elderly patients using patient and ward level factors

Abstract

Hospital readmission among elderly patients in sub-Saharan Africa, where health systems are constrained, is a critical yet poorly understood challenge. Although patient-level risk factors are established, the role of modifiable ward-level organizational factors remains largely unexplored in this context. This study aimed to investigate the patient- and ward-level determinants of 30-day unplanned readmission among elderly Ethiopian patients. A prospective study was conducted from January 2023 to February 2025; involving 504 patients aged ≥ 65 years. Data on patient demographics, clinical characteristics, and ward-level factors were collected from medical records. A GLMM was employed to identify predictors of readmission, presenting adjusted odds ratios (aORs) with 95% confidence intervals (CIs). The overall 30-day readmission rate was 44.8%. The GLMM revealed significant ward-level clustering, with an intra-class correlation coefficient (ICC) of 0.186, indicating that 18.6% of the variation in readmission odds was attributable to differences between wards. Significant patient-level predictors included a higher Charlson Comorbidity Index (aOR 1.27, 95% CI 1.13–1.42), more previous hospitalizations (aOR 1.24, 95% CI 1.08–1.43), longer length of stay (aOR 1.05, 95% CI 1.01–1.09), and cognitive impairment (aOR 1.89, 95% CI 1.21–2.94). Being discharged from a ward with a standardized discharge protocol was associated with a 45% lower odd of readmission (aOR 0.55, 95% CI 0.32–0.94). A significant cross-level interaction showed that this protective effect was strongest for patients with cognitive impairment. In this setting, readmission is driven by a combination of patient-level clinical vulnerability and ward-level organizational capacity. The substantial ward-level variation highlights a key opportunity for intervention. The implementation of discharge protocols represents a powerful, low-cost strategy to significantly reduce readmissions, particularly for the most vulnerable elderly patients with cognitive impairment, thereby improving the quality of care in resource-limited hospitals.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-28060-z.

Introduction

Hospital readmission among elderly patients is a critical public health and clinical concern, often indicating suboptimal quality of care, increased health risks, and higher economic burden on health systems. Globally, 30-day readmission rates in older adults range from 5 to 35%, depending on population characteristics, comorbidities, and healthcare delivery systems^1,2. Reducing preventable readmissions has become a key performance measure in high-income countries, directly linked to hospital reimbursement and patient safety standards³. Predictive models such as the LACE index have been developed to stratify patients by risk of readmission, relying primarily on patient-level factors like comorbidity, length of stay, and prior healthcare use⁴. However, the direct application of these models in low-resource settings is limited, as they fail to capture local epidemiology, constrained resources, and organizational differences⁵.

In sub-Saharan Africa, and Ethiopia in particular, the challenge is compounded by limited hospital resources, shortages of trained health professionals, and high burdens of chronic diseases such as hypertension, diabetes, and heart failure^6,7. Older adults (aged 65 years and above) are particularly vulnerable to frequent readmissions due to multimorbidity, frailty, functional disability, and inadequate post-discharge care. However, evidence on the determinants of readmissions in this age group is limited in Ethiopia, with most available studies focusing on single diseases (e.g., heart failure, tuberculosis) rather than multilevel predictors across patient and hospital systems^8,9. The limited attention to ward- and system-level factors is a critical gap in the literature. Referral hospitals, such as Debre Markos, frequently experience high patient volumes, limited bed capacity, shortages of essential medicines, and variability in discharge planning protocols. Although these factors may influence readmission risk but remain underexplored^10,11.

Traditional statistical approaches, such as logistic regression, have been widely used to predict readmission. However, these models assume independence of observations, which is inappropriate when patients are nested within wards, leading to underestimated standard errors and biased conclusions¹². Generalized Linear Mixed Models (GLMMs) offer a robust alternative by simultaneously analyzing both patient- and ward-level predictors while accounting for intra-cluster correlation^13,14 . Although such multilevel approaches are increasingly applied in high-income contexts, their use in Ethiopian hospital research remains scarce.

This study advances prior Ethiopian work, such as that of^15,16, by integrating organizational (ward-level) predictors using a multilevel design. Previous Ethiopian studies have primarily focused on patient-level characteristics hospital contextual factors. Similarly, studies from Asia, such as^17,18 in Indonesia, applied single-level regression models but did not quantify the ward-level variation’s contribution to readmissions. Our findings provide novel evidence on the influence of ward-level factors on elderly readmissions in a low-resource setting by explicitly modeling this structure using a generalized linear mixed model (GLMM). These insights are transferable to other hospitals in sub-Saharan Africa and similar low- and middle-income contexts facing staff shortages, inconsistent discharge planning, and limited post-discharge follow-up systems.

Therefore, this study aims to investigate the patient- and ward-level determinants of 30-day unplanned hospital readmission among elderly patients admitted to Debre Markos referral hospital, Ethiopia, using a generalized linear mixed model (GLMM) framework. By integrating multilevel factors, this study seeks to quantify the contribution of ward context to readmission risk and generate evidence to inform strategies for reducing avoidable readmissions, strengthening geriatric care, and improving hospital performance in resource-limited settings.

Methods and materials

Study materials and settings

This was a prospective cohort study conducted at Debre Markos Referral Hospital from January 2023 to February 2025, and all analyses were based on the finalized dataset collected within that period. The study utilized secondary data from the hospital’s medical records, surgical logs, and Health Management Information System (HMIS). All elderly patients (aged ≥ 65 years) admitted to the medical and surgical wards during the study period were identified at admission and constituted the inception cohort. The patients were then prospectively followed for the occurrence of the outcome (30-day unplanned readmission). Data on patient demographics, clinical characteristics, hospital stay, discharge diagnoses, and ward-level factors (e.g., bed capacity, nurse-to-patient ratio, discharge planning practices) were extracted prospectively according to a pre-defined protocol using a structured data extraction checklist. A generalized linear mixed model (GLMM) was employed to identify predictors of readmission.

Eligibility criteria

Patients were eligible for inclusion if they were discharged alive and had complete medical records containing essential information on demographics, comorbidities, hospital stay, and discharge diagnoses. Patients were excluded from the analysis if they were transferred to another acute care facility upon discharge or were lost to follow-up, meaning their 30-day readmission status could not be ascertained.

Outcome measurement and follow-up procedure

The primary outcome was 30-day unplanned hospital readmission, defined as any subsequent admission to Debre Markos Referral Hospital for an unplanned reason within 30 days of discharge from the index admission. Readmissions were identified through active monitoring of the hospital’s Health Management Information System (HMIS) and patient logbooks. A readmission was classified as ‘unplanned’ if it was not a pre-scheduled appointment for a procedure or chemotherapy, following established definitions from the LACE index study⁴. The planned readmissions were excluded from the outcome.

Our team conducted a structured telephone follow-up for all patients at 30 days post-discharge to ensure complete follow-up. For patients lost to phone contact, we cross-referenced regional HMIS databases to check for readmissions to other public health facilities in the zone. Patients for whom no readmission data could be confirmed (n = 12) were considered lost to follow-up and excluded from the analysis, as per our exclusion criteria. Death within the 30-day period was a competing event. Sensitivity analysis was performed using a competing-risks regression model (Fine-Gray sub-distribution hazards model) to assess the potential bias introduced by death as a competing risk. The results of this analysis were consistent with those from the primary GLMM, confirming the robustness of our findings (Supplementary Table S1).

Sample size and sampling technique

A consecutive sampling technique was used to include all eligible elderly patients admitted between January 2023 and February 2025, resulting in a total sample size of 504 patients.

Data structure and clustering

The 504 patients were hierarchically nested within 20 hospital wards. The distribution of patients across these clusters was analyzed to inform the multilevel modeling approach. The ward-level cluster sizes ranged from 11 to 35, with a mean of 25.2 and a median of 26 patients per ward, indicating a reasonably balanced data structure for subsequent generalized linear mixed model (GLMM) analysis.

Data collection tools and procedures

Data were extracted from the hospital’s medical records, surgical logs, and Health Management Information System (HMIS) using a structured data extraction checklist developed by the research team. The checklist captured information on patient demographics, comorbidities, hospital stay characteristics, discharge diagnoses, and ward-level factors such as bed capacity, nurse-to-patient ratio, and discharge planning practices. Trained health professionals familiar with hospital records reviewed and cross-checked the data to ensure accuracy and completeness, under the regular supervision of the principal investigator, who resolved any inconsistencies. Double data entry and periodic audits were performed to minimize errors, and statistical analyses were conducted using SAS version 9.4.

Quality of data

A structured data extraction checklist was used to ensure the quality and reliability of the collected data, and all data collectors were trained health professionals familiar with hospital records. Data were cross-checked for completeness and consistency, and the principal investigator conducted regular supervision to identify and resolve discrepancies. Double-entry of data and periodic audits were performed to minimize entry errors. Any missing or inconsistent values were reviewed and verified against the original records before analysis, ensuring a high level of data integrity for the study.

Variables included in the current investigation

Both dependent and independent variables were included in the study. The dependent variable was 30-day unplanned hospital readmission among the elderly patients. Independent variables were grouped into patient-level and ward-level factors. Patient-level variables included age, sex, marital status, comorbidities, Charlson Comorbidity Index, previous hospitalizations, length of stay, primary diagnosis, polypharmacy, functional dependence, cognitive impairment, nutritional status, smoking status, alcohol use, living situation, and socioeconomic status. Ward-level variables included nurse-to-patient ratio (continuous, calculated as nurses per shift divided by occupied beds), physician coverage, staff experience, ward type (medical/surgical), bed occupancy (continuous, percentage of beds occupied), availability of specialized care, standardized discharge protocols (binary: present if a formal checklist for medication reconciliation, education, and follow-up was used), readmission-prevention programs (binary: present if a structured program with education and post-discharge phone follow-up was in place), follow-up policies, ward location, and infection control practices. These variables were selected to assess both individual and organizational determinants of 30-day readmission using a generalized linear mixed model (GLMM) framework.

About the model

A generalized linear mixed model (GLMM) was employed to examine predictors of 30-day unplanned hospital readmission among elderly patients, which extends generalized linear models by incorporating both fixed effects (patient- and ward-level predictors) and random effects (unobserved variation between wards)^13,14. This approach accounts for the hierarchical structure of the data, where patients are nested within wards, making it suitable for binary outcomes like readmission. Mathematically, the model can be expressed as:

1

where, is the readmission status of the patient in the ward , ​ represents the fixed-effect covariate, ​ is the intercept, are fixed-effect coefficients, and is the random effect capturing ward-level variability. The probability of readmission is then given by:

2

The selection of fixed-effect predictors was primarily guided by clinical plausibility and prior evidence from readmission literature, rather than by purely statistical criteria. A core set of patient-level variables (Charlson Comorbidity Index, previous hospitalizations, length of stay, and cognitive impairment) and the key ward-level factor (standardized discharge protocol) were pre-specified for inclusion in the final model based on their established importance. A random-intercept GLMM was specified. Random slopes were not included due to the limited number of wards (n = 20), which could compromise model stability and interpretability. We assessed multicollinearity among all fixed-effect predictors by calculating Variance Inflation Factors (VIFs); all VIFs were below 2.5, indicating no concerning collinearity. Model fit statistics (AIC/BIC) were then used to compare the fit of nested models that included different combinations of other plausible ward-level factors.

Maximum likelihood estimation with adaptive Gaussian quadrature was used to estimate parameters, using 10 quadrature points to ensure estimation stability. The Model convergence was verified by checking that all gradient components were near zero.

Model selection included clinically relevant patient-level predictors (e.g., comorbidity, prior hospitalizations, functional status) and ward-level factors (e.g., nurse-to-patient ratio, discharge protocol, bed occupancy) based on statistical significance, AIC/BIC model fit criteria, and clinical plausibility¹⁹. Parameters were estimated using maximum likelihood estimation with adaptive Gaussian quadrature, which accounts for clustering within wards and provides unbiased estimates for both fixed and random effects²⁰. The model converged successfully, as confirmed by the SAS convergence criteria (GCONV < 1e-8). Results are presented as odds ratios (ORs) with 95% confidence intervals to quantify the effect of patient- and ward-level variables on readmission risk, and all analyses were performed using SAS 9.4.

Results

The overall 30-day readmission rate was 44.8%. The distribution of continuous patient-level variables for the entire cohort and stratified by 30-day readmission status is presented in Table 1. The data higher acuity and complexity profile among readmitted patients. Those who were readmitted had a significantly greater burden of chronic illness, as evidenced by a mean Charlson Comorbidity Index that was 0.7 points higher than those not readmitted (4.4 vs. 3.7, p < 0.001). They were also more frequent users of hospital care, with an average of 0.5 more previous hospitalizations in the past year (2.3 vs. 1.8, p < 0.001). The index admission for readmitted patients was also notably longer by an average of 1.4 days (10.6 vs. 9.2 days, p = 0.004), suggesting more severe illness or complications during that stay. Finally, readmitted patients had significantly lower functional status scores (63.2 vs. 67.9, p = 0.002), indicating greater dependence on others for activities of daily living, which is a known risk factor for poor post-discharge outcomes. No significant difference was found in age or the number of medications prescribed at discharge.

Table 1.

Continuous baseline characteristics (N = 504).

Variable	Overall (N = 504)	Not Readmitted (n = 278)	Readmitted (n = 226)	p-Value
Age (years)	77.8 ± 7.2	77.5 ± 7.4	78.2 ± 6.9	0.265
Charlson comorbidity index	4.0 ± 1.9	3.7 ± 1.8	4.4 ± 2.0	< 0.001*
Previous hospitalizations	2.0 ± 1.5	1.8 ± 1.4	2.3 ± 1.6	< 0.001*
Length of stay (days)	9.8 ± 5.6	9.2 ± 5.3	10.6 ± 5.9	0.004*
Number of medications	5.1 ± 2.4	5.0 ± 2.3	5.3 ± 2.5	0.178
Functional status (Barthel Index)	65.8 ± 17.2	67.9 ± 16.8	63.2 ± 17.3	0.002*

Data presented as Mean ± Standard Deviation.P-value from independent samples t-test.

*Indicates statistical significance at p < 0.05.

Table 2 compares the categorical demographic and clinical characteristics of the two groups. The most striking finding is the significant difference in cognitive status. Over a third of readmitted patients (36.3%) had a documented cognitive impairment compared to just over a quarter (25.5%) in the non-readmitted group (p = 0.010). This suggests cognitive impairment may be a critical vulnerability in the post-discharge period. Notably, there were no significant differences in readmission rates based on sex, marital status, primary diagnostic category, nutritional status, or living situation. This indicates that these factors, while potentially important, may not be primary independent drivers of readmission risk in this population when considered in isolation.

Table 2.

Baseline characteristics categorical variables (N = 504).

Variable	Category	Overall n (%)	Not readmitted n (%)	Readmitted n (%)	p-Value
Sex	Male	248 (49.2)	133 (47.8)	115 (50.9)	0.502
	Female	256 (50.8)	145 (52.2)	111 (49.1)
Marital status	Married	298 (59.1)	168 (60.4)	130 (57.5)	0.512
	Single	138 (27.4)	72 (25.9)	66 (29.2)
	Other	68 (13.5)	38 (13.7)	30 (13.3)
Primary diagnosis	Heart Failure	145 (28.8)	75 (27.0)	70 (31.0)	0.128
	Diabetes	121 (24.0)	72 (25.9)	49 (21.7)
	COPD	92 (18.3)	44 (15.8)	48 (21.2)
	Other	146 (29.0)	87 (31.3)	59 (26.1)
Cognitive impairment	No	351 (69.6)	207 (74.5)	144 (63.7)	0.010*
	Yes	153 (30.4)	71 (25.5)	82 (36.3)
Nutritional status	Underweight	98 (19.4)	49 (17.6)	49 (21.7)	0.401
	Normal	268 (53.2)	153 (55.0)	115 (50.9)
	Overweight	138 (27.4)	76 (27.3)	62 (27.4)
Living situation	Alone	142 (28.2)	75 (27.0)	67 (29.6)	0.771
	With Family	318 (63.1)	178 (64.0)	140 (61.9)
	Nursing Home	44 (8.7)	25 (9.0)	19 (8.4)

p-Value from Chi-square test.

* indicates statistical significance at p < 0.05.

Table 3 presents the results of a standard multivariate logistic regression model containing only patient-level factors. The model identifies key clinical risk factors: each one-point increase in the Charlson Comorbidity Index was associated with a 32% increase in the odds of readmission (adjusted odds ratio [aOR] = 1.32 95% CI 1.19–1.47). Each previous hospitalization increased the odds by 28% (aOR = 1.28, 95% CI 1.13–1.45). The presence of cognitive impairment nearly doubled the odds of readmission (aOR = 1.81, 95% CI 1.20–2.73). Each additional day of the index length of stay was associated with a 5% increase in odds (aOR = 1.05, 95% CI 1.01–1.09). This model, however, assumes all patient observations are independent and does not account for the potential clustering of outcomes within hospital wards, which may bias the standard errors.

Table 3.

Patient-level predictors of 30-day readmission (Model 1: fixed effects only).

Predictor	aOR	95% CI	p-Value
Age (per year)	1.02	0.99–1.05	0.188
Sex (Female vs. Male)	0.89	0.62–1.29	0.539
Charlson Index (per point)	1.32	1.19–1.47	< 0.001*
Prev. Hosp. (per admission)	1.28	1.13–1.45	< 0.001*
Length of Stay (per day)	1.05	1.01–1.09	0.011*
Cognitive Impairment (Yes)	1.81	1.20–2.73	0.005*
Functional Status (per point)	0.99	0.98–1.00	0.115

aOR Adjusted Odds Ratio; CI Confidence Interval. Model 1 is a standard logistic regression.

*Indicates statistical significance at p < 0.05.

Table 4 displays the results of the generalized linear mixed model (GLMM), which accounts for the clustering of patients within wards by including a random intercept for Ward ID. After this adjustment, the patient-level risk factors remained highly significant, although their effect sizes were slightly attenuated (e.g., CCI aOR = 1.27), confirming their robust independent effects. The major advance of this model is the identification of a significant ward-level protective association. Patients discharged from wards that had implemented a standardized discharge protocol had a 45% reduction in the odds of readmission (aOR = 0.55, 95% CI 0.32–0.94) compared to those from wards using ad hoc protocols. In contrast, a higher nurse-to-patient ratio was associated with a non-significant, modest reduction in readmission odds (aOR = 0.85 per 1.0 unit increase, 95% CI 0.50–1.44). This highlights that hospital systems and processes are modifiable factors associated with patient outcomes, independent of patient clinical severity.

Table 4.

Patient- and Ward-level predictors of 30-day readmission (Model 2: GLMM).

Predictor	aOR	95% CI	p-Value
Patient-level
Age (per year)	1.01	0.98–1.04	0.441
Sex (Female vs. Male)	0.87	0.57–1.33	0.520
Charlson Index (per point)	1.27	1.13–1.42	< 0.001*
Prev. Hosp. (per admission)	1.24	1.08–1.43	0.003*
Length of Stay (per day)	1.05	1.01–1.09	0.028*
Cognitive Impairment (Yes)	1.89	1.21–2.94	0.005*
Functional Status (per point)	0.99	0.98–1.01	0.301
Ward-Level
Discharge Protocol (Std. vs. Ad hoc)	0.55	0.32–0.94	0.029*
Nurse-Patient Ratio (per 1.0 unit)	0.85	0.50–1.44	0.539
Readmission Program (Yes vs. No)	1.25	0.70–2.24	0.449

aOR Adjusted Odds Ratio; CI Confidence interval. Model 2 is a GLMM with Ward ID as a random intercept.

*Indicates statistical significance at p < 0.05.

Table 5, The final model introduces a cross-level interaction to test if the effect of the ward-level factor (standardized discharge protocol) differs based on patient-level vulnerability (cognitive impairment). The significant interaction term (aOR = 0.48, p = 0.040) indicates that it does. The protective effect of a standardized discharge protocol is significantly stronger for patients with cognitive impairment. Patients without cognitive impairment, having a standardized protocol reduces the odds of readmission. But for a patient with cognitive impairment, the protective effect is effectively doubled. This suggests that standardized protocols, which likely include medication reconciliation, patient education, and scheduled follow-up, are exceptionally effective at mitigating the specific risks such as confusion, poor self-management, and missed appointments that lead to readmission in this vulnerable subgroup.

Table 5.

Full model with cross-level interactions (Model 3: GLMM).

Predictor	aOR	95% CI	P-value
Main effects
Charlson Index (per point)	1.26	1.12–1.42	< 0.001*
Cognitive Impairment (Yes)	1.92	1.23–3.00	0.004*
Discharge Protocol (Std.)	0.54	0.32–0.92	0.023*
Cross-level interaction
Cognitive Impairment * Discharge Protocol	0.48	0.24–0.97	0.040*

*aOR Adjusted odds ratio; CI Confidence interval.

To illustrate the clinical magnitude of this interaction, we present the model-predicted probabilities of readmission for representative patient profiles (Supplementary Table S2). The results show that the implementation of a standardized discharge protocol was associated with a substantially greater absolute risk reduction for patients with cognitive impairment (26.3%) compared to those without (12.7%).

Table 6 compares the statistical fit of the three nested models. The key evidence supporting the use of a GLMM is the notable random intercept variance (τ²) in Models 2 and 3. The Intra-class Correlation Coefficient (ICC) of 0.186 (95% CI 0.075–0.385) for Model 2 indicates that 18.6% of the total variability in the odds of readmission is attributable to systematic differences between wards. The 95% confidence interval for the ICC, while wide reflecting the limited number of clusters, confirms that the ward-level clustering is a non-negligible source of variation that must be accounted for, as the lower bound of the interval (7.5%) still represents a meaningful amount of clustering. The superior fit of the multilevel models (Models 2 and 3) over the standard regression (Model 1) is confirmed by their substantially lower Akaike (AIC) and Bayesian (BIC) Information Criterion values. Model 3, which includes the cross-level interaction, has the lowest AIC, indicating it is the best-fitting and most parsimonious model given the data, providing strong statistical justification for the interaction interpretation.

Table 6.

Random effects and model fit statistics.

Statistic	Model 1 (Logistic)	Model 2 (GLMM)	Model 3 (GLMM + Interaction)
Random effects variance (τ2)	–	0.75	0.71
SE (τ2)	–	0.32	0.30
ICC	–	0.186	0.178
ICC (95% CI)	–	(0.075–0.385)	(0.071–0.378)
Log Likelihood	− 305.2	− 283.1	− 281.5
AIC	626.4	592.1	589.0
BIC	662.8	654.3	655.6
N (Wards)	–	20	20

ICC Intraclass Correlation Coefficient; AIC Akaike Information Criterion; BIC Bayesian Information Criterion.

The final GLMM (Model 3) demonstrated good predictive performance. The discriminative ability of the model, measured by a conditional C-statistic (AUC) calculated using the method by²¹, was 0.78, indicating a good ability to distinguish between patients who were readmitted and those who were not. The model calibration was acceptable, as assessed by a Hosmer–Lemeshow-type goodness-of-fit test for GLMMs as described by ²² (p = 0.12), suggesting the predicted probabilities of readmission align well with the observed outcomes.

Figure 1 visualizes the substantial heterogeneity in hospital readmission performance across different wards identified by a generalized linear mixed model. The ordered forest plot reveals a striking range of ward-specific, risk-adjusted readmission probabilities, from a best-performing ward with a rate of 12.9% to a worst-performing ward with 79.2%, indicating that the specific ward to which a patient is admitted profoundly influences their outcome. The intra-class correlation coefficient (ICC) of 0.186 quantifies this finding, indicating that 18.6% of the total variance in readmission rates is attributable to systematic, ward-level differences rather than patient characteristics alone, underscoring the critical importance of targeting quality improvement initiatives at the organizational level.

Fig. 1.

Variation in risk-adjusted readmission probabilities across hospital wards.

Discussion

This study aimed to identify patient- and ward-level determinants of 30-day unplanned readmission among elderly patients at a tertiary referral hospital in Ethiopia using a generalized linear mixed model (GLMM) framework. The key findings reveal that readmission is a multifactorial event driven by a complex interplay between patient-level clinical vulnerabilities and modifiable ward-level organizational factors.

The observed 30-day readmission rate of 44.8% is substantially higher than rates typically reported in high-income countries, where readmission among older adults generally ranges from 10 to 25%, depending on the condition and healthcare system^1,2. This likely reflects the synergistic effect of several factors inherent to our resource-constrained setting. First, the study cohort itself was at very high risk, as evidenced by the high mean Charlson Comorbidity Index and the strong association of cognitive impairment with readmission, a finding consistent with the high burden of multimorbidity among the Ethiopian elderly⁷. Second, structural factors play a critical role: there is a severe lack of community-based geriatric care and post-discharge support services, a common system-level challenge in sub-Saharan Africa that shifts the entire burden of complication management back to the tertiary hospital^6,7. Third, patients often present with advanced disease due to delays in seeking care, leading to index admissions that are more complex and difficult to stabilize fully⁶. Finally, as the primary referral center for a large population, our hospital is likely to capture nearly all returning patients, whereas in more decentralized systems, readmissions may be dispersed across multiple facilities. Therefore, while high, this rate may provide a more accurate estimate of the true burden of readmissions among vulnerable elderly populations in similar low-resource contexts where active prospective follow-up is employed.

First, our study confirms the profound influence of patient-level factors established in the global literature. The strong, independent association between a higher Charlson Comorbidity Index and increased readmission risk is well-documented and reflects the heightened physiological vulnerability and complex care needs of patients with multimorbidity^1,23. Similarly, a history of previous hospitalization is a robust marker of illness chronicity and severity, often signifying an underlying high-risk state that predisposes patients to readmission, a finding consistent with models like the LACE index^4,5. Most notably, cognitive impairment emerged as a powerful predictor, nearly doubling the odds of readmission. This aligns with a growing body of evidence highlighting cognitive dysfunction as a critical, yet often overlooked, risk factor that impairs a patient’s ability to perform self-care, adhere to medication, and recognize worsening symptoms after discharge^24,25.

Our findings confirm established patient-level risk factors, including a higher Charlson Comorbidity Index, prior hospitalizations, and notably, cognitive impairment, which nearly doubled the odds of readmission^{1,4,5,23,24,26}. These factors reflect a profile of high clinical vulnerability and complex care needs that are difficult to manage in a resource-constrained system.

The most significant contribution of this study lies in its multilevel approach. The finding that nearly 19% of the variation in readmission odds was attributable to differences between wards (ICC = 0.186) is a powerful indictment of the lack of clustering in traditional statistical methods. It underscores that a patient’s risk is not solely a function of their diagnosis but is also profoundly shaped by the context of the ward in which they are treated. This finding is consistent with studies in other low-resource settings that have identified system-level inefficiencies as major drivers of poor patient outcomes^7,27.

Within this ward-level context, we identified a potent protective factor: the implementation of a standardized discharge protocol. Patients discharged from wards with such protocols had 45% lower odds of readmission. This finding is critical because it points to a concrete, modifiable factor. The observed association may be explained if standardized protocols mitigate readmission by reducing errors and improving transition quality through structured steps, including medication reconciliation, patient and family education, and scheduling follow-up appointments before discharge^28,29. This is particularly salient in resource-constrained hospitals like ours, where investigating such low-cost, high-impact factors are most needed. Implementing standardized discharge protocols in Ethiopian hospitals is a feasible and cost-effective strategy. This can be achieved through phased integration within existing Ministry of Health quality-improvement frameworks, utilizing checklist-based workflows and focused staff training.

The depth of our analysis is further demonstrated by the significant cross-level interaction, which revealed that a standardized discharge protocol had a significantly stronger protective effect for patients with cognitive impairment. This suggests that these protocols are not just only beneficial but are essential for the most vulnerable patients. This association may be explained if these protocols compensate for the deficits by providing clear instructions and structured support to both patients and their caregivers, thereby closing a critical safety gap^26,30.

Our findings should be interpreted in light of certain limitations. First, the study was conducted in a single referral hospital, which may affect the generalizability of findings to primary hospitals or other regions. While conducted in a single center, Debre Markos Referral Hospital shares many systemic characteristics with other public referral hospitals in Ethiopia, such as comparable patient caseloads, similar constraints in human resources and bed capacity, and a high burden of non-communicable diseases among the elderly^6,7,31. These shared challenges suggest that the identified patient-level risk factors (e.g., comorbidity, cognitive impairment) are likely to be highly relevant in similar settings. Furthermore, the organizational factors we studied, such as variability in discharge protocols and nurse-to-patient ratios, are common across the Ethiopian hospital system^11,32. However, hospitals with significantly different resource levels, staffing models, or established quality improvement programs might observe different magnitudes of ward-level effects. Therefore, while the specific predictors are likely transferable, the exact estimates may vary. Future multi-center studies are needed to validate our model’s performance across diverse Ethiopian healthcare contexts.

Second, the use of secondary data limited our ability to assess other potentially important variables, such as quality of social support and health literacy. Furthermore, potential confounders, such as the quality of social support and level of caregiver involvement, were not captured in our data and may influence readmission risk. Furthermore, the observational nature of the study precludes definitive causal conclusions.

Third, while the GLMM was the most appropriate model for our hierarchically structured data, the number of clusters (20 wards) is at the lower boundary for robust random-effects estimation. Although the distribution of patients across wards was reasonably balanced, the limited number of wards may have affected the generalizability of our ward-level findings and the precision of the estimated between-ward variance.

Finally, residual confounding (e.g., by unmeasured factors such as organizational culture or subtle differences in physician practice) cannot be ruled out despite adjusting for numerous patient- and ward-level confounders. Therefore, the associations we report should be interpreted as robust statistical relationships rather than as definitive causal effects.

Despite these limitations, this study has considerable strengths. Its prospective design, relatively large sample size, and high-quality data collection minimize information bias. The study’s analytic rigor was strengthened by the a priori specification of predictors and by assessing model stability through information-theoretic fit criteria, reducing the likelihood of over-fitting and ensuring that observed associations reflect meaningful clinical relationships. Most importantly, the use of a GLMM is a key methodological strength that allowed for the valid estimation of both individual and organizational effects, providing a more accurate and nuanced understanding of readmission drivers than previous studies in similar settings that used conventional regression³³.

Conclusion and recommendations

Conclusion

In conclusion, this study demonstrates that unplanned 30-day readmission among elderly patients is a multifactorial event that stems from a complex interplay between intrinsic patient vulnerabilities and extrinsic ward-level systems. The use of a multilevel model revealed that a significant portion (18.6%) of the variability in readmission risk is attributable to differences between wards, underscoring the critical limitations of analytical approaches that ignore this clustering. The findings confirm that a higher comorbidity burden, a history of prior hospitalization, and the presence of cognitive impairment are powerful patient-level predictors of readmission. However, the analysis moves beyond identifying high-risk patients to reveal a strong association between a modifiable ward-level factor, the implementation of a standardized discharge protocol, and a 45% lower odds of readmission. Furthermore, the significant protective interaction between cognitive impairment and standardized protocols indicates that these systems are particularly vital for safeguarding the most vulnerable patients. Therefore, the high rate of readmission is not an inevitable outcome but is substantially associated with the quality and standardization of hospital care processes. A potential strategy for this challenge would be a dual-focused approach that targets both the clinical complexity of the patient and the organizational context of the care they receive.

Recommendation

Based on these findings, a series of actionable recommendations are proposed. Firstly, hospital administrators and clinical leaders should prioritize the development, mandatory implementation, and continuous auditing of standardized, multidisciplinary discharge protocols across all medical and surgical wards; these protocols must include medication reconciliation, structured patient and family education, and pre-discharge follow-up appointment scheduling. Secondly, to leverage these protocols effectively, clinicians should adopt early identification of high-risk patients at admission using simple criteria like comorbidity indices and cognitive screening to trigger intensified discharge planning. Thirdly, discharge processes must be tailored specifically for patients with cognitive impairment, utilizing simplified instructions and ensuring mandatory caregiver involvement. From a policy perspective, the Ethiopian Federal Ministry of Health and hospital governance bodies should formally integrate 30-day readmission rates as a key performance indicator to incentivize quality improvement initiatives focused on care transitions. Finally, researchers should build upon this work by designing intervention studies, such as stepped-wedge cluster randomized trials, to rigorously evaluate the effectiveness of implementing these standardized protocols in similar low-resource settings, and further investigate the impact of other organizational factors like staffing ratios and social work integration on readmission risk.

Author contributions

Awoke Fetahi Woudneh conceived and designed the study, collected and analyzed the data, interpreted the results, and drafted the manuscript. He critically revised and approved the final version of the manuscript for submission.

Funding

This study was not supported by any specific funding from public, commercial, or nonprofit organizations.

Data availability

The raw data analyzed in this study are available from the author upon reasonable request. However, the data will not be made publicly available due to concerns about protecting participants’ identities and respecting their privacy rights. Informed consent for the publication of the dataset was not obtained at the time of data collection. Please contact Mr. Awoke Fetahi Woudneh at aweke_fetahi@dmu.edu.et.

Declarations

Competing interests

The authors declare no competing interests.

Ethics approval and consent to participate

The Ethics Committee of the Statistics Department at Debre Markos University provided ethical approval for this study. The committee granted approved of secondary data and issued a formal ethical clearance certificate (Reference No. STAT/30/6/2022). All research involving human data was conducted in accordance with relevant ethical guidelines and regulations. Informed consent was waived by the ethics committee due to the retrospective nature of the study and the use of anonymized secondary data. Patient confidentiality was maintained through the anonymization of all records and secure data storage protocols.

Consent to publication

This manuscript does not include any identifying images or other personal or clinical details of participants that could compromise anonymity.